# Research Papers — 2026-09 · bundle page 165 of 266 > Full-text markdown of papers #8201–8250 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/165 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/164.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/166.md) ### How I learned to stop worrying and love StopGrads: Stationarity, Convergence, and a case study on Flow Map Learning Authors: Max W. Shen, Mark Goldstein, Zichu Wang, Aahlad Puli, Rajesh Ranganath Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.16222v1 Stopgrads are widely used in training machine learning models, but stopgrads can alter the gradient, stationary points and convergence guarantees of the original objective, which can make stopgrad training theoretically ungrounded. We introduce a stopgrad regression principle, which identifies a general template for stopgrad objectives with a closed-form characterization of stationary points and their uniqueness, unifying stopgrad objectives for flow maps, reinforcement learning, and diffusion samplers. We provide theoretical grounding for optimizing stopgrad flow map objectives by showing their unique stationary point is the true flow map, and showing positive convergence results for Eulerian and Lagrangian objectives, including MeanFlow and improved MeanFlow. Remarkably, we show that under functional semi-gradient flow, the learned flow map has a closed-form expression composing the initial flow map and the true flow map. We additionally use our stopgrad regression principle to propose modified stopgrad placements for flow map objectives which reduce training memory by 2x. --- ### A neural-astrocyte architecture implements a hybrid automaton for evidence accumulation Authors: Giacomo Vedovati, Ilya E. Monosov, Thomas J. Papouin, ShiNung Ching Published: 2026-09-14 Categories: q-bio.NC, cs.NE arXiv: https://arxiv.org/abs/2609.16217v1 Astrocytes are non-neuronal glial cells that are receiving widespread attention due to their emerging role in neural computation. In this paper, we propose and study dynamical mechanisms by which astrocytes may augment the ability of neural networks to infer context in reinforcement learning (RL) settings. We construct a biologically inspired, two-level dynamical neural-astrocyte network with distinct spatial and temporal organization. We train this model on a hierarchical multi-context task that requires the agent to infer changes in latent task rules based on derived rewards. We find that in this setting, astrocytes enable evidence accumulation of changes in context and subsequent context-specific modulation of neural dynamics. We show that these functions are implemented via two dynamical mechanisms: (i) reward-induced bifurcations that relocate an asymptotically stable attractor into different, context-specific regions of state space, and (ii) the relative shallowness of these attractors, mediated by the entropy of the environment, giving rise to behavioral stickiness. Together, these mechanisms amount to a hybrid automaton, in which uncertainty accumulates until, eventually, the neural dynamics are switched to a new context. This model provides a neuro-dynamic schema, compatible with neural-astrocyte biology and prior empirical observations, for how astrocytes may integrate information from the periphery and drive contextual changes in neural circuits. --- ### Where Should the KV Cache Live? Placement Policies Across GPU, CPU, and SSD for Long-Lived Sessions Authors: Srikanta Datta Tumkur, Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, Ramesh Nampelly Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.16215v1 GPU high bandwidth memory is scarce and expensive, and KV caches consume much of it as chats, agent loops, and document question answering accumulate state. Systems such as Mooncake, LMCache, FlexGen, InfiniGen, and AttentionStore extend GPU memory with CPU DRAM and SSD. The harder question is which blocks belong in each tier, when to move or evict them, and whether prefetching helps. We study these choices in a discrete event simulator spanning GPU HBM, CPU DRAM, and SSD, calibrated against a random forest execution time predictor. We compare recency, reuse frequency, predicted reuse, and an EWMA predictor with prefetch lookahead across chat, agent, and document question answering workloads. Tiering supports 73.02 times more concurrent sessions per GPU and lowers cost per session by 62.04 times. These gains come from tier capacities of 1 plus 8 plus 64, not placement policy. Decode is compute bound at batch size one in our setup, so placement barely affects throughput. It mainly changes PCIe migration traffic and time to first token. Recency produces 2.30 times less migration traffic than reuse frequency for chat. Reuse frequency performs best for agents and document question answering. The existing predicted reuse policy is byte identical to recency, making its agent recommendation effectively recency. A genuine EWMA predictor changes behavior but still ranks behind reuse frequency on the workloads prediction was expected to help. Prefetching does not justify its bandwidth cost. Across the policy and cache size grid, even an oracle with knowledge of future requests never beats no prefetch on migration traffic. Workload specific placement can reduce data movement, but the predicted reuse and prefetch recommendations are not supported as implemented. --- ### Artificial intelligence and biosecurity: capabilities, threat pathways, and defense-in-depth governance Authors: Candace S. Y. Chan, Aris Karatzikos, Ilias Georgakopoulos-Soares Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.16213v1 Artificial intelligence is reshaping biological research across an increasingly connected digital-to-physical workflow. General-purpose large language models can retrieve and integrate scientific information, support experimental planning, and computational analysis; biological foundation models can predict, optimize, and generate proteins, genes, and genome-scale sequences; agentic systems can coordinate multistep research tasks; automated laboratories can partially close the design-build-test-learn cycle. These technologies could greatly benefit medicine, public health, and biotechnology. However, their biosecurity risk depends not only on what the AI can do, but also on who uses it, their expertise and intent, their access to laboratory tools and materials, and the safeguards in place. Current evidence shows that AI uplift exists but primarily affects digital rather than physical tasks. Frontier systems have exceeded expert baselines on in-silico, and screening-evasion benchmarks, whereas controlled wet-laboratory studies find that tacit knowledge and physical execution remain substantial barriers. This review describes the different biological threats from AI tool use, from information gathering and biological design to procurement, synthesis, testing, scale-up, and potential release. We further examine why alignment techniques for general-purpose models transfer poorly to biological ones, and the emerging role of interpretability in auditing whether hazardous capabilities are genuinely removed. We argue for defense-in-depth governance that links capability thresholds to proportionate responsibilities across the biological AI ecosystem, reducing high-consequence risk while preserving beneficial use. --- ### Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics Authors: Daniela Vega, Paula Cárdenas, Hannah Ceballos, Leonardo Manrique, Pablo Arbelaéz Published: 2026-09-14 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.16207v1 Spatial Transcriptomics (ST) has transformed biomedical research by enabling the spatial mapping of gene expression across tissue sections. However, high operational costs, specialized equipment requirements, and sensitivity to experimental noise limit the accessibility and scalability of ST. Recent computer vision approaches aim to overcome these limitations by predicting spatial gene expression directly from histopathology images. While effective, current approaches often suffer from gene expression over-smoothing and overly uniform predictions across tissue regions, suggesting that further progress depends on learning representations that reflect the hierarchical and asymmetric structure of gene regulation and tissue morphology. To address these issues, we propose Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics (HyCLoST), a hyperbolic contrastive learning model that captures the intrinsic hierarchical relationships within ST data. By leveraging hyperbolic geometry and a gene-to-image entailment loss, HyCLoST learns structured, biologically grounded representations that improve gene expression prediction accuracy, achieving a 6% reduction in MSE and an 8% increase in PCC across 26 ST datasets, over previous methods. Our source code is publicly available at https://github.com/BCV-Uniandes/HyCLoST --- ### Calibrate, Then Route: A Measured Study of Learned Request Routing for Disaggregated LLM Serving Authors: Srikanta Datta Tumkur, Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, Ramesh Nampelly Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.16206v1 Disaggregated LLM serving places compute heavy prefill and memory heavy decode on separate GPU pools. Systems such as DistServe, Splitwise, and Mooncake make this separation fast, but routing still determines which instances handle each request. We study a router that estimates the additional completion time on each instance using exact prompt length, predicted output length, post admission KV cache pressure, and SLO class. We develop the policy in a discrete event simulator and validate it on eight NVIDIA A40 GPUs, each running a vLLM engine, with NIXL transferring KV caches between pools. All workloads run at measured saturation. Across three mixed, bursty arrival traces, the calibrated router achieves the highest mean goodput at 0.864, compared with 0.835 to 0.847 for round robin, least loaded, and a length heuristic. It also shows the lowest variance across traces. It beats round robin and the length heuristic on all three traces and least loaded on two. On the third, it trails by 0.003, within run to run noise. Hardware calibration matters: simulator derived constants cost 4.5 goodput points and roughly 40 percent of the tail latency advantage, reducing the scorer to little more than queue counting. Benefits grow with decode pool size and traffic heterogeneity but disappear in pools with three instances, where queue counts are often enough. Under extreme scarcity, greedy cost minimization concentrates requests on the cheapest scored instance, and blind spreading performs better. With calibrated costs, the learned router matches the goodput of round robin using six GPUs instead of seven. --- ### Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration Authors: Aashiq Muhamed, Mona T. Diab, Virginia Smith Published: 2026-09-14 Categories: cs.LG, cs.CL, cs.CR arXiv: https://arxiv.org/abs/2609.16204v1 Safety guardrails in open-weight language models can be readily bypassed using Refusal Feature Ablation (RFA), a technique that identifies and projects out a linear refusal direction from the residual stream, often achieving a high attack success rate (ASR) while preserving model capability. Defending against these attacks typically requires computationally expensive safety finetuning for every new checkpoint. We introduce Decoy Direction Optimization (DDO), a fast, post-hoc weight-editing defense that requires no base-model finetuning. Our approach is based on a simple mechanistic insight: ablation attacks rely on contrastive estimators to find the refusal direction. Rather than trying to hide the true refusal circuitry, DDO actively injects a high-magnitude, nonlinear decoy signal into the network's MLP neurons. When an attacker attempts to locate the refusal direction, the decoy corrupts their estimator, tricking them into ablating a harmless orthogonal feature while the actual safety mechanism remains intact. We prove a spectral bound formalizing this effect and evaluate DDO across six model families, achieving <10% ASR under standard RFA. On Llama-3-8B-Instruct, DDO remains comparable to trained defenses under adaptive multi-phase attacks (65% vs. 58% worst-case ASR) and reduces Heretic weight-level attack ASR from 88.7% to 18%, all at 30 to 450 times lower optimization cost per configuration than the trained baselines. --- ### A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires Authors: Shreyan Mitra, Mohammadreza Narimani, Parastoo Farajpoor Published: 2026-09-14 Categories: eess.IV, cs.CV, cs.LG, eess.SP, physics.geo-ph arXiv: https://arxiv.org/abs/2609.16199v1 This article describes an open image dataset for developing and evaluating active-fire segmentation methods in satellite imagery. The dataset contains 2,148 image-mask pairs from 25 California wildfires, with acquisitions spanning July 2020 to August 2026. Each image is a 512x512-pixel, three-channel composite derived from Sentinel-2 Level-2A bands B12, B11 and B8A at 20 m spatial sampling. A fixed linear rendering is applied throughout the dataset. Corresponding masks distinguish background, SWIR-rule active fire and invalid observations. The masks were generated from shortwave-infrared brightness and near-infrared contrast, followed by constrained neighborhood growth. The release includes chip-level metadata and an incident-disjoint partition containing 18 training, three validation and four test fires. Among the image pairs, 841 contain active-fire labels; these labels occupy 0.0766% of all grid cells. A mask-blind analyst review covers 233 test chips and provides a separate assessment of the rule-generated labels at chip and connected-component levels. Reference training and evaluation code accompanies the data, including a ResNet-34 U-Net implementation with validation-based checkpoint and threshold selection. The archived images, masks, metadata and review annotations support research on rare-class segmentation, learning from algorithmic labels and transfer across fire incidents. The versioned dataset is deposited on Zenodo, with preparation and reuse software maintained in a public GitHub repository. --- ### Permutation-Based Stegomalware in Large Language Models: Threats and Countermeasures Authors: Danny Wood, James Stringer Published: 2026-09-14 Categories: cs.CR, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.16193v1 The difficulty of training large language models (LLMs), together with their ubiquity, raises the threat of stegomalware, where malicious payloads are embedded into model weights. Recent work has demonstrated the use of permutation symmetry in model weights to mitigate these threats, but failed to show neutralization of stegomalware across all weights for LLMs. In this paper, we demonstrate the full potential of behavior-preserving symmetries as a defense against stegomalware, as well as the risks these symmetries pose when exploited by attackers. For stegomalware neutralization, we improve upon previous work, demonstrating that it is possible to select permutations which displace all model parameters. This contrasts with previous methods which left a significant percentage of weights unaltered in LLMs. When used in an attack, we show that permutation symmetries can encode malware into the weights of a model in a way that is theoretically lossless, requires no retraining after encoding, and needs no payload-specific information in the extraction script---a combination of characteristics not previously seen in any single method. While theoretically lossless, permutation can in practice alter model behavior due to the accumulation of numerical error. We therefore quantify the loss in model performance associated with applying these methods, for both attack and defense, showing it to be minimal. --- ### AI-Driven Feedback Systems, Digital Labour, and Silent Quitting: Transforming African Workplaces Authors: Abayomi O. Agbeyangi, Jose M. Lukose Published: 2026-09-14 Categories: cs.CY, cs.AI arXiv: https://arxiv.org/abs/2609.16192v1 The current trend of digitalisation has revolutionised the organisation of work and the way it is measured and performed across the globe, with AI becoming more common for managing labour and performance, as well as employee communication. In African organisations, where there is increasing adoption of remote work, hybrid models of work, digital collaboration, and data-based HR management, the notion of silent quitting has become more relevant, defined as worker disengagement when employees are still doing their job but do not put any effort into achieving good performance and exhibiting any emotion. This paper investigates how AI-driven feedback mechanisms, including sentiment analysis systems, pulse surveys, chatbots, engagement dashboards, and predictive analytics, are changing African workplaces through offering continuous listening, instant performance information and proactive engagement with employees. The study also explores how AI can assist organisations in identifying early disengagement and enable intervention and better employee communication in both private and public sector organisations in Africa. At the same time, we address the challenges of socioeconomic development and governance posed by AI implementation in developing countries, including digital inequality, infrastructure shortcomings, privacy concerns, algorithmic bias, and the risk of workplace surveillance. By situating silent quitting within wider debates on digital labour and automation, the paper contributes an African-centred perspective to discussions on the future of work and offers practical recommendations for HR professionals, managers, policymakers, and technology developers seeking responsible, context-sensitive approaches to workplace transformation across the continent. --- ### Occupancy Network-Guided Autonomous Robotic Partial Nephrectomy Authors: Ethan Kilmer, Pit Henrich, Jiawei Ge, Paul M. Scheikl, Laura Connolly, Soum D. Lokeshwar, Joseph Chen, Justin D. Opfermann, Kaitlyn Kumar, Lauren Shepard, Ahmed Ghazi, Nirmish Singla et al. Published: 2026-09-14 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2609.16186v1 Autonomous soft-tissue cancer surgery has been limited to interventions on organ surfaces, because current systems cannot perceive and adapt to anatomy once it deforms or is cut. We introduce the first vision-guided autonomous system capable of performing complete tumor resections for partial nephrectomy. Our system integrates conditional occupancy networks, trained entirely in a physics-based simulation, that infer full 3-D anatomy (tumor, margin tissue, and kidney) from single-view partial point clouds. These occupancy networks maintain intraoperative tracking even as tissue is cut and deformed, enabling adaptive planning and execution. The surgical platform combines a depth camera for capturing surface point clouds, dual robotic arms for electrosurgical cutting and vacuum-based tissue manipulation, and an autonomous control strategy for tumor resection. In patient-derived hydrogel phantoms under an open partial nephrectomy setting, the robot performed eight consecutive autonomous tumor resections comprising 77 electrosurgical cuts, with all cuts achieving negative surgical margins and 1.61 $\pm$ 0.48 mm mean absolute margin error. This work demonstrates, for the first time, a foundation for supervised autonomous closed-loop, imaging-driven, margin-negative tumor removal in phantoms. --- ### Anatomy of Associative Recall in Fixed-State Recurrences: A Matched-State Decomposition, an Interference Wall, and a Curriculum That Breaks It Authors: Julian Boesch, Andrew Wee Published: 2026-09-14 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.16183v1 Fixed-state recurrences--linear attention and state-space models--are reported to lag behind attention on associative recall, but whole-architecture comparisons cannot say which ingredient is responsible. We decompose masked multi-query recall at a fixed state budget along three single-knob axes: a short causal convolution, the transition structure (rank-1 delta rule vs. diagonal), and decay. The convolution dominates (~+0.5 recall in both families under matched training): comparisons that pit convolution-free cells against a convolution-equipped Mamba measure the missing convolution, not the recurrence. The rank-1 transition beats its diagonal ablation by +0.19/+0.32 at 16/32 pairs, but the margin shrinks to +0.03 once both cells carry the convolution, and a state-matched Mamba-2 ties the unarmed rank-1 cell: no class claim survives. Cells that solve 32-pair recall degrade gracefully with load yet fall to chance retrieving 4 pairs from a distractor haystack--flat across lengths and transitions. Interference under sparse supervision, not capacity: a distance curriculum takes the unchanged architecture from 0.021 to 1.000. Training is a lock-in lottery--a seed either locks in or does not--and the curriculum is the lever. Lock-in rises from 1/10 to 7/10 (p=0.02); dense supervision adds nothing; at L=256 a shaped ramp reopens a boundary the uniform curriculum cannot (4/5 vs. 0/9); and at L=512, where the ramp collapses (0/6), gating it on measured accuracy locks in 6/6 (p=0.001). Bidirectional denoiser cells, reading the query before the haystack, show no measurable advantage over causal training (ten seeds), and collision-key retrieval needs two layers. Arming for recall is free on an S_5 state-tracking guardrail--the armed cell is significantly better at every depth (p<=0.0044). These replace "recurrent models are bad at recall" with a measured decomposition and two cheap interventions. --- ### Skeletal Prototypes on Iterative Nerve Expansions Authors: Jordan Eckert, Henry Schenck Published: 2026-09-14 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.16170v1 Prototype reduction replaces a training set with a smaller representation, and the established methods return a finite set of points. We propose Skeletal Prototypes on Iterative Nerve Expansions (SPINE). The model for each class is an embedded 1-complex rather than a point set. Its initial edge set is a class-conditional Mapper graph, so the data decide which localized clusters are joined. Later phases fit the vertices under a classification objective, and an observation is assigned to the class whose complex is nearest. The segments therefore enter the decision rule and not only the fitting. We evaluate SPINE on seventeen benchmark datasets under stratified 10-fold cross validation, against seven other prototype reduction methods at a matched budget. SPINE attains the highest mean accuracy and the best average rank. It is significantly better than five of the seven competitors under Wilcoxon signed-rank tests with Holm correction. A budget sweep shows that the decision rule using the entire graph segments contribute most when prototypes are scarce, while the method as a whole competes best at moderate budgets. Construction cost places SPINE with the discriminative methods, and it is faster than generalized learning vector quantization on fourteen of the seventeen datasets. --- ### Multisource Remote Sensing and Geospatial Analysis of Vineyard Wildfire Impacts and Resilience: The 2019 Kincade Fire Authors: Parastoo Farajpoor, Mahla Ardebili Pour, Mohammad Bagher Ghiasi, Mohammadreza Narimani Published: 2026-09-14 Categories: eess.IV, cs.CV, physics.ao-ph, physics.geo-ph, stat.AP arXiv: https://arxiv.org/abs/2609.16168v1 Working agricultural landscapes are often treated as background to wildfire disasters, even though they are managed fuel mosaics, productive assets, and parts of regional infrastructure systems. We examine vineyard wildfire resilience during the electrically initiated 2019 Kincade Fire in Sonoma County, California, using an open, event-anchored geospatial framework spanning 4,581 vineyard fields (8,813.2 ha), wildland vegetation, surveyed structures, roads, overhead smoke, and post-fire greenness. Sentinel-2, OpenET, gridMET, soils, terrain, NOAA smoke polygons, an ignition-date OpenStreetMap network, and three-dimensional data inventories were analyzed at native decision scales. Vineyard pixels showed substantially lower descriptive dNBR than wildland pixels inside the perimeter (means 0.130 and 0.337). This contrast did not identify a universal vineyard firebreak effect: a segment-clustered boundary model gave a small negative contrast at 100 m (tau = -0.0166) but changed across bandwidths, failed slope continuity, disappeared in a 100 m donut, and produced a wrong-signed placebo. A 250 m spatial GAM reversed the unconditional pattern: after conditioning on location, terrain, and water use, vineyard fraction was positively associated with dNBR, while residual Moran's I remained 0.519. Beyond spectral impact, all mapped vineyards intersected overhead smoke on at least one day (mean 7.78 potential smoke-days per field), 34.2% of road-network nodes were dead ends, and inside-perimeter vineyards showed a larger greenness deficit through 2021 (recovery ratios 0.815 inside and 0.854 outside). Lower immediate spectral impact therefore did not imply complete resilience. The study provides a reproducible urban-rural informatics template separating descriptive contrasts, conditional associations, exposure indicators, and recovery evidence for decisions in working landscapes. --- ### GPEvac: GNN-Based PPO for Adaptive Evacuation Routing During Shooting Events Authors: Daniel Perkins, Subhadeep Chakraborty Published: 2026-09-14 Categories: cs.AI, cs.CY, cs.LG, cs.MA, eess.SY arXiv: https://arxiv.org/abs/2609.16163v1 The sharp increase in mass shootings underscores an urgent need for systems that guide victims to safety in real time. An effective evacuation system must minimize threat exposure while also accounting for adversarial uncertainty and crowding dynamics. Current methods in the literature are rigidly constrained to layout-specific policies and computationally intractable in large-scale layouts, while practical guidelines simply advise victims to "run", "hide", or "fight". We propose GPEvac: a GNN-based PPO framework that computes adaptive evacuation routes during shooting events. To capture both local and long-distance dependencies, we introduce an edge-first sequential message-passing scheme with a learnable virtual global node. The resulting graph embeddings are integrated into a permutation-invariant scoring mechanism that allows a single learned policy to operate across building layouts of diverse topologies and sizes. Through extensive simulation, we show that GPEvac outperforms intelligent baselines across distinct architectural layouts, significantly reducing total threat exposure. Crucially, the system computes global evacuation routes in just 14.73 ms on local CPU hardware, enabling seamless integration with live surveillance systems. In addition to saving lives during shooting events, the methodologies developed are transferable to other graph-structured decision-making domains, including critical infrastructure, intelligent transportation systems, and adaptive sensor networks. --- ### LLM Inference in a Flash! Authors: Sebastian Zhao, Minseo Kim, Coleman Hooper, Luca Manolache, Michael W. Mahoney, Yakun Sophia Shao, Kurt Keutzer, Amir Gholami Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.16161v1 Large Language Models (LLMs) have shown impressive capabilities across a range of natural language processing tasks, and LLM inference has emerged as a critical workload for enabling downstream applications. The demands of serving LLM inference are becoming increasingly challenging as requests shift toward longer sequences and heavier inference, driven by retrieval-augmented generation, inference-time compute scaling, and long-context applications. Additionally, these challenges are compounded by hardware trends, as memory capacity and communication bandwidth are not scaling as fast as increases in workload complexity. Compute-in-Flash is a promising solution to address memory bandwidth limitations by moving computation close to memory, and to exploit the large capacity of SSD technologies. However, it is challenging to deploy LLMs on these systems as they lack support for high-precision floating point operations and have limited write endurance. In our work, we aim to address these challenges by designing inference algorithms to enable LLM inference on Flash compute-in-memory devices. We present an end-to-end integer-only quantization approach to eliminate expensive floating-point computations. To address the limited write endurance, we design a dictionary-based KV cache compression strategy based on sparse dictionary coding that represents each KV vector as a linear combination of static dictionary vectors. These algorithmic improvements enable us to exploit the benefits of Compute-in-Flash for both model weights and KV cache, and to minimize expensive data transfer operations. Across Llama-3.1-8B and Qwen-2.5-7B, our combined method exhibits limited accuracy degradation while reducing dynamic KV cache traffic by 15$\times$. --- ### Computer-assisted global regularity across nonlinear families of three-dimensional periodic Navier-Stokes flows Authors: Jose Luis Lima de Jesus Silva Published: 2026-09-14 Categories: physics.flu-dyn, cs.LG arXiv: https://arxiv.org/abs/2609.16157v1 Numerical simulations reveal how vortices stretch and transfer energy, but establishing smooth evolution requires bounds that remain valid beyond the simulated resolution. Here I develop a computer-assisted framework that establishes global regularity for continuous families of three-dimensional periodic Navier-Stokes flows. Its central construction combines finite reference trajectories with a common error bound that covers an interval of centre fields and infinitely many smooth perturbation modes. The method retains the complete nonlinear residual before spectral truncation and controls the evolution until viscous decay guarantees regularity for all subsequent times. Applications to cyclic-shear, Arnold-Beltrami-Childress and three-component Taylor-Green fields yield explicit perturbation radii and include initial conditions outside the direct Fourier-Wiener smallness criterion. A parameter-uniform extension covers a connected family of non-Beltrami Taylor-Green centres without repeating the proof for individual parameter values. An ensemble of 4,096 configurations, supplemented by 1,600 refinement trajectories and public turbulence data, connects the mathematical observables to spectral transfer and vortex geometry. Matched neural-operator experiments show that physics-informed training improves physical prediction, while also revealing that these gains do not necessarily improve the discovery of proof-limiting initial conditions. Together, these results provide a reusable method for establishing regularity across prescribed flow families and a quantitative setting for evaluating how learned predictions can assist rigorous computation. --- ### LLMs as Master Forgers: Generating Synthetic Time Series Data for Manufacturing Authors: Mantek Singh, Jeshwanth Challagundla, Prateek Karnal, Gagan Ganapathy, Vineet Shah, Ridam Arora Published: 2026-09-14 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.16155v1 This paper presents a novel framework leveraging Large Language Models (LLMs) to generate synthetic time series data for manufacturing processes. Motivated by the scarcity of labeled time-series data in real-world manufacturing settings, which hinders the development of robust machine learning models, we explore the potential of LLMs to learn complex temporal dependencies and generate realistic synthetic data. Our approach involves fine-tuning pre-trained LLMs on manufacturing process instructions and employing a Retrieval Augmented Generation (RAG) technique to enhance data diversity and realism. We evaluate our method against traditional time series modeling techniques like ARIMA and LSTMs, using quantitative metrics, PCA analysis, and downstream task performance (anomaly detection). Results demonstrate that our LLM-driven framework outperforms these baselines, generating high-quality synthetic time series data that effectively captures temporal dependencies and statistical properties of real manufacturing data, leading to improvements in downstream task performance. --- ### DenseFace: Bias Mitigation in Face Recognition via Density-Aware Probabilistic Matching Authors: Mansur Bultygov, Vadim Seliutin, Dmitry Nekhaev, Ivan Laptev Published: 2026-09-14 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.16149v1 Despite steady progress in face recognition, current face recognition models still suffer from significant demographic biases. While approaches for bias mitigation have been proposed, existing methods often impose constraints on the training procedure and result in the degradation of recognition accuracy. To address this issue, we here introduce a method that reduces racial bias in pre-trained face recognition models without compromising their accuracy. To this end, we model face embeddings of each person by von Mises-Fisher (MF) distribution. We next observe the dependency between demographic attributes and the density of MF distributions, and propose DenseFace, a probabilistic face matching procedure that accounts for differences in MF distributions. Our extensive experiments demonstrate DenseFace to consistently reduce racial bias in strong face recognition models varying in network architectures, training datasets and loss functions. Notably, DenseFace preserves recognition accuracy and requires no retraining of the underlying face recognition model. Our work also investigates previously adopted bias measures and makes suggestions. --- ### Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation Authors: Gautam Kishore Published: 2026-09-14 Categories: cs.AI, cs.CL, cs.LG, cs.NE arXiv: https://arxiv.org/abs/2609.16145v1 We study a practical question: can a small correction module fix errors in a frozen language model's outputs without degrading its base capabilities? We propose CRN v2, a lightweight logit-level correction module (~34M trainable parameters, 0.73% of the 4.65B text module) that sits atop a fully frozen Gemma 4 E2B model. The base model is never updated; only the correction module learns, via supervised fine-tuning followed by reference-free DPO on 83,400 error-correction pairs. On a 60-question domain exam (CEHRI: Certified Human-Robot Intelligence, covering facts, arithmetic, and implicit-goal reasoning), CRN v2 corrects 53.3% of base-model errors (reworded variant: 43.3%) while showing no degradation on tested capability benchmarks (MMLU/BoolQ N=200; car-wash N=8). A LoRA baseline at the matched CRN v1 budget (6.6M params, rank 19) achieves 83.3% correction but suffers 30-75% capability loss on the same benchmarks -- the correction-capability tradeoff. An ablation shows that the KL preservation term (lambda=0.1) is critical: lowering it to 0.01 degrades correction to 35.0%. A hidden-state injection variant at earlier layers (1.6M params, SFT-only) reaches 50.0%/55.8% but does not exceed logit correction; shallower injection (layer 4) drops to 30.0%/28.3%; multi-depth logit correction (~35M) reaches only 40%; and longer training (5,000 SFT + 2,000 DPO) stays at 53.3% -- none of the alternative configurations we tested exceeded the rank-128 logit result, consistent with a best-achieved result of ~53% rather than a floor. This is a study of a design principle (frozen base + logit correction + KL anchoring), not a claim of architectural novelty. All code, main-result weights, and evaluation scripts are released (deep variant as code only -- no trained deep checkpoints). --- ### Optimal Pruning for Neural Architectures using Fisher Information Distances Authors: David S. Berman, Yen-Yu Fu, Edward Hirst, Thelma Chiwete Obirai Published: 2026-09-14 Categories: cs.AI, cs.IT, math.DG arXiv: https://arxiv.org/abs/2609.16129v1 A new scheme for parameter pruning is introduced, derived from the differential-geometric distance in model space. Pruning a parameter sets its value to zero, representing a displacement of the model to the hypersurface on which that parameter vanishes. The minimal distance from the unpruned model to this hypersurface is naturally computed via the geodesic distance in the model space as determined by the Fisher information metric. This distance determines the true change in the model, and its performance, under pruning. By analysing progressively more faithful approximations of this geodesic distance a natural hierarchy of optimality for pruning methods is determined. This starts with the traditional magnitude pruning, then develops into new more sophisticated and effective pruning schemes. The method is demonstrated for both fully-connected networks and vision transformers, on MNIST and CIFAR-10, over the complete $0$-$100\%$ pruning range and across five random seeds. It outperforms pruning by parameter magnitude and by the local Fisher information alone in every architecture and dataset combination considered, on both accuracy and the Matthews correlation coefficient. Additionally, analysis of different levels of geodesic approximation produces intermediate pruning schemes that are computationally efficient and maintain near-optimal performance. This geometric picture supplies not only a state-of-the-art pruning methodology for AI models, but also a verified and mathematically-motivated justification for pruning schemes. --- ### A Decision-Support Audit Protocol for Supervision Drift in Proxy-Labeled Credit-Risk Prediction Authors: Mehrdad Shoeibi, Muhammad Shabanpour, Waldemar Karwowski, Niloofar Yousefi Published: 2026-09-14 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.16102v1 Credit-risk models are trained on proxy labels and deployed under temporal and segment change, yet no single transfer metric separates base-rate shift, probability-scale shift, and feature-label relationship change. We contribute a design-science artifact: a locked, multi-signal audit protocol for supervision drift in proxy-labeled credit-risk prediction. Five layers (transfer performance, an oracle-gap probe, a calibration diagnostic, feature-label stability, and a synthetic positive control), thresholds, and decision rules were locked before interpretation; a bounded reading is a designed outcome. On a public LendingClub dataset (temporal 2013 to 2016 and cross-segment transfer), ranking is stable and oracle gaps are small; the clearest temporal signal is a prevalence and probability-scale mismatch that intercept-only diagnostic recalibration largely reduces, though its cause is not identifiable from the available release. The positive control responds only to larger injected shifts; subtler drift cannot be excluded. Mapping diagnostic patterns to governance actions is conceptual guidance, not validated here. --- ### SWB-DM: A Calibrated Sliced-Wasserstein-Barycenter Aggregator with Delayed-Momentum Caching for Byzantine-Robust Federated Learning under Partial Participation Authors: Saranraj S, Saranya M S, Alex David S, Ajay Kumar A Published: 2026-09-14 Categories: cs.LG, cs.CR, cs.DC arXiv: https://arxiv.org/abs/2609.16099v1 Robust aggregation methods for federated learning quietly rest on a fragile assumption: that whoever shows up in a given round is a fair sample of the full population. In practice, they rarely are. When only a handful of clients participate per round, even a modest fraction of adversaries can dominate that sample and silently invalidate the finite-sample guarantees that coordinate-wise median, Krum, Bulyan, and trimmed mean all depend on. We introduce SWB-DM to address this directly. SWB treats each slice of a client update as a one-dimensional distribution, computes a trimmed Wasserstein barycenter across clients, and recovers coordinate identity via a medoid-based gauge-fixing step -- a heuristic we developed and do not claim it belongs to standard optimal-transport theory. DeMoA-style delayed momentum then caches updates across the full client population each round, decoupling robustness from whoever happened to be sampled. Trim ratio calibration is not cosmetic: under-trimming causes collapse at corruption levels a properly calibrated model survives. Across 448 CIFAR-10 configurations, plus CIFAR-100, FEMNIST, and a 500-client scalability run, we find several mechanistically distinct failure modes. Even-sample coordinate-wise median degrades to a deterministic wrong answer. Krum silently violates its own n greater than 2f+2 precondition and diverges without warning. Bulyan's n greater than or equal to 4f+3 threshold produces a sharp pass/fail boundary. On attacks, IPM defeats order-statistic defenses -- including SWB -- more reliably than ALIE, confirmed through delta-space measurements against a convergence bound. SWB-DM's cache carries a real warm-up cost, but extending all baselines to the same round budget shows its CIFAR-10 gains are disproportionately large. On CIFAR-100, FLTrust benefits more -- for reasons entirely unrelated to caching. --- ### Universal Defenses for Tool-Integrated LLM Agents Against Adversarial Attacks Authors: Xiaoyan Li, Yunli Wang Published: 2026-09-14 Categories: cs.CR, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.16098v1 Large Language Model (LLM) agents have demonstrated impressive capabilities across a variety of domains, particularly when integrated with external tools for multi-step task completion. However, they are increasingly vulnerable to adversarial attacks, including direct prompt injection, indirect prompt injection, memory poisoning, and backdoor attacks, which exploit the model's openness to prompt injection and tool manipulation. In this work, we explore practical and generalizable defense strategies within a unified framework across these four attack types. We introduce two universal tool-based defenses: Attacker Tool Filtering, which uses anomaly detection (e.g., Isolation Forest) to identify and remove suspicious tools, and Normal Tool Recalling, a white-box method that restores the agent's original toolset prior to planning. Additionally, we incorporate prompt-based defenses: Chain-of-Thought prompting and self-reflection techniques to enhance reasoning and task paraphrasing to mitigate attacks. Experimental results across both four open-source LLMs (Gemma2-9B, Qwen2-7B, LLaMA3-8B, and LLaMA3.1-8B) and three proprietary LLMs (GPT-3.5, GPT-4, and GPT-5) show that our methods significantly reduce the Attack Success Rates (ASR), achieving 0% ASR in many settings, while preserving or even improving the original task success rate. These findings highlight the promise of simple, modular, multi-layered defenses for strengthening the security and robustness of tool-integrated LLM agents. The code is available at https://github.com/Xiaoyan-Lisa/Defenses-for-Tool-Integrated-LLM-Agents-Against-Adversarial-Attacks. --- ### Coaching Qwen3 Coder 30B to Think Like a CodeClash Arena Agent Authors: Ivy Ning Zhang Published: 2026-09-14 Categories: cs.SE, cs.AI arXiv: https://arxiv.org/abs/2609.16096v1 Large language model coding agents have recently become useful for software tasks, but weaker or open-weight agents still struggle to reliably interpret user intent and execute complex multi-step workflows. This gap is especially visible in long-horizon settings, where an agent must repeatedly inspect prior outcomes, diagnose failure, and choose the next code edit under interaction constraints. It motivates a natural question: what can we do to improve the thinking process of a weak code agent? We study this question in CodeClash, a code-arena benchmark where the original work evaluates 8 commercial coding agents across 6 arenas through multi-round tournaments. Since Qwen3 Coder Plus ranks last among them, we take the open-weight Qwen3-Coder-30B as a case study and investigate how to improve it with distilled knowledge from stronger agents. Our analysis shows that Qwen3-Coder-30B is not well optimized for arena-style interaction: it frequently produces syntax and protocol-breaking errors and exhibits weak strategic adaptation across rounds. These failures are difficult to correct with vanilla instruction tuning alone, since offline SFT cannot directly verify whether a generated action is valid or beneficial. To address this, we propose ReAct SFT, which rewrites teacher trajectories into explicit (obs)(thought)(act) chains, and trajectoryquality weighted SFT, which reweights samples to encourage post-edit checking. ReAct SFT substantially improves strategic behavior, and our fine-tuned model outperforms the original Qwen3 Coder Plus in tournament evaluation. --- ### RAG-CT: Mitigating Privacy Risks on Retrieval-Augmented Generation Systems via Scanning Prompt Distribution Authors: Xingyu Lyu, Jiayimei Wang, Jianfeng He, Ning Wang, Yidan Hu, Yimin Chen Published: 2026-09-14 Categories: cs.CL, cs.AI, cs.CR arXiv: https://arxiv.org/abs/2609.16095v1 Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for improving the quality of generated contents of Large Language Models (LLMs) by grounding responses in external knowledge, thus reducing hallucinations and factual errors. However, recent studies have highlighted a critical vulnerability: adversaries can exploit the retrieval process to extract personally identifiable information (PII) from the underlying corpus. To mitigate this risk, we propose a novel defense, RAG-CT, that identifies malicious queries by analyzing their entropy and margin distributions and using a score-based detection method. Extensive experiments with four state-of-the-art attack strategies and four defense baselines on two datasets show that our approach significantly reduces PII leakage while outperforming existing defenses. This work provides a lightweight yet effective mechanism to protect RAG systems against PII leakage without requiring modifications to the underlying LLM or retriever. --- ### Evaluating Open-Weight E-Commerce Agents with Environment-Grounded Verification Authors: Nimit Shah, Haitz Sáez de Ocáriz Borde Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.16093v1 A shopping conversation has many routes to the same cart, and a task-success rate reduces all of them to one score. We build a deterministic and reproducible e-commerce environment that precommits each trial's customer and trajectory parameters, including the persona, difficulty, target cart, and an item reveal schedule. A simulated consumer attempts to buy a target cart from the environment with assistance from the evaluated model. The environment guides the simulator's actions and records every assistant action alongside the environment state at that point. After the trial, these records allow the evaluator to assess individual parts of the conversation against the retained evidence. For example, the evaluator penalizes a search for failing to surface a target product only when the customer has already mentioned that product. We further use this evidence to apply different penalties to tool calls depending on how the assistant's actions compare with an expected tool-call set. Our environment also interacts with the simulator bidirectionally, reading its output to stop the trial when the simulator determines that the customer has become too frustrated and injecting directives in real time that specify when to explore, defer buying an item, or recall a previous exchange. This interaction creates an open-ended and verifiable simulation. Across eight open-weight agents from 20B to 35B parameters, with 160 trials per agent and 44 metrics, the resulting capability profiles distinguish under-action, over-purchase, unsupported product attributes, and poor search, all of which terminal success obscures. --- ### Distilling Foundation Models for Agentic What-If Reasoning:Cost, Latency, and Governance in a Hybrid LLM+SLM Architecture Authors: Sourish Dey, Aditya Kumar Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.16091v1 Tabular foundation models deliver strong zero-training predictive performance via in-context learning, but their high inference latency makes them impractical as hot-path decision backends in interactive agentic loops. We distill a TabPFN teacher into a compact feed-forward student across a business-decision simulation on UCI Adult and five OpenML benchmarks: the classification head compresses 53.2M parameters to 8,546 (6,220x); the deployed two-head loan pipeline compresses 111.4M parameters to 17,059 (6,532x). The student retains 95.4-100.5% accuracy and 96.8-100.0% AUC, with the lowest accuracy retention on credit-g at 95.4%; an alpha = 0 hard-label control shows that the teacher's soft targets provide a 2.1-7.0 AUC point gain. --- ### MUUNRiver-Bench: Diagnosing Relation-Dependent Music Retrieval with Multimodal Instructions Authors: Zhancheng Guo, Congren Dai, Shangda Wu, Jianhuai Hu, Danni Zhao, Xiaobing Li, Maosong Sun Published: 2026-09-14 Categories: cs.SD, cs.AI arXiv: https://arxiv.org/abs/2609.16090v1 Music retrieval is relation-dependent: given a reference track, a listener may seek its style with a new theme, a cover, or a comparable voice, and these intents demand contradictory rankings. We present MUUNRiver-Bench, a diagnostic benchmark whose reference-audio queries use natural-language instructions to define relevance. A pipeline combining expert genre priors, LLM-generated prompts and lyrics, synthesis, and expert review yields 3,440 tracks spanning 13 genres and 116 sub-genres, and seven tasks: similar-music, style-preserving lyric-rewriting, lyric-preserving style-rewriting, cover, vocal-timbre, isolated-vocal, and segment retrieval. Across six models in eight configurations, task-wise rank reversals reveal complementary biases: acoustic encoders favour local identity, whereas text-aligned encoders favour semantic relations. Frozen encoders diagnose default similarity preferences; instruction-aware and audio-text fusion systems provide exploratory tests of textual conditioning, with neither simple fusion scheme consistently improving its backbone --- ### Is INT8 Portable? A Cross-Platform Measurement Study of Quantized Inference on Embedded and Automotive Accelerators Authors: Yuyeong Shin Published: 2026-09-14 Categories: cs.AR, cs.LG, cs.PF arXiv: https://arxiv.org/abs/2609.16085v1 Eight-bit integer (INT8) post-training quantization is the default recipe for edge deployment, under a widely held assumption: INT8 makes inference faster at a small, predictable accuracy cost, and a model quantized once can be carried to any target. We test that assumption with a controlled measurement study across seven hardware classes -- ARM and x86 CPUs, a discrete GPU, an NVIDIA Jetson AGX Orin iGPU and its NVDLA cores, and two vendor NPUs (Qualcomm Hexagon HTP, DEEPX DX-M1) -- holding the ONNX artifact and the quantization scales fixed so the integer kernel or ISA is the only free variable. Portability fails on three axes. (1) The sign of the INT8 speedup is set by the CPU's dot-product ISA (ARM dotprod/SDOT, x86 VNNI): cores that have it speed up by up to 2.1x, cores that lack it slow down by 1.7x, for the identical model and runtime. (2) INT8 outputs are not portable, and the rule is an invariance rather than a gradient: FP32 predictions are bit-identical for every pair (1000/1000), while INT8 predictions agree 1000/1000 exactly when two targets share an integer kernel and 958-965/1000 whenever they do not -- independent of whether the boundary is CPU<->CPU or CPU<->accelerator, and invisible to top-1 accuracy, which is preserved. (3) Vendor NPUs own quantization: a bring-your-own QDQ graph fails silently on one NPU (external scales ignored, accuracy 0.75 -> 0.005 while it compiles, profiles and runs without error) and loudly on the other (the compiler refuses the graph), so only the vendor's native path yields a correct engine. We further show that edge-NPU latency regimes are set by output/device-to-host transfer size rather than compute, and locate the transition with a fixed-compute sweep. We release the scripts and 32 reports. "Quantize once, deploy anywhere" is unsafe for embedded and automotive deployment, where per-input determinism and redundancy matter. --- ### Predicting Social Media Engagement using Machine Learning Authors: Ritwik Singh, Mayukh Majumdar, Subodha Kumar Published: 2026-09-14 Categories: cs.SI, cs.LG arXiv: https://arxiv.org/abs/2609.16082v1 Social media platforms are popular channels for disseminating information, owing to their large user bases and ease of access. Companies also use social media as an important aspect of the advertising process. By creating high-quality posts, companies can strengthen their engagement metrics and increase their follower count. While a growing body of research has examined social media engagement, fewer studies have jointly examined the visual, textual, and temporal features of image posts, even though these features collectively determine the performance of content on social media. To understand the important drivers of social media engagement, we collect image posts of furniture firms on Facebook and extract visual, temporal, and textual features from them using text and image analytics methods. We evaluate several machine learning models - including Random Forest, Light Gradient Boosting Machine (LightGBM), and eXtreme Gradient Boosting (XGBoost) - to assess the drivers and the prediction power of social media engagement using the features from our data. Our research quantifies the extent to which these features are associated with interactions and provides recommendations that organizations may consider. --- ### Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection Authors: Keertana Chidambaram, Andrew Ilyas, Vasilis Syrgkanis Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.15989v1 Chain-of-thought (CoT) monitoring is a safety strategy where the reasoning of a large language model "actor" is inspected by a "monitor" (often another language model) for signs of unsafe planning, deception, or misalignment. We find that planting harmful but benign-sounding reasoning in the actor's context can steer it to perform adversarial actions while evading monitors, an attack we term "plan injection". We initially discover this attack in the multiple-choice question-answering monitorability setting proposed by Lanham et al. (2023), using the investigator-agent elicitation framework of Li et al. (2025). We generalize the attack and show that the discovered behavior scales to harder tasks (achieving 25-33% monitor evasion rates across different monitorability benchmarks) and larger models such as DeepSeek-R1. Across the settings we study, actor models not only follow injected plans but also paraphrase them as their own reasoning, without explicit attribution to the injections. Finally, we find cases where extra monitor resources cause harm - giving the monitor access to the injected plan drops detection by as much as 50% in the Bio-Math task and in a case study on monitor reasoning budget, we find transcripts where additional thinking tokens are spent rationalizing the injected plan rather than flagging it. --- ### Bellman Policy Optimization Authors: Zhuoqing Song, Haotian Xu, Xikun Zhang, Lidong Bing Published: 2026-09-14 Categories: cs.LG, cs.CL, math.OC arXiv: https://arxiv.org/abs/2609.15987v1 Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models (LLMs). We introduce Bellman Policy Optimization (BPO), a critic-free method derived from Policy Mirror Descent (PMD). For autoregressive generation with terminal rewards, BPO uses the Bellman equations to reformulate PMD as a trajectory-level objective. The reformulation avoids estimating state values at intermediate states. We prove that it has the same unique optimal solution as the original PMD objective. We derive the practical BPO loss by approximating this objective. Its mismatch-correction weight is a smoothed ratio of complementary token probabilities. Experiments on mathematical reasoning benchmarks demonstrate the effectiveness of BPO. --- ### Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and Theoretical Computer Science Authors: Honghao Lin, David P. Woodruff, Yuan Deng, Jieming Mao, Song Zuo, Vahab Mirrokni Published: 2026-09-14 Categories: cs.AI, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.15983v2 Language models can produce plausible short proofs, but may still be unreliable on long-horizon research problems, where progress depends on a sequence of uncertain and interdependent decisions. We introduce Stellar Colosseum, a model-agnostic harness for allocating inference across research in mathematics and theoretical computer science. Colosseum explores alternative strategies before proof construction, uses a readiness gate to decide when a route is mature enough to decompose, represents the proof plan as interdependent section-level subproblems, and routes verifier findings back to the affected part of the argument. Across these stages, it generates candidates in parallel, attacks them with targeted falsification, and combines candidates and their critiques into a single research artifact through overlapping random-sample tree aggregation. The Colosseum workflow has been integrated into Google Antigravity's Teamwork framework as the Long Proof pattern. We demonstrate the capabilities of Colosseum through open-ended research and evaluations on theorem-proving and competitive programming benchmarks. Using Colosseum with Gemini 3.1 Pro, we obtain several new results that address open problems arising from papers published at top venues such as FOCS and JMLR. On TCS-Bench, a benchmark of research-level theorem-proving tasks drawn from papers published at FOCS, STOC, and SODA, Colosseum achieves 71.0% accuracy using Gemini 3.1 Pro and Gemini 3.7 Flash. In a separate Codeforces evaluation using Gemini 3.1 Pro, the proof-oriented pipeline with execution feedback solves 218 of 222 problems. --- ### The Router Within: Eliciting Native Skill Routing from a Frozen LLM Authors: Ruishuo Chen, Xun Wang, Yu Chen, Zhuoran Li, Longbo Huang Published: 2026-09-14 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.15982v2 Skills extend an LLM agent beyond its parametric knowledge, and the gain they promise rests on picking the right one. Deployed harnesses route by preloading every skill's metadata into the context, which disperses the agent's attention and caps the library size. Retrieval pipelines move the selection out of the context, but also out of the agent's capability. We show that the frozen agent LLM already carries the routing signal in its own forward passes, and that two linear maps suffice to read it out with no skill text in the context. Our Gavel (Glance And Verdict from a frozen LLM) reads it in two steps. A glance scores the full library by matching the task's mid-layer states against a compact bank that one forward pass builds for each skill at installation, with the two maps as the only trained parameters. A verdict then resumes each shortlisted skill's forward pass, reads the model's own likelihood and yes/no judgment, and fuses both with the glance as a product of experts. Trained once, Gavel transfers zero-shot to three public benchmarks and SkillTraj, our new benchmark of 372 simulated agent trajectories. On Qwen3-32B it outperforms progressive disclosure and retrieve-and-rerank pipelines that add 1.2B to 16B external parameters, by up to 13.4 points on written tasks and up to 21.9 when the need for a skill arises mid-rollout. Routing accuracy improves as the backbone does, and in a bash-agent harness Gavel lets the 32B trigger the right skill on Skill-Use more often than models of up to 1.6T parameters in Codex. --- ### A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models Authors: Xingyun Wang, Haomin Zheng, Man Yuan, Leqian Yang, Ziming Liu Published: 2026-09-14 Categories: cs.LG, cs.CV arXiv: https://arxiv.org/abs/2609.15980v1 When a video model generates physically incorrect motion, did it fail to learn the correct motion, or did it learn it but fail to use it? We show the latter: the correct motion remains available inside the model and can still be made to control the generated video. We train on videos where red masses oscillate slowly and blue masses oscillate quickly, then test a red mass with fast observed motion. Even when the model generates slow motion in this conflicting case, a low-dimensional edit predicted from simple physical variables restores the correct fast motion. We call this ability causal writability. At fixed strength, we find a sharp depth boundary: the same edit changes the video before the boundary but not after it. This closure marks commitment for that write. The motion signal nevertheless remains, and a stronger downstream write can restore physical motion, while excessive gain overshoots. Early causal writability predicts which errors training later corrects: those errors are writable at more network depths than errors that persist. We reproduce both causal writability and its sharp closure in a pretrained 1.3B video model, supporting generality across model scale and training regime. --- ### Disentangling Representation Evolution in Transformers through Directional Decomposition Authors: Shwai He, Haichao Zhang, Shen Yan Published: 2026-09-14 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.15975v1 Transformer representations evolve through learned additive transformations that either preserve their current direction or redirect it. We study this evolution as a functional geometry, decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial parallel components beyond the residual identity path. We then apply the decomposition in two spaces: to attention and MLP updates relative to the hidden state, and to attention value aggregation relative to the current token's value. Targeted edits reveal a strongly space-dependent asymmetry: exclude-self value-space parallel manipulation is markedly more robust than residual-space and perpendicular counterparts, preserving the direct self message while scaling only the non-self aggregate. The same decomposition gives a component-resolved description of compression-induced update error: perpendicular error separates compression methods more clearly than parallel error. Extensive experiments further demonstrate that full-aggregate parallel suppression during from-scratch pretraining lowers validation-loss trajectories and improves downstream averages, with the value-space variant strongest. Together, these results connect representation geometry to editing robustness, compression diagnosis, and training-time intervention. Code is available in the \href{https://github.com/Shwai-He/Transformer-Geometry}{project repository}. --- ### Discovery Foundation Models: Toward Open-Ended Discovery Intelligence Authors: Ling Yang, Zhenfei Yin, Yingcheng Wu Published: 2026-09-14 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.15973v1 Foundation models have progressed from learning and reasoning over existing knowledge, to increasingly learning through action, tool use, and outcome feedback. We argue that the next frontier is a further transition: from solving and acting within problems specified by humans to participating in the process by which new problems, representations, explanations, and knowledge are created. We refer to this capability as Discovery Intelligence. We formulate Discovery Foundation Models (DFMs) as general-purpose model systems for open-ended discovery. A DFM operates over a revisable research state and supports seven coupled capabilities spanning problem discovery, formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual discovery improvement. We instantiate this framework with Zetema, which couples explicit research-state dynamics, verification and experimental gating, external grounding, and cross-task Discovery Skill evolution. We further ground the framework with GALILEO, a real therapeutic-discovery system in which Dry-Lab reasoning, robotic and hands-on Wet-Lab experimentation, external biological evidence, and iterative hypothesis and design revision form a closed physical discovery loop. We then formulate a unified approach to capability formation and process-centered evaluation, enabling discovery behavior to be trained, improved, and measured beyond final-answer performance. Together, these components establish discovery as a learnable, executable, and evaluable capability of foundation-model systems. We view this shift as a broader progression in intelligence scaling: from learning over existing knowledge, to learning from action outcomes, and ultimately to participating in the construction, testing, and revision of the structures through which new knowledge is discovered. Code: https://github.com/Gen-Verse/DFM-Plans --- ### Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States Authors: Zixuan Wang, Yufan Zhou, Jinzhou Tang, Xinle Yu, Chengjun Wu, Lyumanshan Ye, Zhaoxiang Feng, Letian Peng, Adyasha Patra, Fan Bai, Enze Ma, Zhengding Hu et al. Published: 2026-09-14 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.15972v1 As language models become more capable, long-term collaboration in learning, reasoning, and decision-making calls for a deeper understanding of the people they serve. Yet training such human-aware language models faces a fundamental supervision gap because current datasets for LLM assistant training contain few if any well-informed responses explicitly grounded in users' unspoken beliefs and goals. Scaling such supervision is inherently constrained, as users' underlying states are not directly observable. We thus propose the Mind2Dialogue framework to mitigate this gap by simulating users' mental states and turning them into privileged supervision for human-aware training. Specifically, we first propose a psychology-guided simulator that preserves personal characteristics while updating mental states through interaction to generate coherent conversations. The key idea is to enforce a shared evolving mental state that drives user behavior and guides an Oracle assistant's responses. Our privileged distillation then trains models on the Oracle's well-informed responses to assist users without direct access to their mental states at deployment. Moreover, we propose to evaluate human-aware learning by combining personalization and theory of mind, examining how models understand people and act on that understanding. Training on the full Mind2Dialogue corpus improves every reported personalization metric over the corresponding Qwen, Llama, and OLMo instruction-tuned baselines, including gains of 26.6 to 40.9 percentage points in preference-following generation. The gains extend to belief and action reasoning on Qwen and Llama, beyond personalized assistance. Looking forward, Mind2Dialogue makes user simulation a foundation for genuine AI collaborators that understand beliefs and intentions behind people's words and support their long-term goals across education, work, and everyday life. --- ### Verifiable by Construction: Claim-Level Evaluation of Verbatim Citation in Clinical Question Answering Authors: Jiashuo Zhang, Yuling Chen, Yvonne Commodore-Mensah, Michael Oberst Published: 2026-09-14 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.15964v2 Large language models (LLMs) have been widely adopted for clinical question answering (QA). Current systems can attach citations to their answers, but these often point to broad texts, leaving time-pressed clinicians unable to verify them efficiently. An alternative is to ensure that responses are verifiable by construction: providing fine-grained verbatim quotes from reference material that substantiate claims, so users can verify an answer without opening other documents. In this paper, we evaluate the ability of current models to perform this task end-to-end: from providing citations for every factual claim, to producing verbatim quotes, to ensuring that those quotes fully substantiate the claims. To do so, we build a standardized harness over four clinical practice guidelines and evaluate twelve LLMs on 222 synthetic clinical questions, measuring each of these stages separately. We find that most models can attach verbatim quotes to over 90% of their claims from prompting alone, apart from some lightweight models such as claude-haiku-4.5. Yet these quotes often fail to substantiate every detail of the claims they accompany. For instance, claude-opus-5 produces verbatim quotes for 98.0% of its claims, but fully substantiates only 37.1%. Our work provides insights into the current capability gap of LLMs in building verifiable clinical QA systems, along with artifacts for future research. --- ### Privacy-Aligned Personalized Federated Learning with Compact Adaptation and Variable-Length Gaussian Communication Authors: Yilin Xu, Chun Hei Michael Shiu, Chih Wei Ling, Linqi Song Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.15950v1 Record-level differential privacy exposes a structural misalignment in personalized federated learning when client-specific variation is low-dimensional while training repeatedly releases high-dimensional updates. In this paper, we address this misalignment by releasing a private client context once and confining repeated adaptation to a fixed coefficient space. Beyond dimensionality reduction, the factorized generator induces an adaptive optimization geometry that reshapes noisy updates, and controlled ablations show that most of its private-training gain is retained by radial evolution. To further reduce the communication cost, we realize the Gaussian mechanism for coefficient updates directly through variable-length quantization with finite expected code length, so that the quantization error itself serves as the required privacy perturbation rather than extra distortion. Across MNIST and CIFAR-10, our design matches or outperforms full-model private adaptation across privacy budgets and client heterogeneity, while reducing protected uplink by a factor of 2.67 at \(\varepsilon=16\) on CIFAR-10 with comparable future-client accuracy. --- ### Vulnerability Localization Benchmark: Measuring Agentic Security Analysis at Repository Scale Authors: Aman Priyanshu, Supriti Vijay, Kimia Majd, Xuhong He, Fraser Burch, Takahiro Matsumoto, Jianliang He, Baturay Saglam, Arthur Goldblatt, Zhuoran Yang, Amin Karbasi Published: 2026-09-14 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.15939v1 Language-model agents increasingly operate over complete software repositories, yet cybersecurity evaluations primarily measure whether they can detect, reproduce, or repair vulnerabilities rather than whether they can locate the relevant code. We study vulnerability localization: given a weakness class and an unfamiliar repository, identify the implementation files associated with that weakness. We introduce the Vulnerability Localization Benchmark (VLoc Bench), comprising 500 real world vulnerabilities from 290 repositories across six package ecosystems and 147 CWE categories. Each task pairs repository snapshots immediately before and after a security fix. On the vulnerable snapshot, an agent receives only the CWE description and read-only terminal access and must return the affected files; on the patched snapshot, it must determine that the recorded vulnerability is no longer present. We evaluate 27 language models and four static-analysis tools under a common agent interface. Repository-scale vulnerability localization remains difficult: the strongest system achieves 0.229 File F1, and 38.4% of tasks receive no correct localization from any evaluated model. We further find that stronger localization does not imply reliable behavior after remediation: systems that identify vulnerable files effectively can still report unsupported locations on patched repositories. These results establish vulnerability localization as a distinct repository-scale capability and provide a setting for studying both how security agents search for vulnerable code and when they should refrain from reporting it. --- ### HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses Authors: Jieyuan Liu, Mengzhou Hu, Jefferson Chen, JungHo Kong, Pratibha Jagannatha, Yiming Gao, Dexter Pratt, Hsin-Yuan Lee, Zhiting Hu, Trey Ideker, Wei Wang, Eric P. Xing et al. Published: 2026-09-14 Categories: cs.CL, cs.CE, cs.MA, cs.NE arXiv: https://arxiv.org/abs/2609.15938v1 Scientific agents contribute to hypothesis discovery by synthesizing evidence, assessing proposals, and developing new explanations. Recent systems combine scientific agents with evolutionary search through critique, comparison, and revision. However, how different forms of agent collaboration affect hypothesis quality remains an open question. Answering this question requires separating the effects of agents' scientific capabilities from those of their collaboration. A framework must therefore preserve agents' scientific roles and support rules for combining, revising, and retaining hypotheses. Building on this view, we introduce HypoEvolve, which makes collaboration explicit through successive updates to a hypothesis population. Specifically, we propose a generational genetic algorithm to coordinate specialized large language model (LLM) agents that integrate mechanistic arguments, reconsider assumptions, and assess evidence and testability. Each generation specifies how scientific judgments and new proposals reshape the population, making collaboration effects on hypothesis quality directly testable. Moreover, we design our evaluation around scientifically meaningful hypotheses that explain how a proposed intervention could work. Drug repurposing links these explanations to target-level biological claims assessed against external evidence. Specifically, we adapt DepMap and Open Targets into complementary external measures grounded in experimental, genetic, and clinical evidence. Across 34 cancer types, HypoEvolve achieves the highest scores against six baselines on both measures. DepMap selectivity reaches 0.171, versus 0.115 for the strongest baseline. Gains over single-pass generation also generalize to held-out cancer types. HypoEvolve advances a vision of autonomous science in which AI research teams achieve a capacity for discovery beyond that of individual models. --- ### Recurrent Graph Neural Networks with Set-Based Aggregation Authors: Blai Bonet Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.15932v1 Recurrent GNNs iterate message passing to convergence, and their logical characterizations to date rely on multi-set aggregation, graded (counting) logics, and halting or acceptance conditions that cannot be verified from the network's parameters. We study recurrent GNNs with set-based aggregation and identify sufficient conditions checkable from the weights for networks to compile into formulas and formulas into networks. The main result is an effective, two-directional equivalence between a class of networks and the Boolean closure of reachability and safety properties, the fragment B$Σ^{\circ}_1$ of the modal $μ$-calculus. The fragment is not an artifact: it is the exact expressive level of stabilization over finite vocabulary, which supports fixed points of a single polarity and Boolean combinations thereof, but not the composition of fixed points of opposite polarities. The correspondence needs no counting logic, no external halting signal, and no non-effective acceptance condition, yielding a verifiable path from weights to symbolic explanations for networks meeting the conditions. --- ### Pilot Early, Commit Late: A Real-Options Model of Enterprise AI Adoption under Rapid Technological Progress Authors: Gaurav Tewari Published: 2026-09-14 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.15919v1 Artificial intelligence presents firms with an unusual timing problem. The technology frontier is improving rapidly, implementation is partly irreversible, and organization-specific capabilities are accumulated through action. This paper develops a two-period decision model of AI deployment under uncertainty in which a firm chooses among immediate deployment, a limited pilot, and waiting. Deployment earns current operating value but exposes the firm to architectural obsolescence; waiting preserves the option to adopt after the frontier is observed; a pilot sacrifices current operating value to build organization-specific learning without full commitment. The model yields five central timing results and a sixth comparative result on where learning occurs. First, a mean-preserving increase in frontier uncertainty raises the value of waiting and piloting but leaves immediate deployment unchanged when its payoff is affine in the frontier. Second, faster expected frontier progress can reduce the relative attractiveness of immediate deployment when deployed architecture captures only a limited share of future improvement. Third, a pilot dominates waiting exactly when the expected value of the capability it builds exceeds its cost. Fourth, sufficiently valuable organization-specific learning creates a nonempty region in which "pilot early, commit late" is optimal. Fifth, there is a closed-form modularity threshold above which immediate deployment dominates the best outside option. Sixth, production learning and pilot-specific learning affect the timing margin differently. A continuous-time extension recovers the standard result that uncertainty raises the adoption threshold while capability and modularity lower it. The paper separates deploying, experimenting, and waiting, and shows why rapid progress can rationally increase experimentation without justifying irreversible commitment. --- ### Safe Meta-Reinforcement Learning via Information Space Reachability Authors: Zeyang Li, Sunbochen Tang, Navid Azizan Published: 2026-09-14 Categories: cs.LG, eess.SY arXiv: https://arxiv.org/abs/2609.15915v1 Meta-reinforcement learning (meta-RL) enables agents to adapt to unseen tasks with limited experience. Despite its promise, the application of meta-RL in real-world tasks is hindered by safety requirements, which have been underexplored in prior work. In this paper, we propose a safe meta-RL framework that explicitly accounts for safety during adaptation. Our key insight is to reason about safety in the information space, which captures both the physical state and the agent's belief over the underlying task. Within this space, we introduce a safety value function that measures the probability of the agent avoiding unsafe regions indefinitely. We show that this function satisfies a self-consistency condition and a Bellman equation, which make it learnable via meta-RL. Based on this formulation, we develop a safe meta-RL algorithm that learns the safety value function and leverages it for safety filtering and constrained policy optimization. Experiments on meta-RL benchmarks demonstrate the effectiveness of the proposed method. --- ### SlipSense: Multimodal Tactile Learning for Low-Latency and Generalized Slip Detection Authors: Tong Jian, Aditya Thurvas Senthil Kumar, Xinyi Li, Ziling Chen, Tianyu Dai, Ali Sengul, Matteo Grimaldi, Wenjie Lu, Saleh Nabi, Tao Yu Published: 2026-09-14 Categories: cs.RO, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.15910v1 Slip detection is fundamental to dexterous manipulation, yet existing systems often lack precise characterization of detection latency and cross-platform generalization. We present SlipSense, a multimodal tactile slip-detection framework built on TacV5, a compact sensor integrating a $32 \times 32$ piezoresistive array operating at 240 Hz and a 3-axis MEMS accelerometer operating at 8 kHz. The piezoresistive array captures spatial pressure distributions, while the accelerometer captures friction-induced vibrations, providing complementary slip cues. The framework performs modality-specific encoding, intra-sensor fusion, and cross-modal attention with causal temporal prediction at 240 Hz. Experiments on a dataset of 1.4 million frames spanning 37 objects demonstrate the complementarity of the two modalities. SlipSense achieves 96.7% Macro F1 with a false-positive rate below 1.6%, detecting 76% of slip events within 23.1 ms. When trained solely on UMI data, SlipSense generalizes zero-shot to a Tesollo dexterous hand, transferring across unseen objects, distinct sensor units, and robotic platforms without retraining. --- ### Discrete Beckmann Transport Models for One-Step Language Modeling and Reasoning Authors: Sophia Tang, Shiyi Wang Published: 2026-09-14 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.15903v2 Discrete diffusion and flow models are a promising alternative to autoregressive language models, but compressing many-step sampling into fewer steps typically requires distilling a pretrained teacher model. This caps the student at the teacher's quality and requires a costly two-stage training pipeline. We introduce Discrete Beckmann Transport Models (DBTM), built on a time-independent flow whose autonomous transport map provably carries any point in the ambient space to a fixed point on the vertices of the simplex in a single step. We show that this fixed-point property is characterized by a conservation equation whose residual can be minimized directly from data, removing the requirement for a teacher flow and time conditioning. Under this construction, a partially trained map corresponds to the flow truncated at finite time, so generation reduces to iterating one map until it reaches a fixed point. We further extend the map to a partial-context interpolant where additional function evaluations act as refinement steps rather than ODE integration steps. On language modeling and reasoning tasks, DBTM enables one- and few-step generation that improves quality and accuracy over discrete diffusion and continuous flow baselines. --- ### Bridging Control, Inference, Transport, and Thermodynamics: From Theory to Applications in Learning Authors: Emmy Blumenthal, Nikolas Claussen, Benjamin Eysenbach, Catherine Ji, Gautam Reddy, Colin Scheibner, Benjamin Sorkin Published: 2026-09-14 Categories: cond-mat.stat-mech, cond-mat.dis-nn, cond-mat.soft, cs.LG, physics.bio-ph arXiv: https://arxiv.org/abs/2609.15897v1 The last decade has seen the development of powerful methods for learning complex structure from high-dimensional data. These advances have brought to the foreground fundamental connections between subdisciplines of physics, applied mathematics, and machine learning. In this review, we bring together some of these ideas, often expressed in different languages, to highlight a conceptual thread that links five distinct fields: control theory, optimal transport, probabilistic inference, non-equilibrium thermodynamics, and machine learning. A common theme is the optimization of free-energy-like functionals under dynamical or statistical constraints. We offer a guided tour through this thread and present selected applications in reinforcement learning, variational inference, and generative modeling. The review does not assume prior familiarity with these topics, and begins with principles originating from physics. --- ### Quenched Ensemble Sampling Authors: David Yallup Published: 2026-09-14 Categories: stat.ML, cs.LG, stat.CO arXiv: https://arxiv.org/abs/2609.15894v1 Some of the sharpest challenges in sampling from the energy functions of physical systems arise at phase transitions, where the density of states changes abruptly and many sampling algorithms stall. Nested sampling is a particle method that traverses the density of states under a hard energy constraint and is known to be robust to such transitions, but its application in high dimension is limited by the difficulty of sampling under that constraint. In this work we introduce Quenched Ensemble Sampling, which generalises the hard constraint to a family of repulsive potentials at the energy boundary. This preserves the quenched path of monotonically decreasing energy while making the constrained target amenable to scalable gradient-based kernels. We demonstrate on synthetic models of phase transitions that our method estimates the marginal likelihood and draws posterior samples across a first-order transition where popular alternatives such as tempering fail. We apply the procedure to marginal likelihood estimation in Bayesian neural networks, enabling model comparison between network architectures. Finally, in a high-dimensional continuous lattice field theory, we show that this method traverses a first-order transition and estimates the partition function. ---