# Research Papers — 2026-09 · bundle page 237 of 266 > Full-text markdown of papers #11801–11850 of 13298 published in 2026-09, newest first. AI-readable mirrors of the full 13298-paper 2026-09 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2026-09/page/237 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/236.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/238.md) ### $τ^τ$-Bench: An Environment for End-To-End, Realistic Agent Construction Authors: Quan Shi, Keshav Dhandhania, Karthik Narasimhan, Victor Barres Published: 2026-09-04 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.04611v1 LLM agents are rapidly becoming production software, deployed to handle customer service, adjudicate disputes, and operate internal systems. Notably, the work of building them is increasingly handed to coding agents, yet existing benchmarks say little about whether an AI system can deliver one under the conditions of a real client engagement. We introduce $τ^τ$-bench (pronounced hyper-tau-bench), a benchmark that makes agent construction the task. A developer agent is given the records a business actually keeps, a client who holds requirements, a production API that operations must run through, a codebase to inherit, and limits on serving cost and models: the same starting point a real engagement provides. From these it must deliver a complete customer-service agent, scored by deploying that agent against held-out simulated users. Across 53 tasks spanning four domains, the strongest configuration, Claude Opus 5 under Claude Code, passes just 23.9% of evaluation simulations. Meanwhile, an expert-authored reference ceiling scores 82.2%. The failures mirror ones human agent developers see: models issue shallow queries in place of deep comprehension of the records, communicate almost nothing to the client, and experiment too little with agent architecture and serving spend, shipping the first design that runs. We aim for $τ^τ$-bench to turn the work of cooperative agent building into a measurable target for coding agents. --- ### An Evaluation Framework for Generating Multi-View Images of a Person in a Scene Authors: Mahir Majid, Young Kyung Kim, Guillermo Sapiro Published: 2026-09-04 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.04603v1 Recent generative image-editing Diffusion Transformers (DiTs) demonstrate impressive semantic editing capabilities but still struggle with spatially consistent camera angle changes. A primary bottleneck in training foundation models to execute free-form, promptable camera angle changes is the lack of specialized training data. While multi-view datasets exist for generic 3D environments and objects, there remains an absence of paired, multi-view datasets featuring human subjects at fixed locations in natural scenes, including frontal and side-profile views. Capturing such multi-camera data in unconstrained environments is logistically challenging and unscalable. In this paper, we first experiment with multiple state-of-the-art image editing models to create this data synthetically, but find that the outputs are frequently prone to hallucinations involving how much the subject's head turns relative to the background, often producing inconsistent environments. To address this issue, we propose the Head Scene Rotation Difference (HSRD) metric to quantitatively evaluate camera movements around a person. The proposed metric operates by decoupling camera movement from localized head pose manipulation. As demonstrated by the extensive experimentation, HSRD provides the pipeline necessary to evaluate 3D spatial parallax for a person in a scene, paving the way to reliably construct high-quality multi-view synthetic datasets. --- ### PetQA: Benchmarking Veterinary Knowledge and Clinical Reasoning Authors: Taegyun Kim, Youngwook Ham, Jungwook Rhim, Ju-Hyun An, Sungkyu Park, Kunwoo Park Published: 2026-09-04 Categories: cs.CL, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.04598v1 We introduce PetQA, a Korean long-form question-answering (QA) benchmark for evaluating veterinary knowledge and clinical reasoning in large language models (LLMs) and large vision-language models (LVLMs). PetQA contains 10,076 text-only and 8,751 multimodal QA pairs derived from real-world questions about dogs and cats, paired with answers from expert veterinarians. Its test split, PetQA-Bench, further includes annotations for question types and clinical conditions. We evaluate eighteen models using ROUGE, BERTScore, and LLM-as-a-judge metrics for factuality and helpfulness under three settings: zero-shot inference, retrieval-augmented generation (RAG), and supervised fine-tuning (SFT). The benchmarking results provide an overview of the strengths and limitations of current models in addressing veterinary clinical queries and highlight the need for more effective adaptation methods to develop clinically reliable AI systems for veterinary care. To facilitate broader use, we additionally provide translated versions of PetQA-Bench in five languages. --- ### GNN-Guided Graph Coarsening and Adaptive QUBO Penalties for the Capacitated Vehicle Routing Problem with Time Windows on a Quantum Annealer Authors: Youssef Kamel Rezk, Paweł Gora Published: 2026-09-04 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.04593v2 Graph coarsening reduces the large Quadratic Unconstrained Binary Optimization (QUBO) formulations arising when vehicle-routing problems are solved by quantum annealing. Nearby customers with compatible time windows are merged into super-nodes, the reduced problem is solved, and the solution is expanded to the original graph. For the Capacitated Vehicle Routing Problem with Time Windows (CVRPTW), existing coarsening heuristics require family-specific tuning and remain unreliable on random instances. We address these limitations on the Solomon benchmark using simulated annealing and a D-Wave Advantage2 processor. We first introduce adaptive penalty calibration. Uniform penalty scaling has little effect, whereas controlling the internal coefficient range substantially improves raw samples. Removing non-binding constraints, normalising binding ones, and scaling the remaining penalties reduces mean raw constraint violations from 33.0 to 0.06 at the same solver budget (p=3.7e-11, n=56). A variable-count-preserving control attributes this gain to conditioning rather than problem size. Second, we replace the hand-tuned merge score with a graph neural network (GNN) using one configuration across all families. At N=10, it achieves 100% feasibility across all Solomon families, including R-type (100% vs. 80% for the tuned heuristic). Across N=10,...,100, feasibility is 83% vs. 69%, with the GNN better or tied on 85/90 instance-size pairs. At N=80,100, the difference is significant (p=0.002; 25/25 pairs), while the QUBO remains approximately 5-6 times smaller. Finally, hardware experiments reproduce the conditioning effect at fixed logical variable count: feasible samples increase from 0.02% to 39% across 13 instances. Classical repair with local search remains a reference bound for end-to-end solution cost. --- ### Hidden In Plain Gaze: Gaze Representations as Privacy Controls for Utility and Re-identification Risk in XR Authors: Cory Ilo, Brendan-David John, Doug A. Bowman Published: 2026-09-04 Categories: cs.CV, cs.ET, cs.HC, cs.LG arXiv: https://arxiv.org/abs/2609.04592v1 Intelligent extended reality (XR) systems increasingly use eye and head tracking to infer user intent, task, and attention, but the same signals can also reveal biometric identity. We study whether gaze data representation choice can serve as a lightweight privacy control at feature extraction, before adding perturbation or formal privacy mechanisms. Using the egocentric HoloAssist dataset, we compare three gaze representations under matched model capacity: raw gaze, spatial attention heatmaps, and engineered eye-movement features. We evaluate each representation on action recognition as task utility and closed-set user re-identification as privacy leakage. Representation choice substantially changes the privacy-utility tradeoff. Engineered features retain roughly 85% of raw gaze's action-recognition accuracy while reducing re-identification by about an order of magnitude, to roughly four times the chance rate across 206 identities. This reduction attenuates rather than eliminates identity leakage, and the differences across representations show that abstraction alone does not guarantee privacy. Engineered features expose interpretable and auditable structure, giving designers a transparent privacy lever that complements mechanisms such as differential privacy. --- ### Dual-Part Multi-Lateral Branched Network for Multi-Class Segmentation in Cardiovascular Catheterization Angiograms Authors: Olatunji Omisore, Ahmed Elazab, Ali Shahidinejad, Fariza Sabrina Published: 2026-09-04 Categories: cs.CV, cs.AI, cs.RO arXiv: https://arxiv.org/abs/2609.04590v1 Catheterisation image processing requires segmentation models that are fast, accurate and explainable. While most of the existing studies usually focus on binary segmentation, there is a recent demand for simultaneous segmentation of multiple structures found in catheterization scenes. In this study, a dual-part MLBNet architecture is designed with multi-lateral encoder blocks and multi-head decoder branches for class-aware segmentation in cardiovascular catheterization scenes. Lateral branches in the encoder enables repeated feature extraction to learn diverse shared representations, while multiple decoder heads are used to introduce class-skewed branches that specialize in different structural properties in catheterization scenes. To analyze the performances of the dual-part MLBNet architecture, several multi-class segmentation angiogram data obtained during cardiovascular catheterization in phantom models, synthetic human-simulated aorta, and animal model are used for model training and evaluation. Results obtained showed the dual-part models could effectively separate guidewire, catheter, vessels and background pixels to their classes of memberships with high probability. The results demonstrate that all models were able to distinguish the dominant background class from foreground structures with high overall accuracy. --- ### JLIR: A Julia-Native MLIR-Inspired Intermediate Representation with Automatic JACC Kernel Extraction Authors: Narasinga Rao Miniskar, Seyong Lee, Keita Teranishi, Jeffrey S Vetter Published: 2026-09-04 Categories: cs.PL, cs.CL arXiv: https://arxiv.org/abs/2609.04585v1 The Multi-Level Intermediate Representation (MLIR) has made reusable compiler infrastructure practical for domain-specific computation. However, MLIR's strong compile-time type requirements and low-level (C++) extension model can be a poor match for high-level, dynamically specialized languages such as Julia. MLIR has several drawbacks for dynamic programming languages in terms of the type system and level of abstraction. It is thus extremely challenging for non-compiler or scientific computing users to introduce new programming abstractions and express algorithm implementations in a form that remains both natural and optimizable. As a result, library interfaces for linear algebra, mesh processing, partial differential equations, and related domains often sit outside the compiler optimization path. We present JLIR (Julia-native Level Intermediate Representation), a Julia-native intermediate representation framework that brings the main benefits of MLIR-style multi-level, dialect-oriented compilation into the Julia ecosystem while remaining usable as ordinary Julia code. JLIR represents Julia programs before low-level lowering, supports extensible operations and transformation passes through Julia's language mechanisms, and allows partially typed programs to remain transformable until concrete types are known. The framework includes built-in dialects for arithmetic, control flow, functions, structured loops, and memory operations, and it also includes a lightweight mechanism for adding new domain operations without modifying the core system. To demonstrate JLIR's capabilities, we applied it to automatic Julia for Accelerators (JACC) kernel generation. --- ### Representation Redundancy and Structural Complexity in Finite-Field Inversion Authors: Zheng Zhang, Na Zhang Published: 2026-09-04 Categories: cs.LG, math.RA arXiv: https://arxiv.org/abs/2609.04583v1 The representation chosen for a mathematical operation can affect both its algebraic form and its empirical learning difficulty. We study this phenomenon for inversion over \(\mathbb F_{2^n}\), with field elements expressed in varying ordered \(\mathbb F_2\)-bases. We prove that two ordered bases induce the same coordinate inversion map if and only if they belong to the same Galois orbit. Since every orbit has size \(n\), the correspondence between ordered bases and distinct inversion maps is exactly \(n\)-to-one. We then analyze three Boolean formulations of inversion. The reference formulation has algebraic degree \(n-1\) and joint ANF leap \(1\), the mixed representation formulation has degree \(2(n-1)\) and joint ANF leap \(2\), and the complete raw formulation has degree at most \(3(n-1)\) and joint ANF leap at least \(n\). Exhaustive computations agree with the theoretical results and bounds in the cases considered. Controlled experiments with multilayer perceptrons show the same ordering in learning difficulty, while Galois orbit redundancy provides only a limited generalization benefit under the tested conditions. These results show that exact redundancy among representations can coexist with changes in Boolean structure and learning behavior when the representation is exposed as part of the input. --- ### When Do Internal Probes Beat Reading the Answer? Miscalibrated Readouts and Behavior-Concealed Knowledge in Language Models Authors: Gnaneswar Villuri, Hashmath Shaik, Alex Doboli Published: 2026-09-04 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.04582v1 A 0.6B language model, asked to verify 1,200 logical conclusions (half valid, half corrupted by a single semantic edit), answers YES every time. Judged by behavior it discriminates nothing; linear probes on its hidden states read the correct verdict at 0.96 AUC, transferring to unseen logical structures and separating foils built from exactly the words of the true conclusion (0.90). We ask where the verdict is lost, and find the dominant failure is a single scalar. The verdict survives to the model's own output logits (margin AUC 0.89) along a well-aligned readout direction; a saturated decision threshold, offset by +4.6 sigma, erases it. The diagnosis generalizes: across 90 semantic-label configurations of a five-model, three-family factorial, behavioral accuracy collapses onto a single function of threshold offset (Spearman -0.93) while margin ranking moves far less. Across a 13x scale range, internal knowledge saturates while free-form behavior is non-monotone: an 8B model underperforms its 4B sibling through an answer-channel failure rather than the threshold; forced-choice accuracy is monotone. The diagnosis is actionable: a one-parameter correction, never fit on evaluated structures, repairs behavior from 50% to 81% (0.6B); calibrated margin decoding recovers 94% at 8B; few-shot prompting works the same way, recentering the threshold (+4.6 sigma to 0.0 sigma) while preserving ranking. Comparing probe to margin separates three regimes: concealed, miscalibrated, and undetected. On a maze task built so foils carry no surface cues, the audit correctly reports the third. In the standard generation setting, answer-surface features and heuristic labels reproduce published probing results without any internal access. --- ### Does the Selected Object Reach the Reader? Auditing Identity Handoffs in Grounded Language-Model Pipelines Authors: Siddharth Vohra, Runmin Jiang, Xiaomo Li, Min Xu Published: 2026-09-04 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.04579v2 Grounded language-model pipelines can be divided into three stages: selecting an object, retrieving passages for it, and using that evidence to answer. If the selected object must reach the reader, losing it breaks the handoff. Benchmark recall checks the dataset-linked object, which can differ. We audit 600 HybridQA questions across three selector families. On 1,463 resolvable records where the selected object matches the dataset-traced passage, exact key lookup and exact title matching return the object every time. With every ranked rule given the same decoded selected title, body-only BM25 omits it on 389 records (26.6%) at cutoff five, while hybrid retrieval with reranking omits it on 14 (1.0%). The two identities differ on 329 of 1,792 resolvable records. With original-question rankings, their top-five checks disagree on 106 records (5.9%). Frozen reader comparisons associate the aligned object's presence with 28.6 to 31.0 points higher exact match. In a deliberately selected 64-item cohort, removing that passage sharply lowers exact match, while removing a similar-length comparison passage does not reproduce the drop. We release the Returned-Object Profile (ROP), an executable record of the target, returned-ID field, cutoff, membership rule, and complete expected population, with data and an offline replay. --- ### Centered Permutation Prefixes for SGD with Random Reshuffling: Sharp Rates, Hölder Geometry, and Composite Proximal Extensions Authors: Jiaxiang Li Published: 2026-09-04 Categories: math.OC, cs.LG arXiv: https://arxiv.org/abs/2609.04578v1 We study stochastic gradient descent with random reshuffling for finite sums \( F(x)=\frac1n\sum_{i=1}^n f_i(x). \) For fresh reshuffling with a constant component stepsize, if each $f_i$ has an $L$-Lipschitz gradient and the average $F$ is $μ$-strongly convex with a Lipschitz-continuous Hessian, we prove the last-epoch rate \( \mathbb E(F(y_K)-F(x_\star)) =\widetilde O\!\left(T^{-2}+n^2T^{-3}\right), \qquad T=nK, \) matching the known quadratic lower bound in its $(n,K)$-dependence. The components may be nonconvex, and no componentwise Hessian continuity or separate bounded-iterate assumption is required. More generally, a $ν$-Hölder-continuous average Hessian adds only $\widetilde O(n^{1+ν}T^{-2-2ν})$, so every $ν\ge 1/2$ preserves the quadratic rate. Under convex components, a decreasing-stepsize result removes the large-epoch requirement and recovers the same two-term scale once $nK$ exceeds the condition-number scale. We also analyze epoch-wise ProxRR for $\mathcal P=F+ψ$. Writing $x^\dagger$ for the composite minimizer and $β_\star=\|\nabla F(x^\dagger)\|$, we prove \( \mathbb E\|y_K-x^\dagger\|^2 =\widetilde O\!\left( \frac{β_\star^2}{K^2} +T^{-2}+n^2T^{-3} +n^{1+ν}T^{-2-2ν} \right). \) For $ν\ge 1/2$, we show that the $β_\star^2/K^2$ splitting term is unavoidable and obtain a matching lower bound up to logarithms in the stated constant-stepsize regime. --- ### Optimizer Memory Schedules for Outscaling the Overtraining Axis Authors: Katie Everett, Shikai Qiu Published: 2026-09-04 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.04577v1 We investigate how optimizers scale across the overtraining axis and show that relative optimizer performance and optimal hyperparameters change substantially with training horizon. In particular, we study how matrix-preconditioned methods (Muon and SOAP) and a momentum-scheduled method (ADANA) scale relative to AdamW. We compare these four optimizers across models from 51M to 253M parameters and overtraining (OT) factors from 1x to 256x, sweeping the base learning rate at every setting. The preferred learning rate schedule can reverse across the overtraining axis, the best weight decay coefficient scales approximately as sqrt(OT), and longer horizons generally favor longer fixed memory. ADANA's scaling advantage over AdamW persists after tuning AdamW's fixed memory separately at each horizon. Log-time weight decay and momentum cooldown provide substantial gains for ADANA that compound as training increases. With this treatment, ADANA outscales AdamW with an exponent advantage close to that predicted by DANA theory on power-law random features. Muon and SOAP instead provide roughly constant token-efficiency advantages over AdamW across most of the measured range, although SOAP may gain further at the highest overtraining factors. ADANA begins behind both matrix-preconditioned optimizers but closes the gaps as training increases, surpassing Muon and becoming competitive with SOAP at our highest OT factors. These results establish training horizon as an essential axis for optimizer evaluation and design. --- ### Training-Free Halving of Activated Experts in Fine-Grained Mixture-of-Experts Models Authors: Xing Chen, Hengshuai Yao Published: 2026-09-04 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.04575v1 Modern fine-grained Mixture-of-Experts (MoE) models route each token to a small number of experts and renormalize their router probabilities. We show that this renormalization implicitly calibrates expert output gain to the training top-$k$: reducing $k$ at inference changes not only which experts are used but also the strength of the expert branch. We separate these effects by activating the top $k_1$ experts while normalizing by the probability mass of the top $k_2$ experts, introducing one integer with no parameters, training, or measurable compute overhead. On Qwen3.6-35B-A3B, reducing from 8 to 4 experts causes a 4.65-point MMLU drop under standard renormalization but only 0.35 points with $k_2=16$, while halving routed-expert compute. The result replicates on the $11\times$ larger Qwen3.5-397B-A17B, where reducing from 10 to 5 experts loses only 0.55 points with an appropriate reference set. Removing renormalization entirely is catastrophic, showing that preserving a suitable reference mass is crucial. We further find that perplexity and downstream accuracy favor different $k_2$, cautioning against selecting MoE compression settings using unlabeled text alone. Analyses also show that expert identity matters substantially more than expert weighting, while balanced and domain-specialized routing leaves limited room for expert pruning. --- ### MURAL: Multimodal Uncertainty-aware Recommendation via Adaptive edge Learning Authors: Ahmad Mousavi, Majid Alikhani, Yeon-Chang Lee, Roberto Corizzo, Yeganeh Abdollahinejad Published: 2026-09-04 Categories: cs.IR, cs.LG, cs.SI arXiv: https://arxiv.org/abs/2609.04574v1 Multimodal Graph Neural Networks have become standard for recommendation by augmenting sparse interaction data with content features. Yet current architectures face two bottlenecks: structural rigidity, from a reliance on static precomputed similarity graphs that cannot adapt to evolving preferences; and semantic fragility, where noisy modality signals are indiscriminately fused, distorting the collaborative signal. We propose MURAL (Multimodal Uncertainty-aware Recommendation via Adaptive edge Learning), a unified framework that shifts multimodal recommendation from fixed structural augmentation to dynamic topology discovery. To address structural rigidity, an Adaptive Edge Learner combines a differentiable retrieval-augmented strategy with an approximate nearest neighbor search to discover latent item-item correlations that are both semantically adaptive and computationally scalable (O(NlogN)). To address semantic fragility, an Uncertainty-Aware Fusion module models the aleatoric uncertainty of heterogeneous modalities, dynamically down-weighting unreliable features while prioritizing high-confidence signals as a defense against cross-modal noise. We further employ a contrastive teacher-student alignment that anchors modality-specific representations to stable behavioral signals, ensuring optimization stability without gradient leakage. Experiments on large-scale benchmarks including TikTok and Amazon show that MURAL significantly surpasses both structural and generative state-of-the-art baselines, achieving superior accuracy while offering interpretability through domain-specific modality dominance and robustness under extreme data corruption. --- ### Finding Icebergs in Language-Model Workflow: Diagnosing Latent Structural Fragility with Stochastic Semantic Evidence Graphs Authors: Matthew F. Dixon, Bertrand Nortier, Miquel Noguer i Alonso Published: 2026-09-03 Categories: cs.CL, cs.IR, stat.ML arXiv: https://arxiv.org/abs/2609.29703v2 AI-workflow governance cannot be reduced to checking the final answer: an apparently safe answer may rest on a fragile evidence path that ordinary evaluation cannot see, localize or govern. We call this hidden fragility a "structural iceberg": hallucinations and unsupported claims may form its visible tip, while consequential weakness remains submerged. Stochastic semantic evidence graphs (SSEGs) expose these icebergs by preserving workflow channels, propagating local uncertainty and identifying the hidden paths on which an apparently safe output depends. ALCE and RAGTruth show that visible failures at the tip -unsupported citations and hallucinated spans -rest on distinct submerged weaknesses and therefore require different interventions. Across retrieval, tool-use and controlled stress tests, SSEG localizes those weaknesses, supports targeted repair, produces no false automatic passes in 35,000 known-truth cases and reduces ToolSandbox review by 28.8% across 96 executions from two agent models. The same structural view carries into end-to-end governance: in a separately sealed 1,200-case FinGovBench study, adding SSEG to GPT-OSS-20B reduces unsafe releases from 452/660 to 8/660 while releasing all 540 safe cases and correctly distinguishing 592/600 matched workflow pairs. An unchanged-gate transfer to Qwen3-8B releases all 540 safe cases and none of 660 unsafe cases, whereas flat-UQ releases 520 unsafe cases. SSEG therefore moves governance below surface-level output checking, turning hidden evidence dependencies into auditable, path-specific decisions about intervention, revalidation and release. --- ### CoVeR: Coverage-Based Routing of Verifier Calls in Agentic Retrieval Authors: Daeyoung Roh, Donghee Han Published: 2026-09-03 Categories: cs.CL, cs.IR arXiv: https://arxiv.org/abs/2609.26086v1 An agentic retrieval system issues a sequence of search queries and must decide, at each step, whether the evidence collected so far is enough to stop. Delegating that decision to an LLM verifier or a prompt judge makes stopping reliable, but the verifier then reprocesses the growing evidence after every retrieval step, a substantial repeated cost. We show that most of these calls can be skipped without materially changing answer accuracy: a single threshold on a frozen sentence-embedding coverage margin detects the states in which the evidence is still plainly incomplete, and the verifier is called only on the ambiguous remainder, a gate we call CoVeR (Coverage-based Verifier Routing). Across three multi-hop QA benchmarks, with the evaluation protocol fixed before the full-scale run, the CoVeR-gated agent matches the answer accuracy of both the full-budget agent and the always-verify baseline within a fraction of an EM point. It cuts 62-68% of verifier calls, and 93% in a saturated regime. Routers built on evidence counts, lexical overlap, or BM25 relevance, alone or learned in combination, give weaker overall trade-offs, the gate transfers without re-tuning across deciders and agent scales, and its drafter distills into a 921k-parameter head atop the frozen encoder, leaving no LLM in the routing loop. The same signal cannot replace verification: matching a claim is far easier than deciding the claim is supported. --- ### Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks Authors: Avinash Amballa, Yashas Malur Saidutta, Wenbo Li, Lazar Valkov, Srinivas Chappidi Published: 2026-09-03 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.22237v1 Merging low-rank adapters (LoRAs) promises to eliminate the overhead of swapping task-specific weights at inference time. However, existing merging methods assume every layer needs the same rank budget. Further, some methods assume that rank budget needs to be split equally among the tasks too. We show this uniform-budget assumption is a major source of the performance gap between merged and per-task LoRAs. However, rank selection is an NP hard problem. To this end, we introduce Net Utility, a data free metric that first decomposes every task LoRA by its Singular Value Decomposition (SVD) and scores each of those singular directions by its task utility and its interference with other tasks directions. Next, we globally pool these scores to select singular directions with the highest values with a constraint on the total number of directions selected. The proposed Net Utility metric is applied on top of five different merging methods across three different merging spaces. The merging is done over two sets of tasks, vision and language tasks. Net utility based rank allocation outperforms its counterparts without that allocation. On average, over vision tasks it achieves +2.1% improvement in performance, and +2.2% improvement over the language tasks. --- ### BizSage: A Self-Evolving Multi-Agent Framework for Business Research with Efficient Knowledge Retrieval Authors: Yuhe Wu, Guangyu Wang, Jiaxin Liu, Guang Zhang Published: 2026-09-03 Categories: cs.CL, cs.MA arXiv: https://arxiv.org/abs/2609.22235v1 While multi-agent systems based on large language models (LLMs) have shown promise in automating the progressive workflow of academic research, extending them to economics and business research, where specialized domain knowledge spans neighboring disciplines yet remains difficult to access in a structured way, presents two challenges. First, existing methods mostly retrieve at the paper level, yet the evidence needed for research tasks is often distributed across different sections, creating a granularity mismatch that hinders retrieval coverage and precision. Second, these fields demand strict empirical rigor, yet current systems provide limited mechanisms for learning from evaluation feedback. We present \textbf{BizSage}, a multi-agent framework combining corpus-level fine-grained retrieval with quality-driven self-evolution. We build a Lateral Knowledge Graph (LKG) by merging section-level knowledge graphs and apply Personalized PageRank (PPR) to surface semantically relevant and structurally important sections. Seven specialized agents collaborate under a Meta-Review self-evolution mechanism that distills failure modes from evaluation traces into reusable strategies. On a benchmark spanning four domains and three tasks, BizSage ranks first on the majority of metrics, achieves pairwise win-rates above 60\% against six baselines, and produces zero hallucinated citations. We hope BizSage paves the way for reliable research assistance in economics, business, and the broader social sciences. --- ### Seeing Through Conflicts: Improving Instruction Hierarchy Alignment in Vision-Language Models Authors: Nicholas Sansoterra, Zishuo Zheng, Sachin Kumar Published: 2026-09-03 Categories: cs.CL, cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.22234v1 Instruction hierarchy (IH) alignment teaches language models to prioritize higher-level instructions when inputs conflict. While studied primarily in text-only settings, vision-language models (VLMs) introduce new challenges for IH: instructions may be embedded in images, split across modalities, visually transformed, or encountered during agentic tasks. Positing multimodal IH alignment as a reasoning problem, we train VLMs using reinforcement learning with rule-based rewards, comparing text-only, image-only, and mixed-modality supervision. We find that text-only IH training partially transfers to multimodal attacks, failing when models must decode, reconstruct, or reason over instructions across modalities. Image-based training improves robustness beyond text-only supervision, while mixed-modality training performs best overall. Importantly, the benefits generalize beyond the synthetic typographic training setting to real-image and web-agent safety tasks, while largely preserving general multimodal capability, showing that lightweight, verifiable supervision can meaningfully improve VLM robustness under adversarial, cross-modal, and interactive instruction conflicts. --- ### SCALE: Simulation-Calibrated Amortized Learning for Energy Materials (A hybrid architecture connecting deterministic modeling, real-world data, and transformer-scale inference for accelerated energy-materials discovery) Authors: Kuan Huang, Bo Bai Published: 2026-09-03 Categories: cs.LG, cond-mat.mtrl-sci arXiv: https://arxiv.org/abs/2609.22233v1 Energy systems face converging pressures for security, affordability, resilience, and sustainability, creating a need for faster discovery of deployable energy materials. Here we introduce SCALE (Simulation-Calibrated Amortized Learning for Energy Materials), a physics-grounded, real-world-data-calibrated learning architecture that connects deterministic scientific operators, experimental calibration, expanded calibrated label generation, and transformer-scale inference. SCALE converts selected high-cost mechanistic computation and measured evidence into reusable models for rapid screening, ranking, inverse design, and active learning. We formulate the framework, identify ten method-based application regimes, and demonstrate SCALE for solid-state metal-hydride hydrogen-storage capacity prediction. In this implementation, a hydride phase-equilibrium capacity operator is calibrated against 381 measured ML-HydPARK capacity anchors and used to generate 5,000 candidate-condition-prototype teacher labels. A crystallographically anchored periodic-graph representation preserves atomic sites, periodic neighbor relationships, and local metal environments absent from formula-only encodings. An edge-biased graph transformer with 2.90 million parameters reproduces calibrated teacher labels with five-fold surrogate fidelity of MAE 0.0582 wt% H2, RMSE 0.0833 wt% H2, R2 = 0.9927, and Pearson r = 0.9963. Post hoc attention analysis suggests that SCALE learns chemically organized element groupings and metal-metal relationships consistent with established hydride chemistry, without chemistry-group labels as supervision. Once trained, SCALE shifts million-candidate evaluation from repeated deterministic workflow execution to batched learned inference, reducing per-candidate screening cost by approximately 10^7-10^8 while retaining links to simulation and experimental evidence. --- ### CNA: An AI-Oriented Comprehensive Normalized Assessment for Healthy Status and Application to Optimize RRT Strategies by Reinforcement Learning Authors: Jiang Liu, Chan Zhou, Yujie Li, Di Wu, Yihao Xie, Peiwei Li, Xin Shu, Jiaqi Zhu, Chunyong Yang, Yuwen Chen, Bin Yi Published: 2026-09-03 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.22232v1 Millions worldwide require Renal Replacement Therapy (RRT) as a treatment essential for survival. However, optimizing RRT strategies via AI is challenging due to heterogeneous patient dynamics, missing data, and the absence of an AI-oriented health assessment criterion. We propose an AI-Oriented Comprehensive Normalized Assessment (CNA) for healthy status and apply it to optimize RRT strategies by using offline reinforcement learning (RL). The key idea of CNA is transforming vital-sign distributions into a standard normal space, enabling a unified, data-driven health-status score defined by deviations from referent intervals, which also provides an AI-oriented criterion to assess strategy quality and supports RL termination. We further design a structured 23-dimensional state representation that integrates 19 indicators with 4 RRT descriptors, and employ matrix decomposition to reconstruct missing vital signs, improving data completeness for learning. These components are incorporated into multiple offline RL algorithms and validated via systematic ablation studies on RRT feature subsets. Compared with physicians' observed treatments, the best learned strategy reduces mortality from 13.2% to 5.0% (reducing 62.24%) and shortens average in-hospital stay from 308.5 to 250.1 hours (reducing 18.93%), demonstrating both methodological innovation and the potential of CNA-guided RL to improve RRT outcomes in nephrology. --- ### EvalMem: An Operation-Level Diagnostic Framework for Long-Term Memory Systems Authors: Zeyu Liu, Jian Zhong, Rongduo Han, Ziyang Wu, Shunye Tang, Chenghao He, Yaxuan Yang, Yihang Qiu, Ailing Wang, Xiao Liang, Guohuan Xie, Xiaokang Xue et al. Published: 2026-09-03 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.22231v1 Long-horizon interactions with LLM-based assistants require memory systems that preserve and update user states, preferences, and interaction histories. Existing evaluations report end-to-end QA accuracy and cannot determine whether errors arise from encoding, retrieval, or generation. We introduce EvalMem, an operation-level diagnostic framework with three parallel Examiners. For each query, the Encoding Examiner checks whether the target fact is stored, the Retrieval Examiner assesses whether the native retriever returns usable evidence, and the Generation Examiner tests whether the model can answer from oracle evidence. Their outputs form fine-grained multi-label defect codes. To improve store-level diagnosis, we adapt agentic RAG with a recall-first strategy that searches using both the query and source evidence, increasing recall of present evidence on LoCoMo from 70.2% to 95.6%. Evaluations of seven memory systems on LoCoMo, LongMemEval-S, and dynamic DynaMem-Bench identify retrieval as the most frequently attributed failure layer; in default LoCoMo, retrieval defects reach 22.1%, compared with 7.7% for encoding and 6.5% for generation. Guided by this diagnosis, MemWiki, a search-friendly auxiliary structure built from each system's memory export, improves mean accuracy by 2.5 and 2.3 percentage points on LoCoMo and LongMemEval-S, respectively. --- ### List Counting Failures Are Not One Phenomenon Authors: Iyad Ait Hou, Saad Mankarious, Aya Zirikly, Rebecca Hwa Published: 2026-09-03 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.22230v1 Counting the items in a bracketed list looks trivial, yet open-weight chat models often get it wrong. Prior work usually blames input bottlenecks such as subword fragmentation or attention dilution, which predict that different models should fail in roughly the same way. Across seven instruct models on identical prompts, however, wrong answers form distinct modes: Qwen and Gemma 27B often flip odd lengths to a nearby even integer, OLMo concentrates errors on a few mid-sized integers, and Llama tends to under-count. These modes are useful labels rather than a stable family law (Gemma 9B does not reproduce Gemma 27B's odd-to-even drop), and heavier subword fragmentation does not make counting harder on our benchmark. When the model answers incorrectly, a linear probe can usually still recover the true count from the residual stream. Matching the same odd-to-even error also does not imply the same late-MLP magnitude fix: scaling a late MLP output helps Qwen modestly but is near null on Gemma 27B under the same protocol, while residual steering can move both only by trading odd gains for even losses. These results caution against transferring that magnitude fix across models without a transfer check. --- ### A Synthetic Multivariate Refrigerator Time-Series Dataset for Predictive Maintenance Authors: Islam Benamirouche, Feriel Fass, Djemel Ziou Published: 2026-09-03 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.22229v1 We generated synthetic multivariate time series for 27 refrigerators with a simplified physicsinspired simulator at one-minute resolution. The simulator includes ambient-temperature variation, door use, thermostat and compressor operation, heat exchange, defrost, electrical consumption, and six progressive degradation types. Each refrigerator provides 15 to 20 sensor outputs according to its configuration. The dataset contains 7,066,161 rows in 27 time-series files and 27 failure logs. The release also includes the Python generator, refrigerator configurations, and documentation. The data can support failure prediction, degradation analysis, and learning across refrigerators with different sensor-output sets. --- ### Assessing Adversarial Robustness of Latent Reasoning Models Authors: Shaolong Chen, Ang Li, Mingjie Li, Yisen Wang Published: 2026-09-03 Categories: cs.CL, cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.22228v1 Large language models increasingly rely on long chain-of-thought (CoT) trajectories for complex reasoning, but autoregressive generation brings substantial memory and inference costs. Latent reasoning models (LRMs) offer a more efficient alternative by compressing intermediate reasoning into a small number of continuous latent vectors. Despite their efficiency, however, the adversarial robustness of LRMs remains largely underexplored. In this work, we systematically evaluate the robustness of latent reasoning across textual and multimodal settings, covering eight models and six benchmarks. We find that, across our evaluated settings, LRMs are generally less robust than explicit CoT baselines under adversarial perturbations, with particularly severe degradation under white-box attacks. Further analysis reveals distinct failure modes across modalities: textual latent states exhibit brittle dynamics and high sensitivity to specific input patterns, while latent states in multimodal models can remain largely invariant to input perturbations and have limited influence on final predictions. These findings expose robustness limitations of current latent reasoning approaches and highlight the need to jointly consider efficiency and robustness when designing implicit reasoning systems. We have open-sourced our code to facilitate reproduction of our research https://github.com/PKU-ML/latent-reasoning-model-assessment. --- ### Guiding the coarse levels of semantic IDs makes the fine levels learnable Authors: Bin Wang, Zhengyu Zhang Published: 2026-09-03 Categories: cs.IR, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.22227v1 Generative retrieval represents each item by a short Semantic ID and casts recommendation as autoregressive generation of that sequence. Because the tokenizer is trained independently to reconstruct an item embedding, its codes are aligned with neither the downstream LLM nor the end task. Nearly every SID system therefore spends extra effort to bridge this gap--alignment corpora, reasoning/RL, or per-token encoders to make codes legible, or learned tokenizer supervision to make them task-aware--yet the recovered meaning is content-derived and may not be the meaning the task needs. We introduce Guided SID, which instead makes the levels that matter most meaningful by construction: we force the coarse RQ-VAE levels to encode a predefined categorical attribute--chosen to be text-grounded (hence legible to the LLM) and task-relevant--by deterministic supervised index assignment (overriding nearest-neighbor selection with the attribute label) while keeping the codebooks learnable (they still receive reconstruction gradients). A trie-merge construction maps any high-cardinality or set-valued attribute onto the fixed code budget while keeping merged buckets semantically coherent. Guiding costs nothing intrinsically: collision and reconstruction match or beat the vanilla baseline despite pinning the coarse level. In a matched end-to-end A/B differing only in the SID encoding, the guided retriever improves recall@k at every list length we measure (1.36x at k=1, 1.39x at k=10), raises mean reciprocal rank from 0.0260 to 0.0355, and predicts the pre-defined attribute 4.2x more often. --- ### Swiss-Knife: A Framework for Reconfigurable Externalised Multi-Objective Alignment at Decode Time Authors: Agnibh Karmakar, Mayur Parvatikar, Shreyash Dhoot, Amit Dhanda, Aman Chadha, Kapil Wanaskar, Vinija Jain, Amitava Das Published: 2026-09-03 Categories: cs.CL, cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.22226v1 Decode-time alignment methods steer a frozen language model by scoring candidate continuations with an external reward and selecting the maximiser. We argue that this shared design is a single degenerate point in a much larger space. We introduce Swiss-Knife, a framework for externalised multi-objective alignment in which the alignment specification is a first-class runtime object: hot-swappable scoring blades, a batch normaliser, a pairwise aggregation operator, and a selection rule. Six axioms characterise the admissible aggregation operators, and we prove a representation theorem: every operator satisfying them has the form $R_i = \sum_{j \neq i} g((μ_i - μ_j)/s(σ_i,σ_j))$, a two-parameter family containing probit and logistic comparison rules and pointwise argmax as named coordinates. Within it, pairwise aggregation is Lipschitz-stable under adversarial reward contamination while argmax is not, and Candidate-Batch Normalization (CBN) makes the weight simplex invariant to the rescalings under which reward models are only ever identified. Our reference instantiation pairs DPO-LoRA blades with an uncertainty-aware pairwise tournament. Sweeping the helpfulness/honesty/harmlessness simplex, it attains the best balanced frontier of six decode-time methods (harmonic $F_1$ 0.797 vs. 0.750 for the strongest baseline, $p < 10^{-14}$) with the lowest refusal rate and highest helpfulness of any arm, and reconfigures its objectives in 0.05 ms with no gradient computation. Ablating CBN costs 0.217 $F_1$, and the metadata confirms the predicted mechanism: the lowest-variance blade retains 9% of its nominal 33% influence, and the collapse follows the predicted ordering. --- ### Cascaded Non-Line-of-Sight Imaging Authors: Diego Royo, María Peña, Forrest B. Peterson, Andreas Velten, Julio Marco, Diego Gutierrez Published: 2026-09-03 Categories: cs.GR, cs.CV arXiv: https://arxiv.org/abs/2609.16017v1 Time-of-flight non-line-of-sight (NLOS) imaging recovers information from hidden objects by analyzing the time of flight of indirect photons scattered on a visible (relay) wall. Most methods make the simplifying assumption that photons travel exclusively three-bounce paths, thus ignoring other useful information encoded in higher-order photons (with, e.g., four- or five-bounce paths). We present a novel cascaded NLOS imaging approach that leverages higher-order information and allows imaging a broader range of single- and multi-corner scenarios. We combine ultra-fast laser scanning with recent time-gated 2D sensor arrays to capture the scene's impulse response on a visible relay wall. From the captured impulse response, our method computes an analogous virtual impulse response at any other hidden wall. This effectively allows us to concatenate a second, virtual NLOS imaging system that leverages higher-order illumination. We validate our cascaded imaging method both in simulation and with a real prototype, demonstrating NLOS imaging with fourth- and fifth-bounce illumination of objects in challenging orientations and hidden around two corners. We also analyze how wave-based NLOS imaging interacts with rough hidden walls, which explains and helps overcome existing visibility limitations. We further illustrate how to image hidden objects from different perspectives, thus observing previously unseen features, by relying on multiple hidden walls. --- ### GzDRL: Reproducible and Scalable Deep Reinforcement Learning with Gazebo Authors: Amal Dev Haridevan, Junjie Kang, Jinjun Shan Published: 2026-09-03 Categories: cs.RO, cs.LG arXiv: https://arxiv.org/abs/2609.13243v1 We present GzDRL, a novel single-process reinforcement learning (RL) framework for Gazebo that overcomes longstanding bottlenecks in scalable, reproducible robotics experimentation. Unlike conventional middleware-based RL-Gazebo integrations that suffer from nondeterminism and irreproducibility, GzDRL introduces a systematic, middleware-free environment-stepping mechanism that directly synchronizes agent actions and physics updates. This design enables deterministic, high-throughput data collection, efficient vectorization, and reproducible RL training and evaluation. Comprehensive benchmarks demonstrate that GzDRL achieves the highest workstation throughput among the evaluated frameworks while remaining competitive with GPU-accelerated simulators on laptop hardware, and maintains precise agent-environment synchronization, multi-agent scalability, and experiment-level reproducibility. We further validate sim-to-real transfer by deploying learned policies directly onto a physical quadrotor, without fine-tuning. Our results establish GzDRL as an accessible and reproducible platform for advancing RL in robotics and automation. --- ### Calibrating subgrid parametrizations of single-column ocean models via simulation-based inference Authors: Luben M. C. Cabezas, Sacha Wendling, Aurèle Gallard, Gabriel Mouttapa, Julien Le Sommer, Pedro L. C. Rodrigues Published: 2026-09-03 Categories: physics.ao-ph, cs.LG arXiv: https://arxiv.org/abs/2609.13242v1 Subgrid parametrizations of vertical mixing in ocean models depend on free coefficients that cannot be measured directly and must be calibrated against high-fidelity references such as large-eddy simulations (LES). Existing approaches return point estimates and leave the associated uncertainty unquantified, a limitation when the inverse problem is ill-posed or when distinct parameter configurations fit the data comparably well. Simulation-based inference (SBI) addresses exactly this: given a prior and access to the simulator, it approximates the full posterior over parameters without requiring a tractable likelihood, at a cost set by the number of simulator evaluations. We apply it to \texttt{tunax}, a JAX-based single-column ocean model, to calibrate the coefficients of its $k$--$\varepsilon$ closure. A blockwise PCA summary statistic compresses the simulator output along the depth axis while preserving its forcing--horizon--variable structure, making inference tractable at modest budgets. We compare neural posterior estimation and its sequential variants against a recent training-free approach built on a tabular foundation model. The latter recovers informative posteriors from a few hundred simulator calls, outperforming the trained estimators at every budget considered. --- ### ArtSociety: Multi-Agent Multimodal Collaboration for Art Emotion Understanding Authors: Jian Li, Fanfan Ji, Jinxiang Lai, Ying Tai, Jian Yang, Xiao-Tong Yuan, Chengjie Wang, Yabiao Wang Published: 2026-09-03 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.13240v1 The AffectiveArt Multidimensional Art Emotion Understanding task asks to jointly predict an artwork's fine-grained emotion (12 classes, 1549:1 head-to-tail ratio), binary valence/arousal, and five attribute-grounded descriptions -- sub-tasks that exhibit strong empirical trade-offs, so the single-model solutions we tried do not jointly optimize all of them well. We present ArtSociety, a multi-agent framework that assembles heterogeneous multimodal experts -- a DINOv2-Giant vision agent (A1), a scene-grounded CoT fine-tuned MLLM (A2), and three closed-source reasoning agents (A3-A5) -- and coordinates them with two training-free controllers: (i) a rare-class-aware voting arbiter that lowers the agreement threshold for tail emotions, exploiting decorrelated error patterns across agent families; and (ii) a description-first reasoning agent whose DESCRIBE-then-CLASSIFY chain of thought forces visual evidence before label commitment, yielding near-perfect grounded descriptions. A task-routing policy directs the hard emotion task to the full five-agent ensemble while assigning the near-saturated valence/arousal and generative description tasks to the single strongest reasoning agent. On the official test set (1,000 artworks), ArtSociety achieves an Overall Score of 0.8870 (Classification 0.7789, Description 0.9952). An eleven-variant ablation study reveals that, once method and scale saturate at around 0.76, the decisive gains come from agent collaboration and data-side supervision -- a 30B MoE model trained on older data does not outperform an 8B model trained on better data. Code is available at https://github.com/swordlidev/ArtSociety --- ### Abstract-LoRA: Unlocking Single-Image Style Transfer through Targeted U-Net Block Training Authors: Xinglin Hu Published: 2026-09-03 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.13239v1 Diffusion models represent one of the most advanced paradigms in generative modeling. Leveraging their development, a growing number of style transfer methods based on diffusion models have been proposed. However, among these methods, multi-image style transfer approaches that require at least five to ten style examples tend to achieve more satisfactory results. Single-image methods, by contrast, often struggle with either insufficient content preservation or inadequate style fidelity. This greatly limits style extraction from scarce artworks and undermines their artistic value. To address this, we propose Abstract-LoRA, a method that pushes the boundaries of single-image style transfer through lightweight LoRA training on specific U-Net blocks in diffusion models. Specifically, our work is inspired by B-LoRA, a style transfer method that achieves basic style-content disentanglement by training specific U-Net blocks. However, it suffers from a critical limitation: the inability to capture complex backgrounds. Building upon B-LoRA, our method conducts a more refined analysis of U-Net blocks, employing additional U-Net blocks and clustering-based abstraction of style images to better disentangle and balance style and content. Extensive experiments demonstrate that our proposed method not only generates visually more harmonious and satisfying artistic images but also quantitatively improves the preservation of both style and content in the final outputs. --- ### CompEvo: Competition-Induced Evolution for Multi-Agent in News-Driven Time Series Forecasting Authors: Yuxuan Zhang, Yangyang Feng, Yong Guan, Daifeng Li, Kexin Zhang, Junlan Chen, Bowen Deng, Jun Liu, Zehua Zeng Published: 2026-09-03 Categories: cs.NE, cs.MA arXiv: https://arxiv.org/abs/2609.09195v1 News-driven time series forecasting uses evolving textual events together with historical observations to predict future values, supporting applications such as market risk monitoring and resource scheduling. In multi-agent settings, two challenges still remain. The first is degeneration of thought, where agents converge to similar evidence-seeking behaviors. The second is insufficient theoretical grounding, where strategy updates are often heuristic and lack a principled formulation. To address the above challenges, we propose CompEvo, a competition-induced evolution framework for multi-agent news-driven time series forecasting. For theoretical grounding, we introduce an evolutionary game formulation to guarantee equilibrium existence and optimization convergence. Building on this formulation, we construct a trainable multi-agent evolution framework that integrates strategy execution, fitness-based differentiable selection, and competition-induced strategy evolution. CompEvo enables heterogeneous agents to explore diverse news evidence, converts forecasting feedback into differentiable influence weights, and evolves agent strategies under competitive pressure to preserve effective logic while maintaining diversity. Experiments on four real-world datasets show that CompEvo reduces RMSE by 27.3% and MAPE by 26.2% on average over strong baselines. Further analysis indicates that CompEvo successfully maintains diverse and specialized agent behaviors. --- ### Toward Sustainable Distributed LLM Inference: A Systems Synthesis and Research Agenda for an Energy-, Carbon-, and Cache-Aware llm-d Control Plane Authors: Twinkll Sisodia Published: 2026-09-03 Categories: cs.DC, cs.LG arXiv: https://arxiv.org/abs/2609.05565v1 Large language model (LLM) sustainability is increasingly a serving-systems problem, not only a training problem. In production, energy and carbon impact depend on more than model size: workload shape, batching, key-value (KV) cache reuse, prefill/decode placement, model and accelerator choice, power state, geographic carbon intensity, and service-level objectives (SLOs) all matter. Recent systems papers study many of these factors separately. This paper connects those results and asks a practical engineering question: what do they imply when the decision point is a distributed inference control plane such as llm-d? The contribution here is synthesis, not a new set of benchmark results. Reported performance, energy, carbon, and cost improvements remain the results of the cited papers and systems. I group the literature into recurring design patterns and use those patterns to sketch a Sustainable Inference Control Plane (SICP) for llm-d. The proposed control plane would consider latency, energy, carbon, cache reuse, serving cost, and quality when routing and scaling, while keeping TTFT/TPOT SLOs as hard constraints. I also outline an evaluation framework based on SLO-satisfied goodput per joule and per gram CO2e, together with a reproducible experimental plan. The main observation from connecting the literature is that sustainable LLM inference is unlikely to come from one "green" model or one accelerator; it is more naturally treated as a control problem across model, phase, cache, hardware, replica, region, and time. --- ### Rollcast: Proper-Score Gated Rolling Anchors for Adaptive Probabilistic Time-Series Forecasting Authors: Giancarlo Vercellino Published: 2026-09-03 Categories: stat.ML, cs.LG, stat.ME arXiv: https://arxiv.org/abs/2609.05561v1 Rollcast is a probabilistic forecasting method for univariate time series that combines a compact set of rolling statistical anchors rather than relying on a single global model. Rolling means, medians, extrema, regression endpoints, and quantiles define candidate forecast locations and a representation of the current state. A state-dependent softmax gate learns anchor probabilities by minimizing negative log predictive density, while residual distributions retrieved from similar historical states provide local uncertainty. Recursive simulation propagates the resulting mixture over multiple forecast horizons. The method is evaluated in a Monte Carlo study covering eight data-generating processes, including autoregressive, random-walk, local-trend, threshold, regime-switching, stochastic-volatility, heavy-tailed, and variance-break dynamics. Across 2,000 independent fitted series, Rollcast is compared with the true conditional predictive distribution generated by an oracle simulator. Overall empirical coverage is 86.2% for nominal 90% intervals and 91.5% for nominal 95% intervals. Predictive intervals are on average 13.6% wider than the oracle at the 90% level, while CRPS is 14.4% higher than oracle CRPS. Performance is closest to the oracle under autoregressive, threshold, stochastic-volatility, heavy-tailed, and variance-break dynamics, while local trends and regime switching are more challenging. The results indicate that Rollcast can construct competitive probabilistic forecasts from simple, interpretable local summaries, while also identifying limitations in calibration and recursive uncertainty propagation. --- ### LLM-Aided Design for Manufacturing: A Multi-Agent System for Intent-Preserving Redesign of CAD for Improved Manufacturability Authors: Kojo Welbeck, Xiangyu Shi, Zahra Sadeghi, Qi Zhu, Ping Guo Published: 2026-09-03 Categories: cs.CE, cs.CV, cs.MA arXiv: https://arxiv.org/abs/2609.05559v1 We introduce autonomous, intent-preserving Design for Manufacturing (DFM) redesign of CAD parts: given an engineer's CAD model, the method returns a variant that is easier to manufacture without losing its design intent. Generating such a redesign in a single shot is unreliable, since CAD fidelity degrades as parts grow complex; we instead produce it as a sequence of individually verified design transitions. Our DFM-Redesign pipeline realizes this with two coupled agent subsystems driven by a pretrained multimodal LLM: a DFM Reviewer that inspects the current design and proposes one intent-preserving manufacturability improvement at a time, and a CAD Modifier that executes each proposal as an edit to the part's CadQuery program. The CAD Modifier closes a verification loop, compiling every candidate edit and visually checking it against the intended change from multi-view renderings, then re-generating or re-instructing until the edit is accepted or abandoned. Iterating review and verified modification compounds edits into parts more complex than one-shot generators reliably produce, preserves the original intent at each step, and requires no fine-tuning. On a 46-part benchmark scored by chamfer distance to reference geometries, the CAD Modifier reproduces target parts more accurately on average than chain-of-thought single agents given the same tools, and ablations isolate the contributions of the visual review loop and of captioning the design state before each edit. A centrifugal pump casing built from 32 chained transitions illustrates the complexity reachable by compounding verified edits. This is a preliminary report: evaluation of the full review-and-redesign loop, including manufacturability gain and an operational measure of intent preservation, is ongoing. --- ### EdgeMem: LLM-Free Agent Memory Construction and Retrieval via Evidence-Preserving Multi-Anchor Hypergraph Authors: Zeyang Cui, Jiannong Cao, Zhiyuan Wen, Bo Yuan, Junlan Feng, Shengyuan Chen Published: 2026-09-03 Categories: cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.05553v1 Agent memory allows LLM agents to use earlier interactions when answering new queries. Existing methods often compress interaction histories into summaries or other LLM-generated representations. Repeated generation adds cost and can discard answer-bearing details before the system knows what a future query will require. We propose EdgeMem, an agent-memory method built around a simple principle: preserve original interaction turns and organize them through complementary content, temporal, and episodic cues. EdgeMem realizes this principle with a multi-anchor hypergraph constructed by lightweight local processing. Retrieval directly returns source evidence and reserves LLM use for final answer generation, combining structured access to multi-session histories with faithful retention of the original conversation. Experiments on LoCoMo and LongMemEval-S show strong retrieval and memory-grounded question answering; on LoCoMo, EdgeMem achieves the highest strict-judge score among seven reproduced systems under a shared prompt (61.01 versus 58.70), while construction and retrieval require no generative-LLM calls. Overall, EdgeMem shows that preserving and organizing source evidence provides an effective and efficient foundation for agent memory without generative memory management. --- ### An Agent Model Abstraction for Human-AI Teaming Cognitive Coupling Authors: Kolitha Kottagaha W. M, Jos A. C. Bokhorst, Ben Gaffinet, Christos Emmanouilidis Published: 2026-09-03 Categories: cs.HC, cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.05552v1 Industrial environments increasingly rely on collaboration between humans and AI-enabled agents. Effective teamwork requires aligning how agents perceive situations, plan actions to pursue goals, and adapt to changing conditions, yet existing systems lack mechanisms for cross-agent cognitive processes coupling. This paper presents a conceptual cognitive agent model that formalises cognitive coupling through eight components: Input, Process, Output, State, Value, Memory, World Model, and Goal. The model abstracts how agents coordinate and co-regulate their cognitive cycles, providing a basis for analysing distributed cognition and designing cognitively interoperable human-AI systems. --- ### Subject-Relative Micro-Motion and Sleep Dynamics for Near-Infrared Video Sleep Staging Authors: Kunmin Jang, You Rim Choi, Hun Heo, Heonjun Lee, Suahn Bae, Dongik Park, Hyun-Woo Shin, Hyung-Sin Kim Published: 2026-09-03 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.05550v1 Near-infrared (NIR) video is a promising modality for contactless sleep monitoring, but recent video-based sleep staging methods often use it as a route to reconstructed respiratory/cardiac proxies or cross-modal physiological representations. We study video-only sleep staging under labels defined by polysomnography (PSG), where the model infers sleep stages from NIR video alone without explicit physiological proxy reconstruction or auxiliary physiological signal supervision. This tests whether NIR video itself can provide informative sleep-stage evidence, rather than only serving as an input for recovering physiological proxies. We propose ViNUSS (Video-Native Unmediated Sleep Staging), a framework that combines subject-relative micro-motion learning with full-night sleep dynamics modeling. Spatially anchored pre-spatial micro-motion encoding preserves localized temporal variation together with its spatial context. Within-subject stage contrast learns stage cues with respect to each subject's night-specific baseline. Two-scale sleep dynamics modeling captures within-epoch motion evolution and organizes epoch-level evidence into a coherent full-night sleep-stage trajectory. On 475 overnight NIR recordings (~3,250 hours), ViNUSS achieves 0.80 accuracy and 0.78 macro-F1 for four-class sleep staging. Interpretability analysis suggests attention to thoraco-abdominal periodic motion and gross body movements associated with arousals and position changes. These results support NIR video as an independently informative and complementary modality for PSG-defined sleep-stage estimation --- ### Dynamic Adaptation of the LLM Context for Generating Routines with Coupled Semantics Authors: Gnaneswar Villuri, Hashmath Shaik, Alex Doboli Published: 2026-09-03 Categories: cs.SE, cs.AI arXiv: https://arxiv.org/abs/2609.04570v1 LLM-based code generation fails when correctness depends on execution-dependent coupling: the meaning of one routine is defined by the runtime behavior of another, a relationship that cannot be resolved from textual descriptions alone. This limitation, which we call static binding, is not confined to explicitly coupled problems; it appears to varying degrees whenever correctness depends on joint execution behavior across components, from explicit cross-coupled optimizers to subtler joint constraints in packing, routing, and symbolic search. This paper proposes dynamic context adaptation, a sample-efficient validation-generation loop designed for this setting. A validation agent extracts structured diagnostic information from execution traces, providing gradient-like guidance to a generation agent that proposes multiple candidates per iteration. A knowledge graph derived from the problem description supplies semantic constraints to the generation agent. Simulated annealing selects among candidates to avoid greedy collapse. Our method outperforms zero-shot, Reflexion, and OpenEvolve on seven of eight problems at both 300 and 600 evaluations (p < 0.01), a regime where population-based search has not yet accumulated sufficient diversity to compete. Notably, on the primary motivating problem (cross-coupled optimization), our method also achieves the best score at 1000 evaluations, consistent with the hypothesis that structured execution feedback is most beneficial when correctness depends on runtime coupling. Ablation results confirm that structured execution feedback is the primary driver. --- ### Extremely Sparse Supervision Incentivizes Reasoning Ability Authors: Zhishuai Liu, Xingzi Xu, Mehmet Saygin Seyfioglu, Pan Xu, Karim Bouyarmane Published: 2026-09-03 Categories: cs.AI, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.04565v1 Large language models demonstrate increasingly strong reasoning capabilities through effective post-training. Yet, prevailing post-training methods optimize over massive numbers of tokens, implicitly assuming that effective learning must be token-intensive. We revisit this assumption in the on-policy distillation (OPD) setting, which naturally admits dense teacher supervision at every generated token. Using the Qwen3 family, we discover a counter-intuitive phenomenon: reasoning can be effectively incentivized by an extremely small fraction of generated tokens--as few as one or two tokens per reasoning trajectory, corresponding to only 0.05% of all tokens. Surprisingly, this sparse supervision in most cases matches or surpasses full-token training in improving reasoning ability, despite excluding the vast majority of generated tokens from the training objective. This phenomenon is consistently observed across nine teacher--student configurations spanning different model scales on mathematical reasoning tasks, and is further validated on coding reasoning, Llama models and Proximal Policy Optimization (PPO)-based reinforcement learning with verifiable reward (RLVR). Interestingly, such extremely sparse supervision may be closer to the natural learning process: rather than correcting every step word by word, one reflects on a few critical reasoning steps, updates prior understanding, and continues the trial-and-error, avoiding micro-level corrections while remaining remarkably effective. Overall, our results challenge the assumption that effective post-training must be token-intensive and point to a new direction for understanding and designing more efficient post-training algorithms. --- ### La Agente Óptima: Towards Agentic Self-Driving Laboratories Authors: Marcel Müller, Jiaru Bai, Willi Gottstein, Abhijoy Mandal, Mohammad Nazeri, Elia Savino, Yanlin Fang, Sujoy Das, Sergio Pablo García Carrillo, Yeonghun Kang, Juan B. Pérez-Sánchez, Simone Pilon et al. Published: 2026-09-03 Categories: cs.AI, cs.MA, physics.chem-ph arXiv: https://arxiv.org/abs/2609.04564v1 Self-driving laboratories (SDLs) combine automated experimentation with adaptive decision-making to accelerate scientific discovery. Their operation nevertheless often depends on human specialists who translate scientific objectives into executable closed-loop campaigns. Specialists adjust them as data and operating conditions change. Here, we present La Agente Óptima, an agentic framework that constructs and supervises Bayesian optimization campaigns across computational and experimental systems while maintaining a persistent optimization state. By separating large language model (LLM) reasoning from executed campaigns, Óptima runs repetitive optimization loops consistently, returns control to the agent only when progress requires interpretation or campaign revision, and keeps every decision auditable. We evaluate Óptima across ablation studies, five digital discovery tasks, and two physical platforms. Throughout, Óptima maintained executable campaigns as both the scientific problem and execution environment evolved. In a closed-loop contact angle optimization campaign, Óptima identified and corrected a mid-run measurement failure, bringing the contact angle from 71.4 to 67.8 degrees, just above the 64-66 degree range. From this result, Óptima correctly inferred that the target was likely unattainable with the available reagents and recommended changing the formulation. In a five-day multi-objective flow-chemistry campaign, Óptima increased the yield from 30% to 59% over 23 experiments. Despite substantial inference costs, it cost less and used substantially less starting material than a human-directed campaign, while selecting a more mass-efficient operating point. These results show that LLM-based agents can make rigorous, long-running optimization campaigns accessible to domain scientists without specialist setup, expanding the scope of SDLs. --- ### Reducing Hallucinated Transcripts in Whisper via Hallucination Space Projection Authors: Maryam Abbasihafshejani, Murtuza Jadliwala Published: 2026-09-03 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.04561v1 Whisper is a widely used foundation model for automatic speech recognition (ASR), but its generative decoder can produce fluent hallucinated transcripts for inputs containing little or no speech. We propose a training-free, inference-time method to reduce these hallucinations using low-rank projection of decoder activations. A compact hallucination-associated subspace is estimated from non-speech calibration data, and decoder hidden states are projected away from this subspace during inference. We evaluate two variants: always-on, which applies projection to all inputs, and gated, which applies it only when Whisper predicts that an input is likely non-speech. Across non-speech benchmarks, always-on projection reduces average hallucination rate (HR) from 31.31% to 2.44%, a 92.21% relative reduction, while gated projection reduces HR to 3.74%, an 88.05% relative reduction, with lower false rejection of genuine speech. On LibriSpeech, gated projection increases absolute word error rate (WER) by 0.33-4.39 percentage points and yields false-rejection rates (FRR) of 0.41--9.97% across model and split settings. These results show that low-rank activation projection can substantially suppress Whisper hallucinations without retraining, while providing a controllable trade-off between hallucination suppression and speech recognition performance. --- ### IPGeoAI: Transformer-Based Geolocation with LLM Semantic Fusion Authors: Avinash Kadimisetty, Andy Jinqing Yu, Philip Favaloro, Wenlong Liu, Xiaolu Xiong Published: 2026-09-03 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.04559v1 Accurate city-level IP Geolocation is an important enabler for the modern digital ecosystem, underpinning services ranging from local content delivery and targeting to digital rights enforcement. However, traditional heuristic and database-driven methods often struggle to resolve the complex, non-linear allocation patterns of modern network infrastructures, particularly within the exploding IPv6 address space and transient mobile networks. In this paper, we introduce IPGeoAI, a novel deep learning model architecture that reframes geolocation from a static lookup problem to a sequential modeling task. Our approach utilizes the Transformer Encoder to capture hierarchical dependencies inherent in IP subnet structures. We propose a method to resolve geographic ambiguity by integrating unstructured semantic context via a Zero-Shot LLM Feature Extraction pipeline. We utilize Large Language Models to transform raw, noisy Autonomous Systems (AS) descriptions into structured, domain-specific metadata (such as 'University' vs. 'ISP' or 'Global' vs. 'Local') via an offline pre-computation process. By fusing these semantic signals into the network via a Multi-Head Cross-Attention module, we bridge the gap between numerical network topology and real-world semantic identity. Extensive offline evaluation on a proprietary dataset spanning 200,000 cities demonstrates that IPGeoAI significantly outperforms a leading external vendor in city-level granularity. By adopting a hierarchical inference strategy that refines coarse-grained country signals, our model achieves a 6% improvement in city-level accuracy while extending coverage to 100% of the traffic. Furthermore, in large-scale online production tests, the model drove a statistically significant +0.35% improvement in our 1st-tier downstream use cases metric. --- ### A Sim-to-Real Study of Surface-Code Decoder Benchmarking Authors: Shay J. Manor, Leila S. Erhili, Yassine Jebbouri Published: 2026-09-03 Categories: quant-ph, cs.LG arXiv: https://arxiv.org/abs/2609.04557v1 Quantum error-correction decoders are typically benchmarked against synthetic circuit-level noise, under the assumption that a decoder's ranking under such noise transfers to hardware and improves as the noise model becomes more realistic. The Willow processor, the first to operate below the surface-code threshold, allows us to test this assumption. We rank a panel of six decoders using a four-rung ladder of noise models with increasing fidelity, evaluated against real data across three code distances, two bases, and fifteen round counts. Rank agreement with hardware appears once the noise model gives each operation type its own error rate. Calibrating the model to the device improves absolute error rates but not rank agreement. We additionally provide the first independent evaluation of NVIDIA's Ising pre-decoder on hardware, at code distances below its training receptive field and via a mapping onto the lattice on which it was trained. Under these conditions, it holds no accuracy-latency advantage: another panel decoder matches or improves on it in both per-cycle error rate and decode latency in 278 of the 280 evaluations. We release the full pipeline and the per-shot outcome of every evaluation, so future decoders and devices can be compared. --- ### Rhythms of Work: Multi-Scale Interpretation of Human Behavioral Traces for Workplace Agents Authors: Lin Ai, Scott Counts Published: 2026-09-03 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.04556v1 Runtime traces are becoming a central substrate for understanding agentic systems, yet interpretation has focused largely on what the agent did. Workplace agents face the complementary problem: interpreting the human activity that surrounds them. Hours of low-level events carry rich evidence about a user's state but are too granular to reason over directly, and flattening them into one stream or compressing them into a single embedding both treat "summarize the user's behavior" as if it had one correct answer. We argue instead that behavioral interpretation is resolution-dependent: the same trace should admit multiple addressable interpretations at different temporal resolutions. We construct a multi-resolution vocabulary of semantically normalized operators, recurring motifs, coherent episodes, and day-level rhythms, each preserving the structure salient at its own horizon. Applied to 667 million human-attributed events from a large commercial productivity suite (50,000 users, 100 organizations), it yields 120 operator types, thousands of motifs, 25 episode types, and five day-rhythm archetypes. We validate it on real telemetry: re-running the entire pipeline on a disjoint 2,000-user sample recovers the same taxonomy (structural stability), and on held-out users the full representation forecasts a user's next episode more accurately than a flat-operator baseline, a 17% relative macro-F1 gain (predictive validity), so the abstractions preserve future-relevant information rather than merely describe it. A controlled resolution ablation then shows that no single level is optimal across questions: different agent-facing questions about the same trace are best answered at different resolutions. Behavioral trace interpretation for agents should therefore be multi-resolution and query-conditioned: an agent should access the temporal grain a question needs, not one universal summary. --- ### DART: Depth-as-Target Pretraining for Surgical Vision Foundation Models Authors: John J. Han, Adam Schmidt, Muhammad Abdullah Jamal, Jie Ying Wu, Omid Mohareri Published: 2026-09-03 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.04555v1 Vision foundation models (VFMs) are valuable in data-scarce domains such as surgery, where a single pretrained backbone can provide rich representations for many downstream tasks. Yet the dominant self-supervised pretraining paradigm uses only RGB images, leaving readily available complementary signals, such as depth maps, unused. This is a particular missed opportunity in surgery, where natural-image VFMs transfer poorly while the scene geometry is rich and informative. With strong off-the-shelf models now able to produce pseudo-labeled dense depth for any image corpus, we hypothesize that such signals can be folded into pretraining to learn better representations. We present DART, an RGB-D pretraining recipe that builds on DINOv2 with a simple modification: a pixel-space depth reconstruction objective applied to masked iBOT patches, supervised by pseudo-labeled depth. Depth is used only during pretraining, so fine-tuning and inference remain RGB-only. We find that this pixel-level reconstruction head improves representation quality rather than disrupting it. We further show that depth, which encodes scene geometry, is more effective as a target than alternative dense signals such as Canny edges, confirming that the gains stem from depth rather than added supervision alone. Across eight surgical benchmarks spanning segmentation, depth estimation, and image-level recognition, DART outperforms both natural-image and in-domain baselines, including a vanilla DINOv2 trained on identical data, improving dense prediction while also strengthening image-level understanding. More broadly, DART shows that freely available geometric pseudo-labels can strengthen foundation model pretraining without extra labels or added inference cost, pointing toward stronger backbones for surgery. --- ### Continual Field-Adaptive Models (CFAMs) for Post-Deployment Physical AI Authors: Amarjot Singh, Tanmay R. Pancholi, Jainam Kothari, Shrirang Mahajan, Ketan Bansal, Zackory Erickson, Giuseppe Loianno, Alexandre M. Bayen, Jeff Schneider, Vince Nakayama Published: 2026-09-03 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.04552v2 Unattended interactive autonomy - machines that step into danger in place of humans and complete tasks with human tools - remains a missing capability in mission-critical operations. These domains offer scarce training data and only onboard compute, yet deployed systems must face novelty without erasing prior competence. We introduce Continual Field-Adaptive Models (CFAMs), which learn efficiently in the lab and continue learning after deployment through autonomous, gradient-free, on-device updates. CFAM uses a complementary learning architecture with a frozen slow-learning component and a fast-learning Capsule Field. The slow component contains three cortices: Sensor, which maps multimodal input into 3D-grounded geometry; Reasoning, which decomposes tasks into skills and evaluates outcomes; and Action, which executes geometric skills. The Capsule Field stores field learning one-shot and gradient-free as Competence Capsules. Skill installation is few-shot in the lab and continual in the field; open-world novelty is outside scope. We evaluate CFAM across five embodiments: manipulator, quadruped, humanoid, quadrotor, and off-road vehicle. Baselines (pi0, CogACT, SpatialVLA) use the same in-house multi-embodiment dataset for physical-platform comparisons. CFAM reaches the operating point of a standard policy trained on the full prior-training dataset using 40% of the data, or 2.5x fewer trajectories. At test time, autonomous capture of verified near-OOD cases improves action success by 13.9 percentage points. In sequential simulation, backward transfer is -0.5 percentage points versus -11.4 for LoRA. CFAM therefore provides a bounded form of post-deployment physical intelligence: few-shot skill learning, autonomous field growth from verified near-OOD experience, and retention of prior competence. --- ### VISTA: Dense Multi-Label Classroom Coding with Vision-Language Models Authors: Andrew Franck, Brendan Ng, Ben Fitzgerald, Zane Derrod, Chris Cianci, Chris Craney Published: 2026-09-03 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.04550v1 Video-language benchmarks are usually constructed by the dataset authors without published reliability statistics, leaving the noise floor of the construct unknown. We argue that multimodal benchmarking benefits from methods taken from research communities that have already invested in strategies to ensure reliability. We illustrate the case with the Classroom Observation Protocol for Undergraduate STEM (COPUS): a 24-code multi-label observation instrument with a decade of peer-reviewed reliability literature. We recast COPUS as a video benchmark for multimodal foundation models, where it provides a dense set of structured labels (a 24-dimensional binary vector every 2 minutes across a 50-90 minute lecture), an externally validated vocabulary, and established literature that provides a per-code reliability target based on human evaluators. Annotations in our evaluation corpus are produced by a 5-person human-evaluator panel whose consensus matrix is our reference. We propose VISTA, a baseline that runs MiniCPM-V-4.5 over a dense sliding window, refines its per-window outputs with a lightweight multi-layer perceptron (MLP) head trained on top of the frozen backbone, and max-pools the resulting predictions onto the 2-minute COPUS grid. On three held-out chemistry lectures, VISTA reaches 80.1% restricted macro accuracy versus 74.9% for the zero-shot variant, with the largest residual errors on visually similar instructor codes and on rare audio-dependent codes. We characterize three systematic failure modes (audio-partial observability, fine-grained group-work discrimination, long-tail recall) and release the benchmark tooling, prompts and baseline code at https://github.com/ajfranck/VISTA. --- ### Fast Surrogate Modeling of Excitable and Oscillatory FitzHugh-Nagumo Dynamics with Parametric Neural Operators Authors: Andrew Franck, Justin Li Published: 2026-09-03 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.04549v1 The FitzHugh-Nagumo (FHN) system serves as a simplified model of neuronal voltage dynamics, capturing the activator-inhibitor structure behind both isolated action potentials and the rhythmic spiking seen across the brain. Exploring its 5D physiological parameter space is important for neuromodulation and mapping voltage recordings back to biophysics, yet classical finite-difference solvers make rapid parameter sweeps expensive. We train parameter-conditioned Fourier Neural Operators (FNOs) as fast, differentiable surrogates for the FHN voltage and recovery fields on a one-dimensional spatial domain, conditioning each Fourier layer on the parameter vector $λ= (D_u, D_v, a, b, τ)$ via feature-wise linear modulation (FiLM). We apply a single bifurcation analysis that delimits the two distinct regimes the model spans, oscillatory (tonic firing) and excitable (action-potential propagation), and we train one operator in each. In the oscillatory regime the surrogate attains sub-$0.1\%$ relative $L^2$ error on both fields, runs nearly three orders of magnitude faster than the finite-difference baseline, generalizes uniformly across the parameter space, and extrapolates to low single-digit percentage errors outside of the training bounds. In the excitable regime the same operator accurately reproduces the firing threshold and the $c \propto \sqrt{D_u}$ conduction-velocity law and replicates full traveling pulses, fully capturing the excitable bifurcation structure rather than just smoothly interpolating fields. ---