# Research Papers — 2026-09 · bundle page 232 of 266

> Full-text markdown of papers #11551–11600 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/232
[← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/231.md)
[Next bundle →](https://aipod.dev/papers/archive/2026-09/page/233.md)

### Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference

Authors: Mostafa Elhoushi, Alex Pretko, Nolan Dey, Bin Claire Zhang, Gavia Gray, Gurpreet Gosal, Abdulrahman Mahmoud, Shane Bergsma, Joel Hestness
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05275v1

Layer dropout (a.k.a. stochastic depth) has been shown to enable faster training, higher accuracy, and robustness to zero-shot layer pruning in both language and vision transformers. However, as models and datasets have scaled, dropout - particularly layer dropout - has largely disappeared from large language models (LLMs) pre-training recipes. While some prior work has reported that dropout can degrade accuracy, no comprehensive study has quantified, let alone mitigated, this effect. In this study, we show that layer dropout should be used in state-of-the-art LLM training, establishing best practices and scaling analysis for both training and post-training benefits. Concretely, with optimal layer distribution, time schedule, and optimizer hyperparameters, we observe that at the same training FLOPs layer dropout leads to lower loss. For a given number of training steps, LLMs can achieve lower or similar validation loss while saving upto 25% of training FLOPs. Moreover, layer dropout enables significant post-training optimizations, such as early exit, intermediate-layer skipping, and self-speculative decoding, yielding up to 1.5x inference speedup with negligible accuracy loss. Across more than 2400 training experiments, spanning models from 271M to 8.2B parameters and datasets up to 160B tokens, we demonstrate that these findings extend reliably to large-scale training regimes. All pre-training experiments were run on Cerebras CS-3 systems.

---

### How to Speculate about Uncertainty in Agentic Coding? A Draft-Model Gate Method

Authors: Konstantin Grotov, Valentin Malykh
Published: 2026-09-04
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.05274v1

LLM agents deployed for software engineering fail expensively: they act confidently wrong, and bad actions are recognized only after costly execution and retry. We present Speculative Uncertainty (SU), a method that recovers a predictive failure signal for a black-box agent from its output tokens alone, with no access to logits, weights, activations, or repeated sampling. Inverting speculative decoding, a small open-weight draft model scores the agent's already-generated trajectory in a single forward pass. From these speculative cross-likelihoods we extract phase-aware features by separating the reasoning and action spans, and calibrate them against a verifiable objective. SU produces a failure-likelihood score that any downstream policy, such as routing, human intervention, or extra test-time compute, can consume directly. To show the signal is actionable, we instantiate one such policy, a pre-execution veto gate, on software engineering agents Qwen3-Coder-480B and closed-source Claude 3.5 Sonnet, cutting execution error rate by 6-8 percentage points and token cost by 14-19% in deployment, transferring to out-of-distribution benchmarks without retraining, and generalizing across agent models.

---

### AI for Computational Design Science: A Responsible Human-AI Framework and Case Study on Short-Form Video Safety Surveillance

Authors: Wenli Zhang, Jiaheng Xie, Zhihe Pan, Yidong Chai, Xiao Fang, Sudha Ram
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05270v1

Artificial intelligence (AI) is transforming not only what information systems researchers design, but also how design research is conducted. Yet existing literature offers limited guidance for computational design science (CDS) when AI actively participates in problem formulation, resource construction, design search, evaluation, and knowledge abstraction. We develop AI for Computational Design Science (AI4CDS), a five-phase methodological framework in which AI expands problem and design search while researchers retain responsibility for domain grounding, admissibility, verification, and scientific judgment. Collaboration is governed by graduated trust, reversibility, auditability, and differentiated reproducibility. We instantiate AI4CDS through ChildRiskGuard, an interpretable artifact for detecting short-form videos inappropriate for children, while documenting AI interactions, rejected alternatives, corrections, and audit trails. The case translates audience-dependent safety and explanation faithfulness into three technical challenges and develops an artifact that separates generic from child-specific risk, represents distinct developmental-risk mechanisms, and makes concept-level explanations part of the predictive computation. ChildRiskGuard achieves an F1 score of 0.769, substantially outperforming direct application of a general-purpose content-safety model while remaining competitive with strong benchmarks. The primary contribution is AI4CDS as a responsible framework for AI-enabled CDS; ChildRiskGuard provides process and artifact evidence of how AI-expanded, researcher-governed design can generate and evaluate novel computational design knowledge.

---

### CONTINUITY: Security-Context Contracts for Composable LLM Agent Controls

Authors: Chris Zheng, Geng Yang
Published: 2026-09-04
Categories: cs.CR, cs.AI
arXiv: https://arxiv.org/abs/2609.05269v1

LLM agent systems increasingly combine provenance tracking, authorization, policy enforcement, protocol adapters, and execution controls. However, individually correct security mechanisms do not necessarily compose into an end-to-end secure system: security-critical context may be dropped, widened, rebound, or reinterpreted as actions cross component boundaries. We identify this failure mode as security-context discontinuity and introduce CONTINUITY, a framework for verifiable composition of agent security controls. CONTINUITY models each component with an assume-guarantee contract and carries authenticated security context across transitions using signed root grants, provenance commitments, role-bound transition receipts, bounded typed releases, transformation witnesses, and effect-bound execution permits. We formalize end-to-end consequence integrity, requiring every realized external effect to be backed by a valid and current authorization witness linking the principal, task, provenance, delegation, policy state, canonical action, and finality boundary. We implement a reference verifier and deterministic cross-layer fault-injection suite covering 32 fault classes across four application domains. In 2,560 parameterized attack instances spanning 128 fault-domain classes, the full CONTINUITY configuration commits no harmful external effect, while completing all 700 benign tasks and escalating all 200 ambiguous cases. These results show that secure agent execution requires not only sound individual controls, but explicit contracts that preserve their guarantees across the complete instruction-to-effect path.

---

### Shallow neural network approximation in mixed Sobolev spaces

Authors: Yuwen Li, Guozhi Zhang
Published: 2026-09-04
Categories: math.NA, cs.LG
arXiv: https://arxiv.org/abs/2609.05263v1

We investigate the best $L_2$ approximation of mixed Sobolev spaces by shallow neural networks with $n$ neurons and general activation functions. We first establish an activation-independent Fourier-block principle: if an activation has univariate approximation order $ρ$ in the sense of the Fourier-block property, then the global approximation rate has algebraic order $\min\{α,ρ\}$ for target functions of mixed smoothness $α$, up to explicit logarithmic factors. To verify this property for concrete activations, we introduce a structured univariate approximation condition that implies the Fourier-block property with explicit parameters. For $\mathrm{ReLU}^k$, a matching algebraic lower bound identifies $\min\{α,k+1\}$ as the optimal algebraic approximation exponent in any dimension, up to logarithmic factors in the upper bound. The framework also yields the exponent $\min\{α,k+1\}$ for cardinal B-splines and soft-$\mathrm{ReLU}^k$, and the full mixed-smoothness exponent $α$ for ELU and cosine activations, again up to logarithmic~factors.

---

### Self-Supervised Lexical Representation Learning for Fast, Large-Scale Phylogenetic Inference

Authors: Tim Wientzek
Published: 2026-09-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.05262v1

Computational phylogenetics has become an essential tool in historical linguistics, yet its application at a global scale remains constrained by two factors: the labor-intensive manual annotation of cognacy judgments required for character-based methods and the substantial computational cost of inference on large datasets. This paper introduces a fully self-supervised contrastive learning framework that learns lexical representations directly from raw IPA-transcribed wordlists, without requiring cognacy annotations, alignments, or additional expert input. The model employs a dual contrastive objective: a word-level loss that organizes phonetically similar forms into a coherent space, and an auxiliary language-level loss that encourages the lexical space to reflect broader phonological properties of languages. From the resulting word representations, pairwise language distances are derived and used to infer a global phylogenetic tree of 3,399 language varieties. The inferred tree achieves a generalized quartet distance (GQD) to the Glottolog reference tree competitive with multiple baselines, while requiring only minutes of computation on a standard notebook GPU. Furthermore, the same representations capture diachronic concept stability: variance in pairwise distances across languages yields stability rankings that correlate significantly with established rankings. Ablation studies confirm that both the language-level objective and the use of phonetic feature vectors improved the inferred trees topology with regards to GQD. The framework thus provides a computationally efficient and fully automatic alternative for large-scale phylogenetic inference and offers a unified representation supporting downstream analyses at both the language and concept level.

---

### Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents

Authors: Jiazheng Sun, Boyu Yang, Binhao Yuan, Mingxuan Li, Xin Peng
Published: 2026-09-04
Categories: cs.AI, cs.SE
arXiv: https://arxiv.org/abs/2609.05261v1

Large language model agents increasingly rely on execution traces to master complex interactive tasks. However, current paradigms are bottlenecked by shallow trajectory retrieval and flat skill summarization, fundamentally ignoring the temporal dependencies and outcome-conditioned topology of agent behavior. We introduce Trace2Tower, a transition-aware EigenTrace framework that distills raw trajectories into a robust skill hierarchy. Trace2Tower abstracts step-level interactions into canonical events, constructing a unified graph governed by semantic compatibility, transition dynamics, and outcome evidence. Through a novel contrastive spectral decomposition, it isolates stable, success-aligned behavioral modes while rigorously suppressing failure-prone shortcuts. These modes organically populate a dynamic skill tower of action templates, procedural routines, and overarching task strategies, continuously refined via verifier-guided feedback. On ALFWorld, Trace2Tower achieves 87.31% success requiring only 10.35 steps and 0.26 invalid actions; on WebShop, it reaches 50.67% exact success. Across both benchmarks, Trace2Tower significantly outperforms existing baselines in task mastery and context-efficient experience reuse.

---

### Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization

Authors: Sihan Ge, Yichen Lin, Chenyu Zhou, Jianghao Lin, Tao Yao, Dongdong Ge
Published: 2026-09-04
Categories: math.OC, cs.AI
arXiv: https://arxiv.org/abs/2609.05258v2

Large language models (LLMs) are increasingly used to formulate optimization models from natural-language problem descriptions, yet realistic operations research (OR) requests are often incomplete: missing objectives, constraints, or business rules can change the resulting mathematical program. Existing evaluations largely assume a complete specification and therefore overlook whether an agent knows when clarification is needed before modeling. We introduce OR-Clarify, a benchmark for pre-formulation clarification. Each task presents a partial public problem description, withholds structured hidden slots, and evaluates agents through bounded interaction with a simulated user. The benchmark supports both openended and choice-based clarification, and measures slot recovery, stopping behavior, silent assumptions, and interaction cost. We further propose Interactive Optimization (InterOPT), a two-stage framework that identifies unresolved formulation-critical gaps and uses them to guide whether to ask the next question or to stop. In our choice-based experiments, InterOPT substantially outperforms all baselines in exact slot recovery; in the open-ended setting, it remains competitive with strong prior methods. Together, OR-Clarify and InterOPT reframe OR assistance as a selective completeness decision: clarify when needed, stop when ready, and quantify what remains missing.

---

### Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization

Authors: Sihan Ge, Yichen Lin, Chenyu Zhou, Jianghao Lin, Tao Yao, Dongdong Ge
Published: 2026-09-04
Categories: math.OC, cs.AI
arXiv: https://arxiv.org/abs/2609.05258v1

Large language models (LLMs) are increasingly used to formulate optimization models from natural-language problem descriptions, yet realistic operations research (OR) requests are often incomplete: missing objectives, constraints, or business rules can change the resulting mathematical program. Existing evaluations largely assume a complete specification and therefore overlook whether an agent knows when clarification is needed before modeling. We introduce OR-Clarify, a benchmark for pre-formulation clarification. Each task presents a partial public problem description, withholds structured hidden slots, and evaluates agents through bounded interaction with a simulated user. The benchmark supports both openended and choice-based clarification, and measures slot recovery, stopping behavior, silent assumptions, and interaction cost. We further propose Interactive Optimization (InterOPT), a two-stage framework that identifies unresolved formulation-critical gaps and uses them to guide whether to ask the next question or to stop. In our choice-based experiments, InterOPT substantially outperforms all baselines in exact slot recovery; in the open-ended setting, it remains competitive with strong prior methods. Together, OR-Clarify and InterOPT reframe OR assistance as a selective completeness decision: clarify when needed, stop when ready, and quantify what remains missing.

---

### Commonsense Reasoning in Computer Vision: Foundations, Recent Advancements, and Future Directions

Authors: Bahar Uddin Mahmud, Sumit Barua, Guan Yue Hong, Ajay Gupta, Hexu Liu
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05257v1

Commonsense reasoning in computer vision encompasses integrating visual data and contextual knowledge, crucial for enhancing AI's understanding of everyday scenarios. This understanding not only improves machine learning models but also enhances their ability to interact meaningfully with humans and the environment. Unlike CNN-based conventional vision models, which are designed to identify objects within a specific image, incorporating commonsense knowledge enables models to interpret scenes in a more holistic manner, thereby improving their spatial ability to reason about relationships among objects and actions. This integration not only enhances object recognition but also facilitates a deeper understanding of the contextual factors, ultimately leading to more precise predictions and interactions in real-world applications. This paper presents a comprehensive survey of recent developments that integrate commonsense knowledge into computer vision tasks. We systematically review approaches based on knowledge graphs, scene graphs, neuro-symbolic models, and commonsense-augmented transformers. We also outline current limitations related to dataset bias, knowledge incompleteness, and integration challenges. Finally, we highlight prospective research trajectories in cross-modal reasoning, scalable commonsense knowledge injection, and neuro-symbolic hybrid architectures to develop truly intelligent visual systems.

---

### Compact Neural Appearance Models for Efficient Gaussian Splatting

Authors: Florian Hahlbohm, Jorge Condor, Linus Franke, Martin Eisemann, Marcus Magnor
Published: 2026-09-04
Categories: cs.CV, cs.GR
arXiv: https://arxiv.org/abs/2609.05255v1

Explicit primitive-based radiance fields such as 3D Gaussian Splatting typically model view-dependent appearance using low-order spherical harmonics (SH). Although efficient to evaluate, SH coefficients dominate per-primitive storage and memory traffic, while their band-limited basis restricts angular detail. We present a thorough, end-to-end comparison of SH and recent spherical appearance models and introduce an implicit alternative that decodes compact per-primitive latent codes using a tiny shared MLP. We integrate all models into the same optimized pipeline, fusing their forward and backward passes into a differentiable CUDA rasterizer and provide a portable WebGL viewer for laptop and mobile GPUs. Our evaluation across reconstruction quality, memory use, and optimization and rendering performance shows that recent spherical models offer the strongest overall quality-efficiency trade-off. Our neural representation is the most compact model evaluated and, compared to third-degree SH, reduces the per-primitive appearance footprint from 192 to 28 bytes, accelerates optimization by 1.3$\times$, while improving reconstruction quality. We further analyze how appearance parametrization shapes optimization, identifying differences in recovered geometry and the tendency of expressive models to absorb non-static scene content. Together, our framework and analysis provide practical guidance for replacing SH beyond what image metrics alone can capture.

---

### GLASS: Graph-Language Alignment with Spherical Scoring for Transferable Graph-Level Anomaly Detection

Authors: Xudong Wang, Chris Ding, Tongxin Li, Jicong Fan
Published: 2026-09-04
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.05253v1

We introduce GLASS, a framework for graph-level anomaly detection (GLAD) that achieves robust cross-domain transferability through graph-language alignment on the unit hypersphere. GLASS builds a unified representation space by aligning a structure-aware graph encoder with an instruction-aware text embedding via a multi-slice soft cosine objective. Our framework serializes local, global, and semantic graph properties into a compact Graph Descriptor Prompt (GraphDP), creating a text bridge that enables domain-agnostic anomaly scoring. By enforcing multi-scale consistency through Matryoshka representation slices, the model captures anomalous deviations at multiple levels of granularity. For scoring, we formulate anomaly detection as density estimation on the aligned hypersphere and introduce Spherical Multi-Modal Scoring (SMS), which instantiates von Mises-Fisher kernel density estimators in both graph and text embedding spaces. This probabilistic formulation recovers angular k-nearest-neighbor scoring as a high-concentration limiting case and provides a principled fusion of structural and semantic anomaly signals. The shared text embedding space further serves as a cross-domain bridge: by encoding a target domain's GraphDP without target-domain training data, GLASS performs zero-shot anomaly detection, and with only a handful of normal examples, few-shot adaptation via reference-set calibration. Across twelve benchmarks and three meta-domains, GLASS obtains the best average AUROC and rank compared with recent advanced GLAD baselines and enables effective cross-domain transfer.

---

### A Unified Physics-Aware Quantum Machine Learning Framework across Power GaN HEMTs and Logic Nanowire FETs: Predicting Unseen Process Splits and Held-Out Geometry Combinations with Lower Error and Tighter Split-to-Split Variability

Authors: Rushat Rai, Yun-Yuan Wang, Autsada Kakaen, Pei-Jie Chang, Doan Viet Nguyen, Yuan-Chieh Chiu, Doldet Tantraviwat, Niall Tumilty, Simon See, Wen-Jay Lee, Tai-Yue Li, Nan-Yow Chen et al.
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05251v1

We present a unified reinforcement-learning (RL) framework that discovers compact parametrized quantum circuits (PQCs) for data-scarce device modeling. A graph neural network (GNN) policy optimized by proximal policy optimization (PPO) searches circuit architectures using leave-one-group-out cross-validation (LOGOCV) error on held-out process or geometry groups as the reward. The framework achieves the lowest mean absolute error (MAE) on all 11 targets versus six classical baselines, with 59% lower error (Ioff) and 81% tighter fold variability (VTH) for HEMTs and 84% lower error (VTH, SS, Ioff) and 82% tighter fold variability (Ioff) for NWFETs. These results demonstrate the potential of RL-selected, classically simulated PQCs as compact surrogates with low OOD error and improved physical consistency, despite imposing no explicit physical constraints, penalty terms, or device-specific equations, on the two evaluated device datasets.

---

### Proton Irradiation Characterization of an Open-Source ML Accelerator on a Zynq UltraScale+ MPSoC

Authors: Saad Memon, Rafal Graczyk, Jan Swakoń, Leszek Grzanka, Sebastian Kusyk, Mike Papadakis
Published: 2026-09-04
Categories: cs.AR, cs.ET, cs.LG
arXiv: https://arxiv.org/abs/2609.05249v1

As spaceborne computing systems increasingly rely on neural network (NN) accelerators, the opacity of commercial, black-box architectures severely restricts the development of verifiable radiation mitigation strategies. Open-source, register-transfer level (RTL)-accessible accelerators resolve this limitation by enabling user-defined instrumentation, yet few have empirical radiation-response baselines. This work establishes a foundational system-level proton-irradiation baseline for an unmitigated open-source Tensil NN accelerator deployed on a Zynq UltraScale+ SoC executing ResNet-20 inference. Under 20 to 58 MeV proton irradiation, we delivered $4.29 \times 10^{10}$ p/cm$^{2}$ within monitored operational windows. Seven workload interruptions required two restarts of the notebook process, four reboots or board resets, and one power-cycle sequence. Two output-corruption events returned incorrect CIFAR-10 classes without loss of service. In the longer event, the accelerator returned a class absent from the ten-image CIFAR-10 pool for 39 consecutive inputs at normal cadence. The process remained alive, while the kernel log, limited memory test, and sampled power showed no anomaly. Observation of the stuck-class sequence ended with scheduled bitstream reconfiguration. All nine onsets occurred under the nominal 4 cm beam, which exposed the SoC, LPDDR4, and additional board circuitry; none occurred under the 2 cm SoC-centered field. This pattern shows a field association but does not establish LPDDR4 as the cause because field size was confounded with run order and dose. Linux-managed accelerators require end-to-end content checks and recovery that reaches the state in which corruption can persist. This baseline documents availability loss and silent output corruption, supporting future software hardening of COTS FPGA-SoCs for neural-network inference in space systems.

---

### Do LLMs Exhibit Coherent Knowledge Structures in Mathematical Reasoning? A Perspective from Knowledge Space Theory

Authors: Peng Cui, Heejin Do, Mrinmaya Sachan
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05245v1

Human knowledge is inherently structured and interdependent: mastery of a concept requires prior mastery of its prerequisites, a principle formalized by Knowledge Space Theory (KST). While LLMs achieve strong performance on complex reasoning tasks, it remains unclear whether they exhibit coherent, human-like knowledge structure. We introduce a KST-grounded framework for evaluating LLM knowledge structure in mathematical reasoning, using it as a normative framework to analyze whether LLM behavior adheres to principled knowledge dependencies. Evaluating eight open- and closed-source LLMs against real human learners, we find that (1) LLMs do not adhere to human knowledge structure -- they frequently violate knowledge dependencies and fail to leverage related knowledge provided in context to improve performance on dependent questions; (2) LLMs do not share a consistent knowledge structure among themselves, as reflected by low overlap in their knowledge distributions. Furthermore, these structural deficiencies remain largely invisible to accuracy-based and LLM-as-judge evaluations. Together, our results provide behavioral evidence that current LLMs knowledge does not follow a human-like structure.

---

### Few-Shot Video Recognition via Hierarchical Metric Learning

Authors: Jiaxin Zhang, Haoran Gao, Xizhan Gao, Zihao Dong, Tingwei Wang, Sijie Niu
Published: 2026-09-04
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.05242v1

Few-shot action recognition (FSAR) aims to recognize unseen action categories with only a small number of annotated video samples. Recent works typically apply single-prototype supervision at the network output and fail to sufficiently exploit rich cross-frame global spatial information in videos. Even existing multi-level metric schemes only impose parallel prototype constraints on intermediate layers, without progressive supervision along the full feature pipeline, which results in limited generalization ability of the learned class prototypes. Inspired by this, we present a novel method, hierarchical metric learning for few-shot action recognition (HML-FSAR). First, a spatial-enhanced module is developed to capture cross-frame global spatial representations. Combined with temporal MHA, heterogeneous alignment, spatial-temporal feature fusion and dictionary learning modules, it constructs the complete feature processing pipeline. Second, a hierarchical metric learning (HML) strategy is embedded into HML-FSAR. Composed of center metric, alignment metric, contrastive metric, dictionary metric and prototype metric, HML imposes progressive multi-stage complementary constraints from frame-level representations to final class prototypes, so as to jointly optimize feature compactness, heterogeneous spatial-temporal alignment, inter-class discriminability and anti-noise robustness. The proposed HML-FSAR method is validated on five widely-used FSAR datasets, and experimental results fully demonstrate its effectiveness.

---

### Uncensored Open-weight Models: Redistribution as the Persistence Layer

Authors: 10a Labs, :, Juliette Garcia, Hailey May, Bobby McKenzie, David Pham, Matthew Swain, Joshua Valdez, Corie Wieland, Zachary Yahn
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05241v1

A rapidly expanding ecosystem of actors is removing built-in safety guardrails from open-weight AI models. We profile this ecosystem by identifying key producers, downstream reproductions, and emerging applications. Between January 2024 and March 2026, we identified 3,471 original uncensored models on HuggingFace, each repackaged an average of 2.4 times; three actors account for 52% of all 8,164 compressed redistributions. Once quantized and mirrored across separate accounts, formats, and registries such as Ollama, these models persist regardless of upstream removal and become easier to deploy downstream. Of the 1,643 identified GitHub applications integrating uncensored large language models (ULLMs), 25% were classified as explicitly malicious.

---

### Cross-Domain Tracker Adaptation Without Target-Domain Labels via Vision-Language Agents

Authors: Daniel Davila, Ravikumar Balakrishnan, Mike Cochran
Published: 2026-09-04
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.05239v1

We present a system that uses a Vision-Language Model (VLM) as a diagnostic agent for adapting a detect-to-track pipeline to a new target domain without access to target-domain labels. Rather than optimizing against annotated metrics, the VLM directly inspects rendered tracking outputs, identifies visual failure modes, and recommends parameter updates through an iterative tuning loop. We first demonstrate that ground-truth-supervised hyperparameter transfer can be brittle. On MOT17->MOT20, applying a source-derived oracle configuration reduces mean HOTA by 0.090, from a target-domain ceiling of 0.357, to 0.267. Without using any target-domain labels, our VLM-based tuner recovers 67.8% of this lost headroom, finishing within 0.029 HOTA of the target ceiling; on the highest-density target sequence, it recovers up to 86.7%. We further show that label-free Bayesian optimization with handcrafted proxy objectives struggles under large domain shifts and can degrade configurations that are already strong. In contrast, the VLM tuner acts selectively: when its visual diagnosis reveals no clear failure mode, it declines to modify the configuration, preserving performance on easy transfers while improving hard ones. Finally, we characterize the conditions under which this approach succeeds, namely, when domain shift manifests through exposed detection-level parameters, versus where it is less effective, such as MOT17->DanceTrack, where the source oracle is already near-optimal.

---

### PRICE: A Systematic Study of LLM Adaptation Choices for Bitcoin Price Forecasting

Authors: Maryam Fakhari, Mehran Safayani
Published: 2026-09-04
Categories: cs.LG, cs.AI, math.NA
arXiv: https://arxiv.org/abs/2609.05235v1

Cryptocurrency markets exhibit extreme volatility and non-stationary dynamics that challenge conventional forecasting methods. Although Large Language Models (LLMs) have shown promise for time series forecasting, the combined effects of adaptation choices remain largely unexplored in financial settings. This study introduces PRICE, a structured approach for adapting LLMs to short-term Bitcoin price forecasting. Built on a 4-bit quantized LLaMA-3 8B model, PRICE investigates how fine-tuning, numerical representation, prompting, inference, and decoding jointly influence forecasting performance. PRICE integrates Parameter-efficient fine-tuning with Low-Rank Adaptation (LoRA), Recursive multi-step inference, Integer-rounded numerical representation, Context-Task-Format (CTF) prompting, and Exact zero-temperature decoding. Ablation studies show that each component contributes to forecasting accuracy and reliability. LoRA enables efficient training on limited hardware, recursive inference improves accuracy, integer-rounded values reduce errors, CTF prompting outperforms Chain-of-Thought, Implicit Chain-of-Thought (iCoT), and few-shot prompting, and zero-temperature decoding improves stability during recursive forecasting. Comparative evaluation against eight transformer-based and time-series foundation models shows that PRICE achieves the lowest forecasting errors on both validation and test sets while maintaining robust performance across evaluation periods. Despite being based on a model primarily pretrained on text rather than time-series data, PRICE achieves competitive or superior performance relative to specialized foundation models. These findings demonstrate that adaptation choices critically determine the accuracy and robustness of LLMs for numerical time-series forecasting.

---

### Measured Sliders: Learning Continuous Controls from Differentiable Image Measurements

Authors: Yijia Chen, Boyu Wei, Xuanhua Yin
Published: 2026-09-04
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.05234v1

Continuous sliders are useful only when coefficient changes produce predictable image changes. Yet most diffusion sliders derive their axes from text or learned representations, leaving their scales disconnected from observable image properties. Consequently, we cannot tell in advance which attributes are learnable, compare control strengths directly, or anticipate interference when multiple controls are combined. We propose Measured Sliders, a framework that defines continuous controls through closed-form differentiable image measurements. A common measurement space unifies the pipeline. Before training, an observability test identifies usable supervision. During training, a measurement-guided objective learns target movement while suppressing non-target changes. After training, decoded calibration expresses controls in comparable units of realized image change. Multiple LoRA branches are stored in one checkpoint and composed without training on joint activations. Across SDXL and FLUX.1-dev, the resulting controls are ordered, selective, and composable. On 553 prompts, lighting direction reaches rho = 0.995 and 98.9% monotone sweeps. A five-attribute checkpoint achieves average selectivity 2.59, compared with 1.50 for the strongest baseline, and preserves every requested direction in 96.7% of pair and 86.1% of triple compositions. The observability test also separates every subsequently successful measurement from the failed candidate. Overall, image-space measurement provides a common basis for learning, diagnosing, calibrating, and composing continuous generative controls.

---

### Hessian-based molecular conformation augmentation for a scalable and efficient strategy of machine learning interatomic potentials

Authors: Bumju Kwak, Jeonghee Jo
Published: 2026-09-04
Categories: cs.LG, physics.chem-ph
arXiv: https://arxiv.org/abs/2609.05233v2

While machine-learning interatomic potentials (MLIPs) have successfully learned potential energy surfaces (PES) and atomic forces, many practical applications, such as vibrational analysis and transition state search, rely heavily on the PES Hessian. Yet standard MLIPs are trained on energy and forces alone, and existing methods that incorporate the Hessian into training objectives require architectural modifications and incur significant computational and memory overheads from higher-order backpropagation. To address these limitations, we propose two Hessian-derived data augmentation schemes: isotropic Gaussian displacement (\textbf{UniAug}) and normal mode-weighted displacement (\textbf{ModeAug}). Both methods utilize simple Taylor expansions, achieving effective augmentation without altering training objectives or extending the autograd graph. This allows seamless, plug-and-play integration with existing architectures and training pipelines. Comprehensive evaluations across non-equilibrium and equilibrium datasets demonstrate that our approach enhances model accuracy where reference forces are large while providing practical, task-specific guidelines.

---

### Hessian-based molecular conformation augmentation for a scalable and efficient strategy of machine learning interatomic potentials

Authors: Bumju Kwak, Jeonghee Jo
Published: 2026-09-04
Categories: cs.LG, physics.chem-ph
arXiv: https://arxiv.org/abs/2609.05233v1

While machine-learning interatomic potentials (MLIPs) have successfully learned potential energy surfaces (PES) and atomic forces, many practical applications, such as vibrational analysis and transition state search, rely heavily on the PES Hessian. Yet, standard MLIPs tend to be trained on energy and forces alone, leaving Hessian information largely unexploited. Meanwhile, existing methods that explicitly incorporate the Hessian into training objectives require architectural modifications and introduce significant computational and memory overheads due to higher-order backpropagation. To address these limitations, we propose two Hessian-derived data augmentation schemes: isotropic Gaussian displacement (\textbf{UniAug}) and normal mode-weighted displacement (\textbf{ModeAug}). Both methods utilize simple Taylor expansions, achieving effective augmentation without altering training objectives or extending the autograd graph. This allows seamless, plug-and-play integration with existing architectures and training pipelines. Comprehensive evaluations across non-equilibrium and equilibrium datasets demonstrate that our approach enhances model accuracy while providing practical, task-specific guidelines.

---

### Substrate-Aware AI Agents: Execution Context as a First-Class Input

Authors: Manu Agrawal
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05232v1

Autonomous AI agents increasingly select actions in environments whose memory, execution-time, runtime, compute, and operational constraints determine what counts as a suitable plan. We call the absence of this execution context from an agent's planning state substrate blindness. We test this general proposition through numerical code generation, where selected implementation choices and operational consequences are directly observable. Three frontier model configurations--Anthropic Claude Opus 5, OpenAI GPT-5.6-Sol, and Google Gemini 3.7 Flash--generate code for a high-dimensional pairwise Euclidean-distance task either from the task alone or with a 128 MB RAM and 10.0 s wall-time contract. Contract disclosure reduced measured peak process memory in 13 of 14 executable index-aligned task-only versus contract-disclosed comparisons and reduced mean wall time in all three cohorts, making execution up to 3.1x faster. Across the audited corpus, disclosure produced structural code changes including bounded blocking, float32 retention, upper-triangle traversal, and in-place or memory-mapped buffers. At a tighter 96 MB contract, independently sampled contract-disclosed cohorts achieved correct-and-within-budget outcomes of 4/5 for Claude Opus 5, 5/5 for GPT-5.6-Sol, and 3/5 for Gemini 3.7 Flash, compared with task-only outcomes of 0/5, 1/5, and 0/5; cohort mean MaxRSS and wall time were 49-74% and 35-64% lower than their task-only references. These results establish a controlled proof of concept for substrate-aware agent planning: a minimal execution contract induces proactive structural adaptation in generated programs, shifting computation away from unconstrained allocations and substantially improving observed resource-time profiles before execution.

---

### ACE: Adaptive Calibration-Free Expert Skipping for MoE-based LLMs

Authors: Zukang Xu, Zhixiong Zhao, Xing Hu, Jiangyong Yu, Houji Wen, Jun Li, Zhe Jiang, Dawei Yang
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05228v1

Mixture-of-Experts (MoE) architectures provide an efficient paradigm for scaling large language models (LLMs), yet fixed top-k routing activates the same number of expert slots for every token, causing substantial redundant computation. Existing expert-skipping methods often rely on router confidence, calibration data, or additional training, and therefore cannot reliably estimate the actual contribution of routed experts. To this end, we propose ACE, a training-free, calibration-free, and checkpoint-preserving framework for token-adaptive expert skipping in MoE-based LLMs. ACE contains two complementary components: 1) Global Spectral Proxy (GSP), which estimates global transformation capacity from the coupled gate, up, and down projections together with RMSNorm scaling; and 2) Router-Conditioned Refinement (RCR), which constructs expert-specific direction prototypes from centered router weights and evaluates expert responses along routing-preferred directions. During inference, ACE combines both estimates with runtime router gates and skips an expert slot only when both views identify it as low-contribution, while always retaining the top-1 expert. All expert statistics are computed offline, leaving only table lookups and lightweight scalar operations online. Extensive experiments across three MoE-based LLMs and eight benchmarks demonstrate that ACE consistently outperforms existing static and dynamic baselines, with increasingly pronounced advantages under aggressive expert skipping. For instance, at a 50% skipping ratio on Qwen3.6-35B-A3B, ACE reduces WikiText-2 perplexity by 7.96% and improves average downstream accuracy by 4.15 percentage points over the strongest competing method.

---

### CABAL: Multi-Agent Simulacra for Tracing the Effects of Collusive Bidding in Peer Review

Authors: Jicheng Zhou, Kemou Li, Kahim Wong, Zheyuan Li, Zhuan Shi, Fengpeng Li, Haiwei Wu, Jiantao Zhou
Published: 2026-09-04
Categories: cs.AI, cs.CY
arXiv: https://arxiv.org/abs/2609.05227v1

Recent reports during the AAAI-27 review cycle highlight the risk of reviewers coordinating bids for reciprocal assignment advantage. Prior work treats bidding, reviewer assignment, and review manipulation as separate stages, leaving the lifecycle effects of collusive bidding unclear. Real-world analysis is further constrained by typically unobservable collusive intent and the lack of counterfactuals for the same conference. Motivated by this gap, we introduce \alg, an end-to-end multi-agent simulacra framework for studying reviewer assignment integrity by holding the conference environment fixed and configuring LLM-driven reviewer agents with honest or collusive policies. We further develop an affinity-guided collusive bidding strategy that uses mutual reviewer-paper affinities to construct collusion rings and select target papers, producing expertise-consistent rather than arbitrarily targeted attacks. Controlled experiments show that collusive bidding more than doubles target-paper capture and that assigned colluders score target papers about two points higher than honest co-reviewers, while conference-wide effects remain comparatively modest. Evaluated bid-phase detectors provide only limited evidence of collusion: in a fixed-triplet detector stress test, native positive-bid graphs are confounded by benign affinity, while a Very-High-only diagnostic view enables precise but low-coverage local recovery.

---

### First Things First: Teaching LLM-Based Agents to Prioritize Must-Haves before Nice-to-Haves

Authors: Tianjie Ju, Xinyue Xu, Wanxuan Sun, Lingxiao Diao, Gongshen Liu, Zhuosheng Zhang, Cheng Yang
Published: 2026-09-04
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.05224v2

Recent progress in multimodal large language models (MLLMs) has fueled significant enthusiasm in their potential to act as autonomous agents for real-world tasks. However, scenarios requiring agents to fulfill users' complex, structured requirements remain largely underexplored. In this work, we examine reasoning tasks under three distinct requirement scenarios: (i) Must-have requirements uniquely determine a unique feasible solution; (ii) Multiple answers satisfy the must-have requirements and are prioritized via the nice-to-have requirements; and (iii) No candidate solution satisfies the must-have requirements, in which case the agent should abstain from generating a response. We evaluate state-of-the-art MLLMs on 3,649 carefully constructed problems that reflect realistic service scenarios, including e-commerce, booking, and map-based or ride-hailing. Our evaluation reveals that existing MLLMs exhibit catastrophic failures in all scenarios. They frequently misinterpret task requirements, violate must-have requirements, and produce invalid solutions. To address this critical gap, we propose First Things First Reinforcement Learning FTF-rl that explicitly optimizes reasoning over multi-priority user requirements. Experimental results show that our method substantially improves the task success rate compared to strong baselines. Moreover, FTF-rl yields general effectiveness on popular logical and mathematical reasoning tasks, including LogicVista, MathVision, and InfoQA. Our findings suggest that enhancing requirement-aware reasoning capability provides a simple yet effective pathway to improve generalization of MLLM agents. Code and dataset are available at https://github.com/claire62/FTF-RL.

---

### First Things First: Teaching LLM-Based Agents to Prioritize Must-Haves before Nice-to-Haves

Authors: Tianjie Ju, Xinyue Xu, Wanxuan Sun, Lingxiao Diao, Gongshen Liu, Zhuosheng Zhang, Cheng Yang
Published: 2026-09-04
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.05224v1

Recent progress in multimodal large language models (MLLMs) has fueled significant enthusiasm in their potential to act as autonomous agents for real-world tasks. However, scenarios requiring agents to fulfill users' complex, structured requirements remain largely underexplored. In this work, we examine reasoning tasks under three distinct requirement scenarios: (i) Must-have requirements uniquely determine a unique feasible solution; (ii) Multiple answers satisfy the must-have requirements and are prioritized via the nice-to-have requirements; and (iii) No candidate solution satisfies the must-have requirements, in which case the agent should abstain from generating a response. We evaluate state-of-the-art MLLMs on 3,649 carefully constructed problems that reflect realistic service scenarios, including e-commerce, booking, and map-based or ride-hailing. Our evaluation reveals that existing MLLMs exhibit catastrophic failures in all scenarios. They frequently misinterpret task requirements, violate must-have requirements, and produce invalid solutions. To address this critical gap, we propose First Things First Reinforcement Learning FTF-rl that explicitly optimizes reasoning over multi-priority user requirements. Experimental results show that our method substantially improves the task success rate compared to strong baselines. Moreover, FTF-rl yields general effectiveness on popular logical and mathematical reasoning tasks, including LogicVista, MathVision, and InfoQA. Our findings suggest that enhancing requirement-aware reasoning capability provides a simple yet effective pathway to improve generalization of MLLM agents. Code and dataset are available at https://github.com/claire62/FTF-RL.

---

### FedDRAW: Federated Dual Reputation Annealing Weighting for Heterogeneous Multi-Institutional Chest Radiograph Classification

Authors: Maryam Moradpour, Anne-Christin Hauschild
Published: 2026-09-04
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.05223v1

Artificial intelligence models are promising for medical diagnosis, but they require large numbers of unbiased data, which in medicine are distributed across hospitals and cannot be centralized to protect patient privacy. Federated Learning (FL) addresses this, since hospitals train one shared diagnostic model while patient data remain local. Training proceeds in communication rounds, in which each hospital trains the shared model locally and returns it to the server for merging by weighted average. This aggregation weight determines whose institutional knowledge shapes the result. Federated averaging (FedAvg) sets it in proportion to local sample count, so a small but informative hospital is permanently assigned a small influence, andl argest clients could dominate the global model even when they are less informative. We propose Federated Dual Reputation Annealing Weighting (FedDRAW), a server-side aggregation method that combines a data-size prior with the cosine similarity between client and global parameters under two coupled annealing schedules. An inner schedule shifts client reputation from the size prior towards similarity. An outer, deferred annealing schedule on the softmax inverse temperature keeps the weighting selective in the early and middle rounds and relaxes it to uniformity at convergence. We evaluate FedDRAW on 12 simulated client-partition scenarios of two chest radiograph datasets (CheXpert and ChestMNIST), against seven federated baselines under identical local training settings. FedDRAW achieved the highest average rank among all eight methods under both AUC and the geometric mean (GM) of sensitivity and specificity, which a Friedman test with Nemenyi post-hoc analysis confirmed to be a statistically significant difference between the methods. Scheduling two signals, rather than fixing the weights by sample count alone, could enable less biased diagnostic models.

---

### A Verifier-Guided Explainable Reasoning Framework with Gold-Anchored QLoRA, Task-Aware Mixture-of-Experts, and Group-Relative RLVR

Authors: Thi Kim Trang Vo, Nam Tien Le, Thi Kim Nguyet Vo, Minh Khang Tran, Duy Phuong Tran
Published: 2026-09-04
Categories: cs.CL, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2609.05221v1

Large language models (LLMs) show strong reasoning ability, but their explanations can remain inconsistent, weakly grounded, or difficult to verify. We propose a verifier-guided explainable reasoning framework for transparent educational question answering that combines gold-anchored QLoRA, task-aware symbolic routing, and group-relative RLVR. Qwen2.5-3B-Instruct is first adapted with field-weighted QLoRA supervision anchored to authoritative answers. A lightweight router then assigns logic problems to a FOL/Z3 verifier and physics problems to a formula- and unit aware symbolic solver. Verifier feedback is further used to support candidate evaluation, self-revision, and reward construction during RLVR. Candidate responses are evaluated along three complementary dimensions: P1 for answer correctness, P2 for evidence or unit consistency, and P3 for reasoning depth and explainability. At inference, gold-free self-consistency aggregates multiple candidate responses before an optional question-only physics verifier performs conservative system-level correction. On 438 held-out examples, RLVR increases P3 from 50.68% to 72.20%, while hybrid P1 remains approximately stable at 55.94%. Self-consistency improves model only P1 from 48.86% to 50.23%, with symbolic verification providing the remaining hybrid gain. These results indicate that RLVR primarily strengthens explicit reasoning structure, while symbolic verification complements the neural policy by improving answer reliability at the system level.

---

### Dimension-Adaptive Batched Lipschitz Narrowing Without Knowing the Zooming Dimension

Authors: Yasong Feng
Published: 2026-09-04
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.05214v1

The Appropriately Combined Edge-length (ACE) sequence in A-BLiN depends on the zooming dimension $d_z$. This note removes that dependence. The next edge length is selected from the number of cubes that survive the preceding elimination. The resulting Count-Adaptive BLiN algorithm does not use $d_z$ or the zooming constant $C_z$, yet it attains $\widetilde{\mathcal O}_d(T^{(d_z+1)/(d_z+2)})$ regret with $\mathcal O_d(\log\log T)$ batches. Together with the adaptive-grid lower bound in Theorem 10 of the original paper, the optimal batch complexity remains $Θ_d(\log\log T)$ when $d_z$ is unknown.

---

### PAC-Bayesian Reconstruction Guarantees for Time Series Variational Autoencoders

Authors: Chloé Hashimoto-Cullen, Ghislain Agoua, Benjamin Guedj, Sylvain Le Corff
Published: 2026-09-04
Categories: stat.ML, cs.LG
arXiv: https://arxiv.org/abs/2609.05212v1

Forecasting time series accurately is critical for applications with complex data ranging from energy systems to healthcare and finance. Among current state of the art models, generative latent variable models are increasingly implemented; yet principled generalisation guarantees for modern latent variable models remain limited. In particular, while Variational AutoEncoders are widely used for sequential data, their theoretical analysis is largely restricted to i.i.d. settings. In this work, we develop a PAC-Bayesian framework for latent variables models applied to time series. Building on reconstruction-based bounds, we extend PAC-Bayesian guarantees to Markovian latent structures, capturing temporal dependencies through a sequential generative process. These guarantees do not grow with the length of the trajectory. Our bounds depend on assumptions which are common in the literature; we provide an example framework where they would be verified to show that they are not as restrictive as they may seem.

---

### BLASt3R: Bundle Adjustment of Any Image Set with Multi-View Matching and Monocular Priors

Authors: Vincent Leroy, Philippe Weinzaepfel, Lojze Zust, Yohann Cabon, Jérome Revaud
Published: 2026-09-04
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.05210v1

Recent hybrid Structure-from-Motion (SfM) systems combine the robustness of feed-forward 3D reconstruction with the accuracy of traditional bundle adjustment (BA) with pixel matching. They are usually the best performing methods however their scalability and usability remains limited since estimating dense correspondences between views is prohibitively costly, especially considering time constraints inherent to online applications like Visual SLAM (VSLAM). In this paper, we introduce a regularized BA framework that leverages a fast multi-view matcher and monocular priors for initialization and regularization. In contrast to existing systems, our unified approach seamlessly supports both online VSLAM and offline reconstruction from unordered image collections within the same optimization framework and sharing common hyperparameters for all tasks. Extensive experiments across both domains demonstrate improved performance and speed tradeoffs over traditional, feed-forward, and hybrid baselines. Notably for VSLAM, our uncalibrated method outperforms all previous calibrated approaches.

---

### FluxDisco: Symbolic Regression for Stoichiometric Dynamical Systems via Monte Carlo Graph Search

Authors: Cassandra Durr, Alvaro Köhn-Luque, Chris Jewell, Lloyd A. C. Chapman
Published: 2026-09-04
Categories: stat.ML, cs.LG, physics.data-an
arXiv: https://arxiv.org/abs/2609.05207v1

Dynamical symbolic regression methods identify governing differential equations from noisy data, balancing interpretability and predictive accuracy. However, standard methods often produce expressions that violate known physical laws. To address this, we propose FluxDisco, a physics-informed framework tailored for flux-based, stoichiometric ODE systems. By leveraging a known stoichiometry, we reduce the expression search space and ensure physical adherence. Our framework adapts the Monte Carlo Graph Search algorithm for the unique challenges associated with joint flux discovery of stoichiometric systems. We evaluate our method across a range of physical and biological systems, demonstrating its ability to accurately recover governing dynamics through interpretable equations.

---

### Real-World Multi-Modal and Longitudinal Lung Cancer Dataset

Authors: Rita Cordeiro Mendes, Maria Rita Fonseca Verdelho, Carlos Santiago, Catarina Barata
Published: 2026-09-04
Categories: eess.IV, cs.CV
arXiv: https://arxiv.org/abs/2609.05202v1

Multi-modal learning has demonstrated strong potential in medical applications by integrating heterogeneous data sources such as medical imaging, clinical records, and genomics to improve predictive performance and support clinical decision-making. However, advances in this area are often constrained by two key challenges: the limited availability of well-curated, ready-to-use datasets that accurately reflect real-world conditions, where medical data are frequently collected inconsistently and are often incomplete; and the inherent difficulty of integrating heterogeneous data modalities. In this work, we introduce a newly curated multi-center, multi-modal, and longitudinal dataset designed to support the evaluation of a wide range of learning pipelines under realistic conditions. The dataset comprises a total of 1,365 lung cancer patients and has three imaging modalities (whole-slide images, CT scans, and PET scans), structured clinical data, transcriptomic, and longitudinal follow-up and treatment information. For each imaging modality the dataset contains more than one instance. Moreover, the dataset exhibits substantial and non-uniform missingness across modalities, making it well-suited for studying robust multi-modal fusion strategies. We further provide both uni-modal and multi-modal benchmarks on the task of 12-month overall survival prediction, disease-specific survival, as well as longitudinal benchmark of hazard prediction under severe missing data. Our results show that, despite high levels of missingness, integrating complementary modalities consistently improves predictive performance over uni-modal approaches, highlighting the value of multi-modal fusion in realistic clinical settings. The dataset and benchmark code are available at https://github.com/ritacmendes/MMIST-LUNG.

---

### What Matters in On-Policy Distillation? A Perspective on Data Efficiency and Data Selection

Authors: Zhinan Hou, Jiaqi Zhang, Xunliang Cai, Keyou You
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05198v2

On-Policy Distillation (OPD) has emerged as a widely adopted post-training paradigm for enhancing large language models in reasoning domains. However, the data-centric mechanisms in OPD remain relatively underexplored. This paper presents an empirical study of data efficiency and data selection in OPD. We begin by investigating an extreme setting: training OPD on only one example, namely 1-shot OPD. Surprisingly, we find that 1-shot OPD is consistently effective across all sampled training examples and harder examples often yield superior performance gain. We next investigate what actually drives the student model's improvement in the training data. Our analysis reveals that the improvement is not driven by high token entropy, but the longer CoT paths which hard problem naturally generate. Training on longer CoT can help maintain closer alignment with the teacher over a long reasoning horizon, and learn critical thinking patterns usually missing in short CoTs, such as reflection (e.g., "Alternatively"). Based on these insights, we propose a simple data selection method that selects only hard examples for training, where even "unsolvable" examples that completely exceed the teacher's capability can be successfully used. Our experiments conducted on five models ranging from 1.5B to 8B show that training the student model on only 8 selected hard examples matches the performance of the 17K dataset baseline. Code and models will be publicly available.

---

### What Matters in On-Policy Distillation? A Perspective on Data Efficiency and Data Selection

Authors: Zhinan Hou, Jiaqi Zhang, Xunliang Cai, Keyou You
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05198v1

On-Policy Distillation (OPD) has emerged as a widely adopted post-training paradigm for enhancing large language models in reasoning domains. However, the data-centric mechanisms in OPD remain relatively underexplored. This paper presents a empirical study of data efficiency and data selection in OPD. We begin by investigating an extreme setting: training OPD on only one example, namely 1-shot OPD. Surprisingly, we find that 1-shot OPD is consistently effective across all sampled training examples and harder examples often yield superior performance gain. We next investigate what actually drives the student model's improvement in the training data. Our analysis reveals that the improvement is not driven by high token entropy, but the longer CoT paths which hard problems naturally generate. Training on longer CoT can help maintain closer alignment with the teacher over a long reasoning horizon, and learn critical thinking patterns usually missing in short CoTs, such as reflection (e.g., ``Alternatively''). Based on these insights, we propose a simple data selection method that selects only hard examples for training, where even ``unsolvable'' examples that completely exceed the teacher's capability can be successfully used. Our experiments conducted on four models ranging from 1.5B to 7B show that training the student model on only 8 selected hard examples matches the performance of the 17K dataset baseline.

---

### Phase Transition Frequency as a Training Time Predictor of Test Accuracy in ResNets

Authors: Arunan J
Published: 2026-09-04
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.05194v1

The number of discrete class-separability jumps observed during ResNet finetuning is examined empirically as a predictor of final test accuracy. Across 75 experiments spanning four benchmarks (CIFAR-10, CIFAR-100, TinyImageNet, and CIFAR-10-C) and three architectures (ResNet-18, ResNet-50, and ResNet-101), with five to ten seeds per configuration, a strong within-dataset negative correlation is obtained on standard i.i.d. classification benchmarks: \(r = -0.84\) on CIFAR-10 (\(p < 10^{-8}\), \(n = 30\)) and \(r = -0.87\) on CIFAR-100 (\(p < 10^{-5}\), \(n = 15\)). Under distributional stress, the relationship attenuates: TinyImageNet yields \(r = -0.45\), and the CIFAR-10-C corruption benchmark yields \(r = -0.19\). Two additional analyses discipline the empirical claim. A partial correlation controlling for architecture depth, treated as a linear covariate, shows that on CIFAR-100 the transition count retains statistically significant predictive power (\(r_{\mathrm{partial}} = -0.69\), \(p = 0.007\)); the corresponding result under the stricter categorical conditioning is not established at \(n = 15\). A comparison against six alternative training-curve signals shows that transition count achieved the strongest correlation among the evaluated signals on CIFAR-100 and one of the strongest on CIFAR-10, but is dominated by other signals on the two stressed benchmarks. The comparison is restricted to training-curve-level signals; comparisons against effective rank, Hessian sharpness, Fisher information, margin, and neural-collapse measures, which are the strongest competitors in the current literature, are not part of the present study and remain open. The observation is presented as an in-distribution training-quality probe among a family of candidate probes, and an inexpensive detection procedure suitable for logging alongside a standard training loop is provided.

---

### The Mirror Agent Model: a Bayesian Architecture for Interpretable Agent Behavior

Authors: Michele Persiani, Thomas Hellström
Published: 2026-09-04
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.05190v1

In this paper we illustrate a novel architecture generating interpretable behavior and explanations. We refer to this architecture as the Mirror Agent Model because it defines the observer model, that is the target of explicit and implicit communications, as a mirror of the agent's. With the goal of providing a general understanding of this work, we firstly show prior relevant results addressing the informative communication of agents intentions and the production of legible behavior. In the second part of the paper we furnish the architecture with novel capabilities for explanations through off-the-shelf saliency methods, followed by preliminary qualitative results.

---

### Can Large Language Models Anticipate Behavioral Responses to Social Policies? A Case of Pension Enrollment Prediction among China's Flexible Workers

Authors: Yumiao Li, Peixin Liu, Donglin Di, Chen Li, Runhuan Feng
Published: 2026-09-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.05189v2

Assessing the impacts of social policy changes is a widely acknowledged challenge for policymakers. Econometric methods can be unreliable when extrapolating to hypothetical scenarios, while field pilot programs are highly costly. In this paper, we propose using large language models (LLMs) as policy-assessment tools adapted from general-purpose models. We present FlexPension-LLM, the first domain-specialized large language model for a hierarchical pension-enrollment prediction task among flexible workers in China, and introduce DKI-RDistill, which injects policy-grounded cues into the prompt, including Probit-derived marginal effects and hukou-province pension rules. The method then uses LoRA/SFT to distill rationale-augmented supervision into an open-weight MoE student, with teacher errors corrected by regenerating those cases under ground-truth labels. On a CHFS 2019 blind split, FlexPension-LLM achieves 0.9316 Composite F1, surpassing its Claude Sonnet 4.5 teacher and 15 of 17 baselines, and is statistically indistinguishable from Claude Opus 4.6. Across four external surveys, it averages 0.7549 Composite F1 and shows the narrowest performance range among the strongest systems. Component analysis shows that gains come mainly from policy-grounded cue injection and error-filtered supervision, while rationales provide decision traces that can be checked against policy rules.

---

### Can Large Language Models Anticipate Behavioral Responses to Social Policies? A Case of Pension Enrollment Prediction among China's Flexible Workers

Authors: Yumiao Li, Peixin Liu, Donglin Di, Chen Li, Runhuan Feng
Published: 2026-09-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.05189v1

Assessing the impacts of social policy changes is a widely acknowledged challenge for policymakers. Econometric methods can be unreliable when extrapolating to hypothetical scenarios, while field pilot programs are highly costly. In this paper, we propose using large language models (LLMs) as policy-assessment tools adapted from general-purpose models. We present FlexPension-LLM, the first domain-specialized large language model for a hierarchical pension-enrollment prediction task among flexible workers in China, and introduce DKI-RDistill, which injects policy-grounded cues into the prompt, including Probit-derived marginal effects and hukou-province pension rules. The method then uses LoRA/SFT to distill rationale-augmented supervision into an open-weight MoE student, with teacher errors corrected by regenerating those cases under ground-truth labels. On a CHFS 2019 blind split, FlexPension-LLM achieves 0.9316 Composite F1, surpassing its Claude Sonnet 4.5 teacher and 15 of 17 baselines, and is statistically indistinguishable from Claude Opus 4.6. Across four external surveys, it averages 0.7549 Composite F1 and shows the narrowest performance range among the strongest systems. Component analysis shows that gains come mainly from policy-grounded cue injection and error-filtered supervision, while rationales provide decision traces that can be checked against policy rules.

---

### Conserved Immune Topology Improves Pathology Foundation Model Generalization for Cross-Cancer MSI-H Prediction

Authors: Dasari Naga Raju
Published: 2026-09-04
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.05182v1

Pathology foundation models integrated with multiple instance learning achieve competitive accuracy within single-cancer cohorts, yet cross-cancer generalization remains unresolved due to organ-specific histological and architectural differences. In this paper, we propose Conserved Immune Topology (CIT), a lightweight spatial representation for cross-cancer MSI-H prediction that augments foundation-model embeddings with biologically motivated immune descriptors. CIT uses unsupervised clustering to identify immune-associated tiles, then encodes tertiary lymphoid structures, peritumoral immune reactions, multi-scale tumor-infiltrating lymphocyte density, and immune-tumor mixing from frozen foundation-model embeddings and tile coordinates without requiring annotations or target-domain data. The proposed method was evaluated under cross-site and cross-cancer settings using CPTAC-COAD and TCGA-STAD cohorts, which introduce scanner variability, distribution shifts, and organ-specific architectural variations. Zero-shot cross-cancer transfer with CIT increased TransMIL AUC from 0.6627 to 0.7161, an absolute gain of 0.0534 (p=0.003), with consistent improvements across all three MIL aggregators. These results suggest that spatial immune topology provides potentially an organ-invariant representation for MSI-H prediction, supporting cross-cancer generalization of pathology foundation models.

---

### Measuring the Novelty of Biomedical Papers Using the Latent Distances between Knowledge Units

Authors: Yi Zhao, Heng Zhang, Yuzhuo Wang, Wenqing Wu, Tong Bao, Chengzhi Zhang
Published: 2026-09-04
Categories: cs.DL, cs.CL
arXiv: https://arxiv.org/abs/2609.05175v1

Measuring the novelty of scientific papers is a central concern in research evaluation and scientometrics. From a recombination perspective, prior studies have largely focused on the co-occurrence of knowledge units to assess the novelty of scientific papers. However, these studies often overlook other relationships between knowledge units. This narrow view may result in inaccurate or incomplete evaluations of novelty for scientific papers. To fill this gap, this study introduces a comprehensive novelty measurement that incorporates three types of relationships between knowledge units: network, semantic, and hierarchical. These relationships are used to quantify the latent distances among knowledge units. Using a dataset of 142,036 articles published in PLoS ONE and a validation dataset from the H1 Connect platform, our results demonstrate that (1) each relationship type captures distinct latent distances between MeSH terms; (2) compared to the widely used indicators proposed by Uzzi et al. (2013), our measures show stronger alignment with peer judgements; and (3) combining all three distance metrics yields more effective identification of novel papers than using any single perspective alone.

---

### SMILE: Self-Explainable Multimodal Information Bottleneck for Medical Diagnosis

Authors: Yuqing Yang, Alexander Schmatz, Zhaozhao Ma, Changkyu Choi, Robert Jenssen, Shujian Yu
Published: 2026-09-04
Categories: cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2609.05174v1

Explainability is increasingly seen as a crucial requirement in AI-based medical diagnosis, particularly in safety-critical clinical decision-making. Most existing explainability methods in healthcare operate in a post-hoc manner and are predominantly designed for unimodal data, which limits their applicability in increasingly prevalent multimodal diagnostic settings. This paper addresses the problem of self-explainable multimodal diagnosis by formulating it within the information bottleneck (IB) framework. We propose a unified learning paradigm that jointly optimizes predictive performance and modality-specific explainability by identifying the most informative elements inside each modality that contribute to diagnostic decisions. To enable tractable and stable optimization, we employ a matrix-based Renyi's $α$-order entropy functional under the assumption of sufficiently expressive encoders. Extensive experiments on representative medical datasets spanning heterogeneous modalities demonstrate that the proposed method consistently achieves strong diagnostic performance, including an absolute accuracy improvement of 9.1 percentage points on the iCTCF dataset. Moreover, the learned explanations provide transparent and modality-aware insights into feature relevance, thereby improving both the explainability and generalization.

---

### WeAgent-MMGenEdit: A Full-Stack Recipe for Multimodal Agentic Image Generation and Editing

Authors: Hui Zhang, Zongkai Liu, Liqiang Niu, Juntao Liu, Han Li, Zhen Cao, Wenchao Chen, Chengduo Zhao, Fandong Meng
Published: 2026-09-04
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.05171v1

Image generation and editing models have advanced rapidly, yet remain unreliable when prompts require external world knowledge. Bounded and long-tail parametric knowledge prevents direct or reason-then-generate approaches from recovering the required facts and visual appearances. Existing agentic generation and editing methods mitigate this limitation with retrieval tools, yet remain constrained by insufficient visual verification, overloaded policy models, and weak integration of retrieved textual and visual evidence. To address these limitations, we present WeAgent-MMGenEdit, a full-stack recipe including a multimodal harness, a scalable data construction pipeline, a comprehensive benchmark, and post-training methods for the agent policy and image backend. We first introduce WeAgent-Harness, a multimodal runtime with persistent evidence management and dedicated verification and integration tools that organize retrieved multimodal evidence into a dense carrier. Upon this, we develop a scalable pipeline for prompt synthesis and agentic trajectory collection, yielding 23K supervised trajectories and 14.7K RL tasks with three-layer verifiable checklists. We further introduce WeBench-MMGenEdit, a bilingual benchmark covering both knowledge-intensive image generation and multi-image editing. Finally, a two-sided post-training recipe based on SFT and RL improves the agent policy and image backend. Together, WeAgent-MMGenEdit enables a 30B-total/3B-active policy to outperform similarly sized policy models and approach the performance of a 1T-parameter agent.

---

### Conformal Prediction for Offensive Security

Authors: Giovanni Cherubin
Published: 2026-09-04
Categories: cs.CR, cs.LG
arXiv: https://arxiv.org/abs/2609.05165v1

Despite its introduction more than a quarter century ago, Conformal Prediction (CP) has seen surprisingly few applications to the cyber security world thus far. In particular, we observe that, while CP has been employed as a defensive measure in many recent works, its use for carrying out attacks (i.e., for offensive security) is hard to trace in the literature. We explore this gap, by presenting initial findings in two key areas of offensive security: Privacy-Preserving Machine Learning, and network traffic analysis.

---

### AxQM: A Textbook-Scale Benchmark for Formal Proof Synthesis in a Library of Finite-Dimensional Quantum Mechanics

Authors: Weichen Winston Yin, Jacob M. Taylor, Dirk R. Englund, Frank H. L. Koppens
Published: 2026-09-04
Categories: quant-ph, cs.AI, cs.LO
arXiv: https://arxiv.org/abs/2609.05157v1

Formalizing mathematics in a proof assistant, where a machine checks every definition, statement and proof, has set a new standard of rigor. Large language models are now capable of formalizing autonomously, even at the scale of whole textbooks. We bring this standard of rigor to physics, where theoretical arguments carry idealizations that are rarely stated fully, and any logical gaps could have a cascading effect on interdependent results. Recognizing the need to evaluate autoformalization systems for physics, we release AxQM, 1,019 kernel-checkable proof-synthesis tasks over 479 items drawn from the textbook Quantum Computation and Quantum Information by Nielsen and Chuang. The tasks are stated in a custom Lean library of finite-dimensional quantum mechanics. By task count, it is the largest proof-synthesis benchmark in physics by a factor of four. AxQM is derived from a near-complete formalization of the formal portions of the textbook, so every task is guaranteed a solution, which we keep private. Grading of the benchmark is done deterministically by the Lean kernel, which checks that the proof compiles, that no sorry appears in it or in any declaration it depends on, and that it introduces no new axioms.

---

### Compression Beyond the Uncompressed: A Two-Stage Training Recipe for Soft Context Compression in RAG

Authors: Shuyu Guo, Shuo Zhang, Zhaochun Ren
Published: 2026-09-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.05152v2

Retrieval-Augmented Generation (RAG) improves knowledge-intensive generation by conditioning language models on retrieved documents, but processing these documents becomes increasingly expensive as retrieval depth grows. Soft context compression reduces this cost by encoding documents into compact continuous representations that can be precomputed and reused across queries. However, many existing methods train compressed models by distilling from a full-context teacher. When the teacher is wrong, such distillation can reinforce its errors, while teacher imitation provides no direct signal for improving beyond the teacher. We propose DEX-Comp, a two-stage training recipe that separates reliable imitation from targeted exploration. Pure Distillation learns only from teacher-correct questions to mitigate error propagation, while Hard Exploration applies outcome-based reinforcement learning to teacher-failed questions to directly optimize answer correctness. Across five open-domain QA benchmarks and retrieval depths from top-$5$ to top-$30$, DEX-Comp at $16\times$ compression outperforms all evaluated compression baselines and surpasses the untuned full-context RAG model in average accuracy, while reducing time-to-first-token by $4.4\times$--$23.7\times$. Evaluations across additional datasets and backbones further demonstrate its generalization.

---

### Compression Beyond the Uncompressed: A Two-Stage Training Recipe for Soft Context Compression in RAG

Authors: Shuyu Guo, Shuo Zhang, Zhaochun Ren
Published: 2026-09-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.05152v1

Retrieval-Augmented Generation (RAG) enhances language models with external knowledge, but the lengthy retrieved context inflates the input and degrades inference efficiency. Soft context compression encodes each document into a substantially shorter embedding sequence. However, most existing approaches are trained by distilling outputs from uncompressed RAG systems, inherently limiting their performance relative to the original model. To address this limitation, we propose DEX-Comp, a two-stage training recipe: Pure Distillation warm-starts the compression model on the uncompressed RAG's correct responses only, and Hard Exploration then runs reinforcement learning solely on queries the uncompressed RAG fails, forcing the model to explore computation patterns better suited to compressed representations. On five open-domain QA benchmarks at retrieval depths from top-5 to top-30, DEX-Comp compresses retrieved contexts by $16\times$ and accelerates inference by $4\times$--$24\times$, while achieving performance comparable to or exceeding the uncompressed RAG baseline across retrieval depths. Ablations and evaluations across diverse datasets and backbones further confirm the contribution of each stage and the generalization of our approach.

---

### Large Language Models with At Most One Spike per Neuron

Authors: Zhuoya Zhao, Parsa Omidi, Aref Jafari, Richard Naud
Published: 2026-09-04
Categories: cs.NE, cs.CL
arXiv: https://arxiv.org/abs/2609.05151v1

Leveraging their inherent sparse event-driven computation, spiking neural networks (SNNs) offer a promising path toward energy-efficient large language models (LLMs). Time-to-first-spike (TTFS) coding generates at most one spike per neuron within a time window, yielding extremely low firing rates. However, conventional TTFS SNNs are restricted to specific structures, making it challenging to encode certain blocks in LLM -- such as layer normalization and matrix multiplication --using TTFS. To overcome this limitation, we introduce a reference-based strategy specifically to encode the four core LLM components: embedding layers, layer normalization, attention-related operations and dropout. We construct a fully TTFS-based SNN architecture and train it end-to-end. Experiments on modern LLMs like BERT and GPT-2 demonstrate that our approach achieves performance comparable to ANN counterparts on natural language understanding and common-sense reasoning, while a clear gap remains on language modeling perplexity. To the best of our knowledge, this is the first work to scale a spiking LLM to 1.5 billion parameters using TTFS coding. We also report an estimate of spike-related energy; this is a spike-count proxy under an established cost model rather than a measurement on neuromorphic hardware.

---

### Beyond Stationarity in Time Series: Discovering Causal Structures and Latent Regimes via Markov Blankets

Authors: Lei Zan, Charles K. Assaad, Emilie Devijver, Eric Gaussier
Published: 2026-09-04
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.05150v1

This paper introduces Regime-aware Constraint-Based and Noise-Based causal discovery with Markov Blankets (RCBNB-MB), a novel causal discovery algorithm for time series that relaxes the common assumption of a single, time-consistent causal structure. Time series are typically observed at discrete time points and often exhibit regime changes that challenge the assumption of a static causal structure, a limitation in many real-world dynamic systems. To address this challenge, RCBNB-MB identifies latent causal regimes, defined as subsets of time points within which a stable causal structure holds. The algorithm follows an iterative strategy that segments the time series into regimes and discovers the causal graph within each regime. By leveraging the Markov blanket rather than direct parents, RCBNB-MB gains robustness to errors in causal discovery and preserves predictive information. We provide theoretical guarantees for RCBNB-MB's ability to recover both regime transitions and causal graphs under reasonable assumptions. Furthermore, we validate its effectiveness through extensive experiments on simulated datasets with known ground truth and real-world IT monitoring data, where taking into account regime shifts is critical. Empirical results show that RCBNB-MB systematically outperforms baseline approaches in accurately detecting regime changes and their associated causal graphs, positioning it as a robust and versatile framework for non-stationary time series analysis.

---