# Research Papers — 2026-02 · bundle page 167 of 210

> Full-text markdown of papers #8301–8350 of 10493 published in 2026-02, newest first. AI-readable mirrors of the full 10493-paper 2026-02 archive (50 papers per file).

Canonical page: https://aipod.dev/papers/archive/2026-02/page/167
[← Previous bundle](https://aipod.dev/papers/archive/2026-02/page/166.md)
[Next bundle →](https://aipod.dev/papers/archive/2026-02/page/168.md)

### Aligning Large Language Model Behavior with Human Citation Preferences

Authors: Kenichiro Ando, Tatsuya Harada
Published: 2026-02-05
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2602.05205v1

Most services built on powerful large-scale language models (LLMs) add citations to their output to enhance credibility. Recent research has paid increasing attention to the question of what reference documents to link to outputs. However, how LLMs recognize cite-worthiness and how this process should be controlled remains underexplored. In this study, we focus on what kinds of content LLMs currently tend to cite and how well that behavior aligns with human preferences. We construct a dataset to characterize the relationship between human citation preferences and LLM behavior. Web-derived texts are categorized into eight citation-motivation types, and pairwise citation preferences are exhaustively evaluated across all type combinations to capture fine-grained contrasts. Our results show that humans most frequently seek citations for medical text, and stronger models display a similar tendency. We also find that current models are as much as $27\%$ more likely than humans to add citations to text that is explicitly marked as needing citations on sources such as Wikipedia, and this overemphasis reduces alignment accuracy. Conversely, models systematically underselect numeric sentences (by $-22.6\%$ relative to humans) and sentences containing personal names (by $-20.1\%$), categories for which humans typically demand citations. Furthermore, experiments with Direct Preference Optimization demonstrate that model behavior can be calibrated to better match human citation preferences. We expect this study to provide a foundation for more fine-grained investigations into LLM citation preferences.

---

### Extreme Weather Nowcasting via Local Precipitation Pattern Prediction

Authors: Changhoon Song, Teng Yuan Chang, Youngjoon Hong
Published: 2026-02-05
Categories: cs.LG, cs.CV
arXiv: https://arxiv.org/abs/2602.05204v2

Accurate forecasting of extreme weather events such as heavy rainfall or storms is critical for risk management and disaster mitigation. Although high-resolution radar observations have spurred extensive research on nowcasting models, precipitation nowcasting remains particularly challenging due to pronounced spatial locality, intricate fine-scale rainfall structures, and variability in forecasting horizons. While recent diffusion-based generative ensembles show promising results, they are computationally expensive and unsuitable for real-time applications. In contrast, deterministic models are computationally efficient but remain biased toward normal rainfall. Furthermore, the benchmark datasets commonly used in prior studies are themselves skewed--either dominated by ordinary rainfall events or restricted to extreme rainfall episodes--thereby hindering general applicability in real-world settings. In this paper, we propose exPreCast, an efficient deterministic framework for generating finely detailed radar forecasts, and introduce a newly constructed balanced radar dataset from the Korea Meteorological Administration (KMA), which encompasses both ordinary precipitation and extreme events. Our model integrates local spatiotemporal attention, a texture-preserving cubic dual upsampling decoder, and a temporal extractor to flexibly adjust forecasting horizons. Experiments on established benchmarks (SEVIR and MeteoNet) as well as on the balanced KMA dataset demonstrate that our approach achieves state-of-the-art performance, delivering accurate and reliable nowcasts across both normal and extreme rainfall regimes.

---

### GT-SVJ: Generative-Transformer-Based Self-Supervised Video Judge For Efficient Video Reward Modeling

Authors: Shivanshu Shekhar, Uttaran Bhattacharya, Raghavendra Addanki, Mehrab Tanjim, Somdeb Sarkhel, Tong Zhang
Published: 2026-02-05
Categories: cs.CV
arXiv: https://arxiv.org/abs/2602.05202v2

Aligning video generative models with human preferences remains challenging: current approaches rely on Vision-Language Models (VLMs) for reward modeling, but these models struggle to capture subtle temporal dynamics. We propose a fundamentally different approach: repurposing video generative models, which are inherently designed to model temporal structure, as reward models. We present the Generative-Transformer-based Self-Supervised Video Judge (\modelname), a novel evaluation model that transforms state-of-the-art video generation models into powerful temporally-aware reward models. Our key insight is that generative models can be reformulated as energy-based models (EBMs) that assign low energy to high-quality videos and high energy to degraded ones, enabling them to discriminate video quality with remarkable precision when trained via contrastive objectives. To prevent the model from exploiting superficial differences between real and generated videos, we design challenging synthetic negative videos through controlled latent-space perturbations: temporal slicing, feature swapping, and frame shuffling, which simulate realistic but subtle visual degradations. This forces the model to learn meaningful spatiotemporal features rather than trivial artifacts. \modelname achieves state-of-the-art performance on GenAI-Bench and MonteBench using only 30K human-annotations: $6\times$ to $65\times$ fewer than existing VLM-based approaches.

---

### Traceable Cross-Source RAG for Chinese Tibetan Medicine Question Answering

Authors: Fengxian Chen, Zhilong Tao, Jiaxuan Li, Yunlong Li, Qingguo Zhou
Published: 2026-02-05
Categories: cs.AI
arXiv: https://arxiv.org/abs/2602.05195v2

Retrieval-augmented generation (RAG) promises grounded question answering, yet domain settings with multiple heterogeneous knowledge bases (KBs) remain challenging. In Chinese Tibetan medicine, encyclopedia entries are often dense and easy to match, which can dominate retrieval even when classics or clinical papers provide more authoritative evidence. We study a practical setting with three KBs (encyclopedia, classics, and clinical papers) and a 500-query benchmark (cutoff $K{=}5$) covering both single-KB and cross-KB questions. We propose two complementary methods to improve traceability, reduce hallucinations, and enable cross-KB verification. First, DAKS performs KB routing and budgeted retrieval to mitigate density-driven bias and to prioritize authoritative sources when appropriate. Second, we use an alignment graph to guide evidence fusion and coverage-aware packing, improving cross-KB evidence coverage without relying on naive concatenation. All answers are generated by a lightweight generator, \textsc{openPangu-Embedded-7B}. Experiments show consistent gains in routing quality and cross-KB evidence coverage, with the full system achieving the best CrossEv@5 while maintaining strong faithfulness and citation correctness.

---

### First Proof

Authors: Mohammed Abouzaid, Andrew J. Blumberg, Martin Hairer, Joe Kileel, Tamara G. Kolda, Paul D. Nelson, Daniel Spielman, Nikhil Srivastava, Rachel Ward, Shmuel Weinberger, Lauren Williams
Published: 2026-02-05
Categories: cs.AI, math.AG, math.CO, math.GT, math.HO, math.RA
arXiv: https://arxiv.org/abs/2602.05192v2

To assess the ability of current AI systems to correctly answer research-level mathematics questions, we share a set of ten math questions which have arisen naturally in the research process of the authors. The questions had not been shared publicly until now; the answers are known to the authors of the questions but will remain encrypted for a short time.

---

### Double-P: Hierarchical Top-P Sparse Attention for Long-Context LLMs

Authors: Wentao Ni, Kangqi Zhang, Zhongming Yu, Oren Nelson, Mingu Lee, Hong Cai, Fatih Porikli, Jongryool Kim, Zhijian Liu, Jishen Zhao
Published: 2026-02-05
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2602.05191v1

As long-context inference becomes central to large language models (LLMs), attention over growing key-value caches emerges as a dominant decoding bottleneck, motivating sparse attention for scalable inference. Fixed-budget top-k sparse attention cannot adapt to heterogeneous attention distributions across heads and layers, whereas top-p sparse attention directly preserves attention mass and provides stronger accuracy guarantees. Existing top-p methods, however, fail to jointly optimize top-p accuracy, selection overhead, and sparse attention cost, which limits their overall efficiency. We present Double-P, a hierarchical sparse attention framework that optimizes all three stages. Double-P first performs coarse-grained top-p estimation at the cluster level using size-weighted centroids, then adaptively refines computation through a second top-p stage that allocates token-level attention only when needed. Across long-context benchmarks, Double-P consistently achieves near-zero accuracy drop, reducing attention computation overhead by up to 1.8x and delivers up to 1.3x end-to-end decoding speedup over state-of-the-art fixed-budget sparse attention methods.

---

### PoseGaussian: Pose-Driven Novel View Synthesis for Robust 3D Human Reconstruction

Authors: Ju Shen, Chen Chen, Tam V. Nguyen, Vijayan K. Asari
Published: 2026-02-05
Categories: cs.CV, cs.GR
arXiv: https://arxiv.org/abs/2602.05190v1

We propose PoseGaussian, a pose-guided Gaussian Splatting framework for high-fidelity human novel view synthesis. Human body pose serves a dual purpose in our design: as a structural prior, it is fused with a color encoder to refine depth estimation; as a temporal cue, it is processed by a dedicated pose encoder to enhance temporal consistency across frames. These components are integrated into a fully differentiable, end-to-end trainable pipeline. Unlike prior works that use pose only as a condition or for warping, PoseGaussian embeds pose signals into both geometric and temporal stages to improve robustness and generalization. It is specifically designed to address challenges inherent in dynamic human scenes, such as articulated motion and severe self-occlusion. Notably, our framework achieves real-time rendering at 100 FPS, maintaining the efficiency of standard Gaussian Splatting pipelines. We validate our approach on ZJU-MoCap, THuman2.0, and in-house datasets, demonstrating state-of-the-art performance in perceptual quality and structural accuracy (PSNR 30.86, SSIM 0.979, LPIPS 0.028).

---

### Are Open-Weight LLMs Ready for Social Media Moderation? A Comparative Study on Bluesky

Authors: Hsuan-Yu Chou, Wajiha Naveed, Shuyan Zhou, Xiaowei Yang
Published: 2026-02-05
Categories: cs.CL, cs.HC, cs.LG, cs.SI
arXiv: https://arxiv.org/abs/2602.05189v1

As internet access expands, so does exposure to harmful content, increasing the need for effective moderation. Research has demonstrated that large language models (LLMs) can be effectively utilized for social media moderation tasks, including harmful content detection. While proprietary LLMs have been shown to zero-shot outperform traditional machine learning models, the out-of-the-box capability of open-weight LLMs remains an open question. Motivated by recent developments of reasoning LLMs, we evaluate seven state-of-the-art models: four proprietary and three open-weight. Testing with real-world posts on Bluesky, moderation decisions by Bluesky Moderation Service, and annotations by two authors, we find a considerable degree of overlap between the sensitivity (81%--97%) and specificity (91%--100%) of the open-weight LLMs and those (72%--98%, and 93%--99%) of the proprietary ones. Additionally, our analysis reveals that specificity exceeds sensitivity for rudeness detection, but the opposite holds for intolerance and threats. Lastly, we identify inter-rater agreement across human moderators and the LLMs, highlighting considerations for deploying LLMs in both platform-scale and personalized moderation contexts. These findings show open-weight LLMs can support privacy-preserving moderation on consumer-grade hardware and suggest new directions for designing moderation systems that balance community values with individual user preferences.

---

### SpectraKAN: Conditioning Spectral Operators

Authors: Chun-Wun Cheng, Carola-Bibiane Schönlieb, Angelica I. Aviles-Rivero
Published: 2026-02-05
Categories: cs.LG, math.NA
arXiv: https://arxiv.org/abs/2602.05187v1

Spectral neural operators, particularly Fourier Neural Operators (FNO), are a powerful framework for learning solution operators of partial differential equations (PDEs) due to their efficient global mixing in the frequency domain. However, existing spectral operators rely on static Fourier kernels applied uniformly across inputs, limiting their ability to capture multi-scale, regime-dependent, and anisotropic dynamics governed by the global state of the system. We introduce SpectraKAN, a neural operator that conditions the spectral operator on the input itself, turning static spectral convolution into an input-conditioned integral operator. This is achieved by extracting a compact global representation from spatio-temporal history and using it to modulate a multi-scale Fourier trunk via single-query cross-attention, enabling the operator to adapt its behaviour while retaining the efficiency of spectral mixing. We provide theoretical justification showing that this modulation converges to a resolution-independent continuous operator under mesh refinement and KAN gives smooth, Lipschitz-controlled global modulation. Across diverse PDE benchmarks, SpectraKAN achieves state-of-the-art performance, reducing RMSE by up to 49% over strong baselines, with particularly large gains on challenging spatio-temporal prediction tasks.

---

### Towards Worst-Case Guarantees with Scale-Aware Interpretability

Authors: Lauren Greenspan, David Berman, Aryeh Brill, Ro Jefferson, Artemy Kolchinsky, Jennifer Lin, Andrew Mack, Anindita Maiti, Fernando E. Rosas, Alexander Stapleton, Lucas Teixeira, Dmitry Vaintrob
Published: 2026-02-05
Categories: hep-th, cond-mat.dis-nn, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2602.05184v1

Neural networks organize information according to the hierarchical, multi-scale structure of natural data. Methods to interpret model internals should be similarly scale-aware, explicitly tracking how features compose across resolutions and guaranteeing bounds on the influence of fine-grained structure that is discarded as irrelevant noise. We posit that the renormalisation framework from physics can meet this need by offering technical tools that can overcome limitations of current methods. Moreover, relevant work from adjacent fields has now matured to a point where scattered research threads can be synthesized into practical, theory-informed tools. To combine these threads in an AI safety context, we propose a unifying research agenda -- \emph{scale-aware interpretability} -- to develop formal machinery and interpretability tools that have robustness and faithfulness properties supported by statistical physics.

---

### Data-Centric Interpretability for LLM-based Multi-Agent Reinforcement Learning

Authors: John Yan, Michael Yu, Yuqi Sun, Alexander Duffy, Tyler Marques, Matthew Lyle Olson
Published: 2026-02-05
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2602.05183v2

Large language models (LLMs) are increasingly trained in complex Reinforcement Learning, multi-agent environments, making it difficult to understand how behavior changes over training. Sparse Autoencoders (SAEs) have recently shown to be useful for data-centric interpretability. In this work, we analyze large-scale reinforcement learning training runs from the sophisticated environment of Full-Press Diplomacy by applying pretrained SAEs, alongside LLM-summarizer methods. We introduce Meta-Autointerp, a method for grouping SAE features into interpretable hypotheses about training dynamics. We discover fine-grained behaviors including role-playing patterns, degenerate outputs, language switching, alongside high-level strategic behaviors and environment-specific bugs. Through automated evaluation, we validate that 90% of discovered SAE Meta-Features are significant, and find a surprising reward hacking behavior. However, through two user studies, we find that even subjectively interesting and seemingly helpful SAE features may be worse than useless to humans, along with most LLM generated hypotheses. However, a subset of SAE-derived hypotheses are predictively useful for downstream tasks. We further provide validation by augmenting an untrained agent's system prompt, improving the score by +14.2%. Overall, we show that SAEs and LLM-summarizer provide complementary views into agent behavior, and together our framework forms a practical starting point for future data-centric interpretability work on ensuring trustworthy LLM behavior throughout training.

---

### The Single-Multi Evolution Loop for Self-Improving Model Collaboration Systems

Authors: Shangbin Feng, Kishan Panaganti, Yulia Tsvetkov, Wenhao Yu
Published: 2026-02-05
Categories: cs.CL
arXiv: https://arxiv.org/abs/2602.05182v2

Model collaboration -- systems where multiple language models (LMs) collaborate -- combines the strengths of diverse models with cost in loading multiple LMs. We improve efficiency while preserving the strengths of collaboration by distilling collaborative patterns into a single model, where the model is trained on the outputs of the model collaboration system. At inference time, only the distilled model is employed: it imitates the collaboration while only incurring the cost of a single model. Furthermore, we propose the single-multi evolution loop: multiple LMs collaborate, each distills from the collaborative outputs, and these post-distillation improved LMs collaborate again, forming a collective evolution ecosystem where models evolve and self-improve by interacting with an environment of other models. Extensive experiments with 7 collaboration strategies and 15 tasks (QA, reasoning, factuality, etc.) demonstrate that: 1) individual models improve by 8.0% on average, absorbing the strengths of collaboration while reducing the cost to a single model; 2) the collaboration also benefits from the stronger and more synergistic LMs after distillation, improving over initial systems without evolution by 14.9% on average. Analysis reveals that the single-multi evolution loop outperforms various existing evolutionary AI methods, is compatible with diverse model/collaboration/distillation settings, and helps solve problems where the initial model/system struggles to.

---

### Benchmarking Artificial Intelligence Models for Daily Coastal Hypoxia Forecasting

Authors: Magesh Rajasekaran, Md Saiful Sajol, Chris Alvin, Supratik Mukhopadhyay, Yanda Ou, Z. George Xue
Published: 2026-02-05
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2602.05178v1

Coastal hypoxia, especially in the northern part of Gulf of Mexico, presents a persistent ecological and economic concern. Seasonal models offer coarse forecasts that miss the fine-scale variability needed for daily, responsive ecosystem management. We present study that compares four deep learning architectures for daily hypoxia classification: Bidirectional Long Short-Term Memory (BiLSTM), Medformer (Medical Transformer), Spatio-Temporal Transformer (ST-Transformer), and Temporal Convolutional Network (TCN). We trained our models with twelve years of daily hindcast data from 2009-2020 Our training data consists of 2009-2020 hindcast data from a coupled hydrodynamic-biogeochemical model. Similarly, we use hindcast data from 2020 through 2024 as a test data. We constructed classification models incorporating water column stratification, sediment oxygen consumption, and temperature-dependent decomposition rates. We evaluated each architectures using the same data preprocessing, input/output formulation, and validation protocols. Each model achieved high classification accuracy and strong discriminative ability with ST-Transformer achieving the highest performance across all metrics and tests periods (AUC-ROC: 0.982-0.992). We also employed McNemar's method to identify statistically significant differences in model predictions. Our contribution is a reproducible framework for operational real-time hypoxia prediction that can support broader efforts in the environmental and ocean modeling systems community and in ecosystem resilience. The source code is available https://github.com/rmagesh148/hypoxia-ai/

---

### Among Us: Measuring and Mitigating Malicious Contributions in Model Collaboration Systems

Authors: Ziyuan Yang, Wenxuan Ding, Shangbin Feng, Yulia Tsvetkov
Published: 2026-02-05
Categories: cs.CL
arXiv: https://arxiv.org/abs/2602.05176v1

Language models (LMs) are increasingly used in collaboration: multiple LMs trained by different parties collaborate through routing systems, multi-agent debate, model merging, and more. Critical safety risks remain in this decentralized paradigm: what if some of the models in multi-LLM systems are compromised or malicious? We first quantify the impact of malicious models by engineering four categories of malicious LMs, plug them into four types of popular model collaboration systems, and evaluate the compromised system across 10 datasets. We find that malicious models have a severe impact on the multi-LLM systems, especially for reasoning and safety domains where performance is lowered by 7.12% and 7.94% on average. We then propose mitigation strategies to alleviate the impact of malicious components, by employing external supervisors that oversee model collaboration to disable/mask them out to reduce their influence. On average, these strategies recover 95.31% of the initial performance, while making model collaboration systems fully resistant to malicious models remains an open research question.

---

### Enhancing Adversarial Robustness with Signed Distance Fields for Harmonizing Geometric Invariance and Texture

Authors: Zhe Li, Bernhard Kainz
Published: 2026-02-05
Categories: cs.CV
arXiv: https://arxiv.org/abs/2602.05175v2

Deep neural networks demonstrate impressive performance in visual recognition but remain highly vulnerable to imperceptible adversarial attacks. Existing defense strategies such as adversarial training and diffusion-based purification have achieved significant progress but are frequently constrained by high computational cost, information loss, and inference latency. To address these challenges, we propose a Geometric and Texture balancing Purification (GeoTexPuri) framework that enhances adversarial robustness by harmonizing invariant geometric structures with textural features. Specifically, the framework integrates dense geometric guidance into the training phase by transforming discrete image masks into continuous spatial fields via Signed Distance Fields (SDF). This process establishes stable structural anchors that shield the model from local pixel noise. Through a multi-stream training objective, the model learns to internalize purified representations that effectively align semantic textural cues with these underlying geometric invariants. Extensive experiments on ImageNet demonstrate the efficacy of our approach. GeoTexPuri achieves 84.79\% clean accuracy and 83.52\% robust accuracy under the AutoAttack. Crucially, GeoTexPuri functions as a deterministic classifier during inference, requiring only the input image without any auxiliary geometric modules or additional computational costs, thereby ensuring a scalable and efficient solution for real-time applications.

---

### Total Variation Rates for Riemannian Flow Matching

Authors: Yunrui Guan, Krishnakumar Balasubramanian, Shiqian Ma
Published: 2026-02-05
Categories: stat.ML, cs.AI, cs.LG, math.ST
arXiv: https://arxiv.org/abs/2602.05174v1

Riemannian flow matching (RFM) extends flow-based generative modeling to data supported on manifolds by learning a time-dependent tangent vector field whose flow-ODE transports a simple base distribution to the data law. We develop a nonasymptotic Total Variation (TV) convergence analysis for RFM samplers that use a learned vector field together with Euler discretization on manifolds. Our key technical ingredient is a differential inequality governing the evolution of TV between two manifold ODE flows, which expresses the time-derivative of TV through the divergence of the vector-field mismatch and the score of the reference flow; controlling these terms requires establishing new bounds that explicitly account for parallel transport and curvature. Under smoothness assumptions on the population flow-matching field and either uniform (compact manifolds) or mean-square (Hadamard manifolds) approximation guarantees for the learned field, we obtain explicit bounds of the form $\mathrm{TV}\le C_{\mathrm{Lip}}\,h + C_{\varepsilon}\,\varepsilon$ (with an additional higher-order $\varepsilon^2$ term on compact manifolds), cleanly separating numerical discretization and learning errors. Here, $h$ is the step-size and $\varepsilon$ is the target accuracy. Instantiations yield \emph{explicit} polynomial iteration complexities on the hypersphere $S^d$, and on the SPD$(n)$ manifolds under mild moment conditions.

---

### Finite-Particle Rates for Regularized Stein Variational Gradient Descent

Authors: Ye He, Krishnakumar Balasubramanian, Sayan Banerjee, Promit Ghosal
Published: 2026-02-05
Categories: stat.ML, cs.LG, math.ST
arXiv: https://arxiv.org/abs/2602.05172v2

We derive finite-particle rates for the regularized Stein variational gradient descent (R-SVGD) algorithm introduced by He et al. (2024) that corrects the constant-order bias of the SVGD by applying a resolvent-type preconditioner to the kernelized Wasserstein gradient. For the resulting interacting $N$-particle system, we establish explicit non-asymptotic bounds for time-averaged (annealed) empirical measures, illustrating convergence in the \emph{true} (non-kernelized) Fisher information and, under a $\mathrm{W}_1\mathrm{I}$ condition on the target, corresponding $\mathrm{W}_1$ convergence for a large class of smooth kernels. Our analysis covers both continuous- and discrete-time dynamics and yields principled tuning rules for the regularization parameter, step size, and averaging horizon that quantify the trade-off between approximating the Wasserstein gradient flow and controlling finite-particle estimation error.

---

### Path Sampling for Rare Events Boosted by Machine Learning

Authors: Porhouy Minh, Sapna Sarupria
Published: 2026-02-05
Categories: physics.comp-ph, cond-mat.soft, cond-mat.stat-mech, cs.LG, physics.chem-ph
arXiv: https://arxiv.org/abs/2602.05167v1

The study by Jung et al. (Jung H, Covino R, Arjun A, et al., Nat Comput Sci. 3:334-345 (2023)) introduced Artificial Intelligence for Molecular Mechanism Discovery (AIMMD), a novel sampling algorithm that integrates machine learning to enhance the efficiency of transition path sampling (TPS). By enabling on-the-fly estimation of the committor probability and simultaneously deriving a human-interpretable reaction coordinate, AIMMD offers a robust framework for elucidating the mechanistic pathways of complex molecular processes. This commentary provides a discussion and critical analysis of the core AIMMD framework, explores its recent extensions, and offers an assessment of the method's potential impact and limitations.

---

### EBPO: Empirical Bayes Shrinkage for Stabilizing Group-Relative Policy Optimization

Authors: Kevin Han, Yuhang Zhou, Mingze Gao, Gedi Zhou, Serena Li, Abhishek Kumar, Xiangjun Fan, Weiwei Li, Lizhu Zhang
Published: 2026-02-05
Categories: cs.LG, cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2602.05165v3

Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for enhancing the reasoning capabilities of Large Language Models (LLMs). However, dominant approaches like Group Relative Policy Optimization (GRPO) face critical stability challenges: they suffer from high estimator variance under computational constraints (small group sizes) and vanishing gradient signals in saturated failure regimes where all responses yield identical zero rewards. To address this, we propose Empirical Bayes Policy Optimization (EBPO), a novel framework that regularizes local group-based baselines by borrowing strength from the policy's accumulated global statistics. Instead of estimating baselines in isolation, EBPO employs a shrinkage estimator that dynamically balances local group statistics with a global prior updated via Welford's online algorithm. Theoretically, we demonstrate that EBPO guarantees strictly lower Mean Squared Error (MSE), bounded entropy decay, and non-vanishing penalty signals in failure scenarios compared to GRPO. Empirically, EBPO consistently outperforms GRPO and other established baselines across diverse benchmarks, including AIME and OlympiadBench. Notably, EBPO exhibits superior training stability, achieving high-performance gains even with small group sizes, and benefits significantly from difficulty-stratified curriculum learning.

---

### Position: Capability Control Should be a Separate Goal From Alignment

Authors: Shoaib Ahmed Siddiqui, Eleni Triantafillou, David Krueger, Adrian Weller
Published: 2026-02-05
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2602.05164v1

Foundation models are trained on broad data distributions, yielding generalist capabilities that enable many downstream applications but also expand the space of potential misuse and failures. This position paper argues that capability control -- imposing restrictions on permissible model behavior -- should be treated as a distinct goal from alignment. While alignment is often context and preference-driven, capability control aims to impose hard operational limits on permissible behaviors, including under adversarial elicitation. We organize capability control mechanisms across the model lifecycle into three layers: (i) data-based control of the training distribution, (ii) learning-based control via weight- or representation-level interventions, and (iii) system-based control via post-deployment guardrails over inputs, outputs, and actions. Because each layer has characteristic failure modes when used in isolation, we advocate for a defense-in-depth approach that composes complementary controls across the full stack. We further outline key open challenges in achieving such control, including the dual-use nature of knowledge and compositional generalization.

---

### LOBSTgER-enhance: an underwater image enhancement pipeline

Authors: Andreas Mentzelopoulos, Keith Ellenbogen
Published: 2026-02-05
Categories: cs.CV
arXiv: https://arxiv.org/abs/2602.05163v1

Underwater photography presents significant inherent challenges including reduced contrast, spatial blur, and wavelength-dependent color distortions. These effects can obscure the vibrancy of marine life and awareness photographers in particular are often challenged with heavy post-processing pipelines to correct for these distortions. We develop an image-to-image pipeline that learns to reverse underwater degradations by introducing a synthetic corruption pipeline and learning to reverse its effects with diffusion-based generation. Training and evaluation are performed on a small high-quality dataset of awareness photography images by Keith Ellenbogen. The proposed methodology achieves high perceptual consistency and strong generalization in synthesizing 512x768 images using a model of ~11M parameters after training from scratch on ~2.5k images.

---

### SHaSaM: Submodular Hard Sample Mining for Fair Facial Attribute Recognition

Authors: Anay Majee, Rishabh Iyer
Published: 2026-02-05
Categories: cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2602.05162v1

Deep neural networks often inherit social and demographic biases from annotated data during model training, leading to unfair predictions, especially in the presence of sensitive attributes like race, age, gender etc. Existing methods fall prey to the inherent data imbalance between attribute groups and inadvertently emphasize on sensitive attributes, worsening unfairness and performance. To surmount these challenges, we propose SHaSaM (Submodular Hard Sample Mining), a novel combinatorial approach that models fairness-driven representation learning as a submodular hard-sample mining problem. Our two-stage approach comprises of SHaSaM-MINE, which introduces a submodular subset selection strategy to mine hard positives and negatives - effectively mitigating data imbalance, and SHaSaM-LEARN, which introduces a family of combinatorial loss functions based on Submodular Conditional Mutual Information to maximize the decision boundary between target classes while minimizing the influence of sensitive attributes. This unified formulation restricts the model from learning features tied to sensitive attributes, significantly enhancing fairness without sacrificing performance. Experiments on CelebA and UTKFace demonstrate that SHaSaM achieves state-of-the-art results, with up to 2.7 points improvement in model fairness (Equalized Odds) and a 3.5% gain in Accuracy, within fewer epochs as compared to existing methods.

---

### AirGlove: Exploring Egocentric 3D Hand Tracking and Appearance Generalization for Sensing Gloves

Authors: Wenhui Cui, Ziyi Kou, Chuan Qin, Ergys Ristani, Li Guan
Published: 2026-02-05
Categories: cs.CV
arXiv: https://arxiv.org/abs/2602.05159v1

Sensing gloves have become important tools for teleoperation and robotic policy learning as they are able to provide rich signals like speed, acceleration and tactile feedback. A common approach to track gloved hands is to directly use the sensor signals (e.g., angular velocity, gravity orientation) to estimate 3D hand poses. However, sensor-based tracking can be restrictive in practice as the accuracy is often impacted by sensor signal and calibration quality. Recent advances in vision-based approaches have achieved strong performance on human hands via large-scale pre-training, but their performance on gloved hands with distinct visual appearances remains underexplored. In this work, we present the first systematic evaluation of vision-based hand tracking models on gloved hands under both zero-shot and fine-tuning setups. Our analysis shows that existing bare-hand models suffer from substantial performance degradation on sensing gloves due to large appearance gap between bare-hand and glove designs. We therefore propose AirGlove, which leverages existing gloves to generalize the learned glove representations towards new gloves with limited data. Experiments with multiple sensing gloves show that AirGlove effectively generalizes the hand pose models to new glove designs and achieves a significant performance boost over the compared schemes.

---

### GreekMMLU: A Native-Sourced Multitask Benchmark for Evaluating Language Models in Greek

Authors: Yang Zhang, Mersin Konomi, Christos Xypolopoulos, Konstantinos Divriotis, Konstantinos Skianis, Giannis Nikolentzos, Giorgos Stamou, Guokan Shang, Michalis Vazirgiannis
Published: 2026-02-05
Categories: cs.CL
arXiv: https://arxiv.org/abs/2602.05150v2

Large Language Models (LLMs) are commonly trained on multilingual corpora that include Greek, yet reliable evaluation benchmarks for Greek-particularly those based on authentic, native-sourced content-remain limited. Existing datasets are often machine-translated from English, failing to capture Greek linguistic and cultural characteristics. We introduce GreekMMLU, a native-sourced benchmark for massive multitask language understanding in Greek, comprising 21,805 multiple-choice questions across 45 subject areas, organized under a newly defined subject taxonomy and annotated with educational difficulty levels spanning primary to professional examinations. All questions are sourced or authored in Greek from academic, professional, and governmental exams. We publicly release 16,857 samples and reserve 4,948 samples for a private leaderboard to enable robust and contamination-resistant evaluation. Evaluations of over 80 open- and closed-source LLMs reveal substantial performance gaps between frontier and open-weight models, as well as between Greek-adapted models and general multilingual ones. Finally, we provide a systematic analysis of factors influencing performance-including model scale, adaptation, and prompting-and derive insights for improving LLM capabilities in Greek.

---

### CoSA: Compressed Sensing-Based Adaptation of Large Language Models

Authors: Songtao Wei, Yi Li, Bohan Zhang, Zhichun Guo, Ying Huang, Yuede Ji, Miao Yin, Guanpeng Li, Bingzhe Li
Published: 2026-02-05
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2602.05148v2

Parameter-Efficient Fine-Tuning (PEFT) has emerged as a practical paradigm for adapting large language models (LLMs) without updating all parameters. Most existing approaches, such as LoRA and PiSSA, rely on low-rank decompositions of weight updates. However, the low-rank assumption may restrict expressivity, particularly in task-specific adaptation scenarios where singular values are distributed relatively uniformly. To address this limitation, we propose CoSA (Compressed Sensing-Based Adaptation), a new PEFT method extended from compressed sensing theory. Instead of constraining weight updates to a low-rank subspace, CoSA expresses them through fixed random projection matrices and a compact learnable core. We provide a formal theoretical analysis of CoSA as a synthesis process, proving that weight updates can be compactly encoded into a low-dimensional space and mapped back through random projections. Extensive experimental results show that CoSA provides a principled perspective for efficient and expressive multi-scale model adaptation. Specifically, we evaluate CoSA on 10 diverse tasks, including natural language understanding and generation, employing 5 models of different scales from RoBERTa, Llama, and Qwen families. Across these settings, CoSA consistently matches or outperforms state-of-the-art PEFT methods.

---

### Cross-talk based multi-task learning for fault classification of machine system influenced by multiple variables

Authors: Wonjun Yi, Rismaya Kumar Mishra, Yong-Hwa Park
Published: 2026-02-05
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2602.05146v2

Machine systems inherently generate signals in which fault conditions and various variables influence signals measured from machine system. Although many existing fault classification studies rely solely on direct fault labels, the aforementioned signals naturally embed additional information shaped by other variables. Herein, we leverage this through a multi-task learning (MTL) framework that jointly learns fault conditions and other variables influencing measured signals. Among MTL architectures, cross-talk structures have distinct advantages because they allow for controlled information exchange between tasks through the cross-talk layer while preventing negative transfer, in contrast to shared trunk architectures that often mix incompatible features. We build on our previously introduced residual neural dimension reductor model, and extend its application to two benchmarks where system influenced by multiple variables. The first benchmark is a drone fault dataset, in which machine type and maneuvering direction significantly alter the frequency components of measured signals even under the same drone status. The second benchmark dataset is motor compound fault dataset. In this system, severity of each fault component, inner race fault, outer race fault, misalignment, and unbalance influences measured signal. Across both benchmarks, our residual neural dimension reductor, consistently outperformed single-task models, multi-class models that merge all label combinations, and shared trunk multi-task models.

---

### TIDE: Temporal Incremental Draft Engine for Self-Improving LLM Inference

Authors: Jiyoung Park, Hankyu Jang, Changseok Song, Wookeun Jung
Published: 2026-02-05
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2602.05145v2

Speculative decoding can substantially accelerate LLM inference, but realizing its benefits in practice is challenging due to evolving workloads. We present TIDE (Temporal Incremental Draft Engine), a serving-engine-native framework that integrates online draft adaptation directly into high-performance LLM inference systems. TIDE reuses target model's intermediate hidden states generated during inference as training signals for draft adaptation, thereby avoiding additional target model computation and serving-time overhead. It employs adaptive runtime control to activate speculation and draft model training only when beneficial. TIDE exploits heterogeneous clusters by mapping inference and training to appropriate GPU classes. Across diverse real-world workloads, TIDE achieves up to 1.66$\times$ throughput over no-speculation baselines while recovering performance on misaligned workloads where static draft models degrade throughput. TIDE also reduces training time by up to 3.02$\times$ and storage requirements by 24$\times$ compared to existing draft training approaches, and improves system throughput by up to 1.22$\times$ on heterogeneous GPU clusters.

---

### TIDE: Temporal Incremental Draft Engine for Self-Improving LLM Inference

Authors: Jiyoung Park, Hankyu Jang, Changseok Song, Wookeun Jung
Published: 2026-02-05
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2602.05145v1

Speculative decoding can substantially accelerate LLM inference, but realizing its benefits in practice is challenging due to evolving workloads and system-level constraints. We present TIDE (Temporal Incremental Draft Engine), a serving-engine-native framework that integrates online draft adaptation directly into high-performance LLM inference systems. TIDE reuses target model hidden states generated during inference as training signals, enabling zero-overhead draft adaptation without reloading the target model, and employs adaptive runtime control to activate speculation and training only when beneficial. TIDE exploits heterogeneous clusters by mapping decoupled inference and training to appropriate GPU classes. Across diverse real-world workloads, TIDE achieves up to 1.15x throughput improvement over static speculative decoding while reducing draft training time by 1.67x compared to approaches that recompute training signals.

---

### Fairness Under Group-Conditional Prior Probability Shift: Invariance, Drift, and Target-Aware Post-Processing

Authors: Amir Asiaee, Kaveh Aryan
Published: 2026-02-05
Categories: cs.LG
arXiv: https://arxiv.org/abs/2602.05144v1

Machine learning systems are often trained and evaluated for fairness on historical data, yet deployed in environments where conditions have shifted. A particularly common form of shift occurs when the prevalence of positive outcomes changes differently across demographic groups--for example, when disease rates rise faster in one population than another, or when economic conditions affect loan default rates unequally. We study group-conditional prior probability shift (GPPS), where the label prevalence $P(Y=1\mid A=a)$ may change between training and deployment while the feature-generation process $P(X\mid Y,A)$ remains stable. Our analysis yields three main contributions. First, we prove a fundamental dichotomy: fairness criteria based on error rates (equalized odds) are structurally invariant under GPPS, while acceptance-rate criteria (demographic parity) can drift--and we prove this drift is unavoidable for non-trivial classifiers (shift-robust impossibility). Second, we show that target-domain risk and fairness metrics are identifiable without target labels: the invariance of ROC quantities under GPPS enables consistent estimation from source labels and unlabeled target data alone, with finite-sample guarantees. Third, we propose TAP-GPPS, a label-free post-processing algorithm that estimates prevalences from unlabeled data, corrects posteriors, and selects thresholds to satisfy demographic parity in the target domain. Experiments validate our theoretical predictions and demonstrate that TAP-GPPS achieves target fairness with minimal utility loss.

---

### LABBench2: An Improved Benchmark for AI Systems Performing Biology Research

Authors: Jon M Laurent, Albert Bou, Michael Pieler, Conor Igoe, Alex Andonian, Siddharth Narayanan, James Braza, Alexandros Sanchez Vassopoulos, Jacob L Steenwyk, Blake Lash, Andrew D White, Samuel G Rodriques
Published: 2026-02-04
Categories: cs.AI, cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2604.09554v2

Optimism for accelerating scientific discovery with AI continues to grow. Current applications of AI in scientific research range from training dedicated foundation models on scientific data to agentic autonomous hypothesis generation systems to AI-driven autonomous labs. The need to measure progress of AI systems in scientific domains correspondingly must not only accelerate, but increasingly shift focus to more real-world capabilities. Beyond rote knowledge and even just reasoning to actually measuring the ability to perform meaningful work. Prior work introduced the Language Agent Biology Benchmark LAB-Bench as an initial attempt at measuring these abilities. Here we introduce an evolution of that benchmark, LABBench2, for measuring real-world capabilities of AI systems performing useful scientific tasks. LABBench2 comprises nearly 1,900 tasks and is, for the most part, a continuation of LAB-Bench, measuring similar capabilities but in more realistic contexts. We evaluate performance of current frontier models, and show that while abilities measured by LAB-Bench and LABBench2 have improved substantially, LABBench2 provides a meaningful jump in difficulty (model-specific accuracy differences range from -26% to -46% across subtasks) and underscores continued room for performance improvement. LABBench2 continues the legacy of LAB-Bench as a de facto benchmark for AI scientific research capabilities and we hope that it continues to help advance development of AI tools for these core research functions. To facilitate community use and development, we provide the task dataset at https://huggingface.co/datasets/futurehouse/labbench2 and a public eval harness at https://github.com/EdisonScientific/labbench2.

---

### ARIA: Adaptive Retrieval Intelligence Assistant -- A Multimodal RAG Framework for Domain-Specific Engineering Education

Authors: Yue Luo, Dibakar Roy Sarkar, Rachel Herring Sangree, Somdatta Goswami
Published: 2026-02-04
Categories: cs.IR, cs.CL
arXiv: https://arxiv.org/abs/2604.06179v1

Developing effective, domain-specific educational support systems is central to advancing AI in education. Although large language models (LLMs) demonstrate remarkable capabilities, they face significant limitations in specialized educational applications, including hallucinations, limited knowledge updates, and lack of domain expertise. Fine-tuning requires complete model retraining, creating substantial computational overhead, while general-purpose LLMs often provide inaccurate responses in specialized contexts due to reliance on generalized training data. To address this, we propose ARIA (Adaptive Retrieval Intelligence Assistant), a Retrieval-Augmented Generation (RAG) framework for creating intelligent teaching assistants across university-level courses. ARIA leverages a multimodal content extraction pipeline combining Docling for structured document analysis, Nougat for mathematical formula recognition, and GPT-4 Vision API for diagram interpretation, with the e5-large-v2 embedding model for high semantic performance and low latency. This enables accurate processing of complex educational materials while maintaining pedagogical consistency through engineered prompts and response controls. We evaluate ARIA using lecture material from Statics and Mechanics of Materials, a sophomore-level civil engineering course at Johns Hopkins University, benchmarking against ChatGPT-5. Results demonstrate 97.5% accuracy in domain-specific question filtering and superior pedagogical performance. ARIA correctly answered all 20 relevant course questions while rejecting 58 of 60 non-relevant queries, achieving 90.9% precision, 100% recall, and 4.89/5.0 average response quality. These findings demonstrate that ARIA's course-agnostic architecture represents a scalable framework for domain-specific educational AI deployment.

---

### Scaling DPPs for RAG: Density Meets Diversity

Authors: Xun Sun, Baiheng Xie, Li Huang, Qiang Gao
Published: 2026-02-04
Categories: cs.LG, cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2604.03240v1

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by grounding generation in external knowledge, yielding relevance responses that are aligned with factual evidence and evolving corpora. Standard RAG pipelines construct context through relevance ranking, performing point-wise scoring between the user query and each corpora chunk. This formulation, however, ignores interactions among retrieved candidates, leading to redundant contexts that dilute density and fail to surface complementary evidence. We argue that effective retrieval should optimize jointly for both density and diversity, ensuring the grounding evidence that is dense in information yet diverse in coverage. In this study, we propose ScalDPP, a diversity-aware retrieval mechanism for RAG that incorporates Determinantal Point Processes (DPPs) through a lightweight P-Adapter, enabling scalable modeling of inter-chunk dependencies and complementary context selection. In addition, we develop a novel set-level objective, Diverse Margin Loss (DML), that enforces ground-truth complementary evidence chains to dominate any equally sized redundant alternatives under DPP geometry. Experimental results demonstrate the superiority of ScalDPP, substantiating our core statement in practice.

---

### Not All Denoising Steps Are Equal: Model Scheduling for Faster Masked Diffusion Language Models

Authors: Ivan Sedykh, Nikita Sorokin, Valentin Malykh
Published: 2026-02-04
Categories: cs.LG, cs.CL
arXiv: https://arxiv.org/abs/2604.02340v2

Recent advances in masked diffusion language models (MDLMs) narrow the quality gap to autoregressive LMs, but their sampling remains expensive because generation requires many full-sequence denoising passes with a large Transformer and, unlike autoregressive decoding, cannot benefit from KV caching. In this work, we exploit the flexibility of the diffusion framework and study model scheduling, where a smaller MDLM replaces the full model at a subset of denoising steps. Across models trained on OpenWebText and LM1B, we show that early and late denoising steps are substantially more robust to such replacement than middle steps, enabling up to a 17% reduction in FLOPs with only modest degradation in generative perplexity under both unconditional and prefix-conditional generation, while preserving sample diversity. We support these findings with a step-importance analysis based on loss and KL divergence between small and large models across timesteps, as well as an exhaustive search over coarse step segments, both of which identify the middle of the diffusion trajectory as most sensitive consistently across datasets. Our results suggest that simple, architecture-agnostic scheduling rules can significantly accelerate MDLM sampling while largely preserving generation quality.

---

### DF-ACBlurGAN: Structure-Aware Conditional Generation of Internally Repeated Patterns for Biomaterial Microtopography Design

Authors: Rongjun Dong, Xin Chen, Morgan R Alexander, Karthikeyan Sivakumar, Reza Omdivar, David A Winkler, Grazziela Figueredo
Published: 2026-02-04
Categories: cs.CV, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2603.28776v1

Learning to generate images with internally repeated and periodic structures poses a fundamental challenge for machine learning and computer vision models, which are typically optimised for local texture statistics and semantic realism rather than global structural consistency. This limitation is particularly pronounced in applications requiring strict control over repetition scale, spacing, and boundary coherence, such as microtopographical biomaterial surfaces. In this work, biomaterial design serves as a use case to study conditional generation of repeated patterns under weak supervision and class imbalance. We propose DF-ACBlurGAN, a structure-aware conditional generative adversarial network that explicitly reasons about long-range repetition during training. The approach integrates frequency-domain repetition scale estimation, scale-adaptive Gaussian blurring, and unit-cell reconstruction to balance sharp local features with stable global periodicity. Conditioning on experimentally derived biological response labels, the model synthesises designs aligned with target functional outcomes. Evaluation across multiple biomaterial datasets demonstrates improved repetition consistency and controllable structural variation compared to conventional generative approaches.

---

### VeriInteresting: An Empirical Study of Model Prompt Interactions in Verilog Code Generation

Authors: Luca Collini, Andrew Hennesee, Patrick Yubeaton, Siddharth Garg, Ramesh Karri
Published: 2026-02-04
Categories: cs.AR, cs.CL
arXiv: https://arxiv.org/abs/2603.08715v2

Rapid advances in language models (LMs) have created new opportunities for automated code generation while complicating trade-offs between model characteristics and prompt design choices. In this work, we provide an empirical map of recent trends in LMs for Verilog code generation, focusing on interactions among model reasoning, specialization, and prompt engineering strategies. We evaluate a diverse set of small and large LMs, including general-purpose, reasoning, and domain-specific variants. Our experiments use a controlled factorial design spanning benchmark prompts, structured outputs, prompt rewriting, chain-of-thought reasoning, in-context learning, and evolutionary prompt optimization via Genetic-Pareto. Across two Verilog benchmarks, we identify patterns in how model classes respond to structured prompts and optimization, and we document which trends generalize across LMs and benchmarks versus those that are specific to particular model-prompt combinations.

---

### Rethinking Personalization in Large Language Models at the Token Level

Authors: Chenheng Zhang, Yijun Lu, Lizhe Fang, Chunyuan Zheng, Jiajun Chai, Xiaohan Wang, Guojun Yin, Wei Lin, Yisen Wang, Zhouchen Lin
Published: 2026-02-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.06595v1

With large language models (LLMs) now performing strongly across diverse tasks, there is growing demand for them to personalize outputs for individual users. Personalization is typically framed as an additional layer on top of a base NLP task, requiring model responses to meet user-specific needs while still accomplishing the underlying task. From a token-level perspective, different tokens in a response contribute to personalization to varying degrees. Tokens with higher personalization relevance should therefore receive greater emphasis when developing personalized LLMs. However, accurately estimating such personalization degrees remains challenging. To address this challenge, we propose PerContrast, a self-contrast method that estimates each output token's dependence on user-specific information through causal intervention. Building on this mechanism, we develop the PerCE loss, which adaptively upweights tokens with higher estimated personalization degrees during training via a bootstrap procedure, enabling the model to alternate between estimating and optimizing these tokens. Experiments on multiple LLMs demonstrate that PerCE substantially improves personalization performance with minimal additional cost, achieving average gains of over 10% and up to 68.04% on the LongLaMP dataset, along with strong cross-task and cross-scenario transferability. These results highlight the importance of token-level personalization modeling and establish token-aware training as a simple yet effective paradigm for advancing personalized LLMs.

---

### A Coin Flip for Safety: LLM Judges Fail to Reliably Measure Adversarial Robustness

Authors: Leo Schwinn, Moritz Ladenburger, Tim Beyer, Mehrnaz Mofakhami, Gauthier Gidel, Stephan Günnemann
Published: 2026-02-04
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2603.06594v2

Automated \enquote{LLM-as-a-Judge} frameworks have become the de facto standard for scalable evaluation across natural language processing. For instance, in safety evaluation, these judges are relied upon to evaluate harmfulness in order to benchmark the robustness of safety against adversarial attacks. However, we show that existing validation protocols fail to account for substantial distribution shifts inherent to red-teaming: diverse victim models exhibit distinct generation styles, attacks distort output patterns, and semantic ambiguity varies significantly across jailbreak scenarios. Through a comprehensive audit using 6642 human-verified labels, we reveal that the unpredictable interaction of these shifts often causes judge performance to degrade to near random chance. This stands in stark contrast to the high human agreement reported in prior work. Crucially, we find that many attacks inflate their success rates by exploiting judge insufficiencies rather than eliciting genuinely harmful content. To enable more reliable evaluation, we propose ReliableBench, a benchmark of behaviors that remain more consistently judgeable, and JudgeStressTest, a dataset designed to expose judge failures. Data available at: https://github.com/SchwinnL/LLMJudgeReliability.

---

### Hierarchical Embedding Fusion for Retrieval-Augmented Code Generation

Authors: Nikita Sorokin, Ivan Sedykh, Valentin Malykh
Published: 2026-02-04
Categories: cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2603.06593v1

Retrieval-augmented code generation often conditions the decoder on large retrieved code snippets. This ties online inference cost to repository size and introduces noise from long contexts. We present Hierarchical Embedding Fusion (HEF), a two-stage approach to repository representation for code completion. First, an offline cache compresses repository chunks into a reusable hierarchy of dense vectors using a small fuser model. Second, an online interface maps a small number of retrieved vectors into learned pseudo-tokens that are consumed by the code generator. This replaces thousands of retrieved tokens with a fixed pseudo-token budget while preserving access to repository-level information. On RepoBench and RepoEval, HEF with a 1.8B-parameter pipeline achieves exact-match accuracy comparable to snippet-based retrieval baselines, while operating at sub-second median latency on a single A100 GPU. Compared to graph-based and iterative retrieval systems in our experimental setup, HEF reduces median end-to-end latency by 13 to 26 times. We also introduce a utility-weighted likelihood signal for filtering training contexts and report ablation studies on pseudo-token budget, embedding models, and robustness to harmful retrieval. Overall, these results indicate that hierarchical dense caching is an effective mechanism for low-latency, repository-aware code completion.

---

### Hierarchical Latent Structures in Data Generation Process Unify Mechanistic Phenomena across Scale

Authors: Jonas Rohweder, Subhabrata Dutta, Iryna Gurevych
Published: 2026-02-04
Categories: cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2603.06592v2

Contemporary studies in mechanistic interpretability have uncovered many puzzling phenomena in the neural information processing of Transformer-based language models, such as induction heads, function vectors, and the Hydra effect. Some of these individual phenomena have been independently tied to different data distributional properties, while some have been loosely associated with model architecture and how Transformers process information. However, a unified understanding of the relationship between data, model architecture, and optimization remains lacking, failing to answer the fundamental question: why do these three phenomena appear universally across different model families and scales, despite their seeming disconnect? In this work, we answer this question by unifying these three phenomena as consequences of hierarchical latent structures in the data generation process, coupled with decorrelated gradients across additive model components and directional concavity in the representation geometry. We validate our theoretical results in a toy model regime and in a large-scale synthetic data regime, comparing them with language models trained on natural language data.

---

### What Makes Position Zero Special? A Mechanistic Study of Position Zero Attention Sinks in LLMs

Authors: Runyu Peng, Ruixiao Li, Mingshu Chen, Yunhua Zhou, Qipeng Guo, Xipeng Qiu, Yucheng Lu, Chen Zhao
Published: 2026-02-04
Categories: cs.LG, cs.CL
arXiv: https://arxiv.org/abs/2603.06591v2

Transformers frequently allocate disproportionate attention to specific tokens, a phenomenon known as attention sinks. Causal large language models reliably form one at position zero, though its role remains debated. We approach this question from a mechanistic perspective, tracing how the position-zero sink arises from the model's internal computation. We identify a two-block subnetwork responsible for this behavior, which we term the P0-Sink Circuit, and show it arises purely from the structural properties of causal attention, requiring no semantic content. We further validate through from-scratch pre-training experiments that two proposed parameter-free methods effectively accelerate P0 sink formation, and find that earlier sink formation benefits pre-training and improves downstream performance. Both methods outperform the Transformer baseline and achieve performance comparable to Gated Attention across comprehensive settings. Code is available now at https://github.com/Pryest/flash-linear-attention.

---

### ARC-AGI-2 Technical Report

Authors: Wallyson Lemes de Oliveira, Mekhron Bobokhonov, Matteo Caorsi, Aldo Podestà, Gabriele Beltramo, Luca Crosato, Matteo Bonotto, Federica Cecchetto, Hadrien Espic, Dan Titus Salajan, Stefan Taga, Luca Pana et al.
Published: 2026-02-04
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2603.06590v1

The Abstraction and Reasoning Corpus (ARC) is designed to assess generalization beyond pattern matching, requiring models to infer symbolic rules from very few examples. In this work, we present a transformer-based system that advances ARC performance by combining neural inference with structure-aware priors and online task adaptation. Our approach is built on four key ideas. First, we reformulate ARC reasoning as a sequence modeling problem using a compact task encoding with only 125 tokens, enabling efficient long-context processing with a modified LongT5 architecture. Second, we introduce a principled augmentation framework based on group symmetries, grid traversals, and automata perturbations, enforcing invariance to representation changes. Third, we apply test-time training (TTT) with lightweight LoRA adaptation, allowing the model to specialize to each unseen task by learning its transformation logic from demonstrations. Fourth, we design a symmetry-aware decoding and scoring pipeline that aggregates likelihoods across augmented task views, effectively performing ``multi-perspective reasoning'' over candidate solutions. We demonstrate that these components work synergistically: augmentations expand hypothesis space, TTT sharpens local reasoning, and symmetry-based scoring improves solution consistency. Our final system achieves a significant improvement over transformer baselines and surpasses prior neural ARC solvers, closing the gap toward human-level generalization.

---

### Same Input, Different Scores: A Multi Model Study on the Inconsistency of LLM Judge

Authors: Fiona Lau
Published: 2026-02-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.04417v1

Large language models are increasingly used as automated evaluators in research and enterprise settings, a practice known as LLM-as-a-judge. While prior work has examined accuracy, bias, and alignment with human preferences, far less attention has been given to how consistently LLMs assign numerical scores, an important concern for many production workflows. This study systematically evaluates scoring stability across five commonly used models, GPT-4o, GPT-4o-mini, Gemini-2.5-Flash, Claude-Haiku-4.5, and Claude-Sonnet-4.5, two temperature settings, and real enterprise question-answer pairs drawn from a retrieval-augmented generation (RAG) system. We address three questions: how stable a model's scores are across repeated runs, how differently models score identical inputs, and how temperature affects scoring consistency. Temperature controls the determinism of an LLM's output. Despite expectations of stability at temperature=0, we observe substantial variability across models, with completeness scoring showing the largest fluctuations. Cross-model comparisons reveal systematic differences in strictness and interpretive style, leading to divergent ratings for the same answers. Lower temperatures improve stability for some models, notably GPT-4o and Gemini, but have limited or inconsistent effects for Anthropic models. These findings have important implications for enterprise pipelines that rely on LLM-generated scores for routing, triage, gating, or quality control. Identical inputs can receive different scores depending on model, family, or temperature, raising concerns around fairness, reproducibility, and operational reliability. Our results highlight the need for monitoring, robust parsing, and hybrid human-LLM evaluation strategies to ensure dependable use of LLM-as-a-judge in production environments.

---

### Optimizing What We Trust: Reliability-Guided QUBO Selection of Multi-Agent Weak Framing Signals for Arabic Sentiment Prediction

Authors: Rabab Alkhalifa
Published: 2026-02-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.04416v1

Framing detection in Arabic social media is difficult due to interpretive ambiguity, cultural grounding, and limited reliable supervision. Existing LLM-based weak supervision methods typically rely on label aggregation, which is brittle when annotations are few and socially dependent. We propose a reliability-aware weak supervision framework that shifts the focus from label fusion to data curation. A small multi-agent LLM pipeline, two framers, a critic, and a discriminator, treats disagreement and reasoning quality as epistemic signals and produces instance-level reliability estimates. These estimates guide a QUBO-based subset selection procedure that enforces frame balance while reducing redundancy. Intrinsic diagnostics and an out-of-domain Arabic sentiment transfer test show that the selected subsets are more reliable and encode non-random, transferable structure, without degrading strong text-only baselines.

---

### Dual Tuning for Reasoning Efficacy-Driven Data Curation in Multimodal LLM Training

Authors: Ruobing Zheng, Tianqi Li, Jianing Li, Qingpei Guo, Yi Yuan, Jingdong Chen
Published: 2026-02-04
Categories: cs.CL, cs.CV
arXiv: https://arxiv.org/abs/2603.04415v2

Reasoning post-training improves Large Language Models (LLMs) on complex tasks such as mathematics and coding, but its benefits across diverse multimodal tasks remains uncertain. The trend of releasing parallel "Instruct" and "Thinking" models by leading teams is both resource-intensive and user-unfriendly. Prior work finds that the gains from reasoning training are influenced by multiple factors, such as base model capabilities, task characteristics, and Chain-of-Thought (CoT) data quality. However, principled criteria for determining when reasoning post-training is beneficial and which data should support it are still lacking. In this paper, we propose Dual Tuning, a reasoning efficacy-driven data curation framework for multimodal LLMs training. Given a target task and a base model, Dual Tuning jointly evaluates whether the training data is beneficial and whether reasoning training with current CoT content yields positive gains over non-reasoning alternatives. We apply Dual Tuning across spatial, mathematical, and multi-disciplinary tasks, and further analyze how reinforcement learning and thinking patterns affect reasoning efficacy. The Dual Tuning results guide data curation by identifying data that benefit reasoning training, data better suited to direct-answer training, and data that are detrimental under both training modes. Our work provides quantitative criteria for selecting appropriate training data and matching post-training strategies.

---

### GRIP: Geometric Refinement and Adaptive Information Potential for Data Efficiency

Authors: Changhao Wang, Jiaolong Yang, Xinhao Yao, Yunfei Yu, Peng Jiao, Lu Yu, Junpeng Fang, Riccardo Cantoro, Qing Cui, Jun Zhou
Published: 2026-02-04
Categories: cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2603.00031v1

The performance of Large Language Models (LLMs) is increasingly governed by data efficiency rather than raw scaling volume. However, existing selection methods often decouple global distribution balancing from local instance selection, compromising the hierarchical integrity of the training set. We introduce \textbf{GRIP} (Geometric Refinement and Adaptive Information Potential), a framework that unifies these dimensions by modeling the corpus as an information-dense geometric space. GRIP employs a \textbf{Rapid Adaptation Probe (RAP)} to quantify the information potential of semantic clusters, dynamically re-allocating the sampling budget to regions with the highest representation deficits. Subsequently, we perform Intra-Cluster Selection using a \textbf{length-rectified geometric prior} to counteract embedding density artifacts and preserve long-tail logical sequences. Extensive evaluations on Mixture-of-Experts (MoE) models up to 300B tokens demonstrate that GRIP consistently outperforms state-of-the-art baselines, \textbf{surpassing the performance of models trained on $3\times$ larger uncurated datasets}. Our work establishes a robust geometric foundation for adaptive data curation in large-scale pre-training.

---

### SimpleTool: Parallel Decoding for Real-Time LLM Function Calling

Authors: Xiaoxin Shi, Jiaxin Wan, Linkang Dong, Wei Jiang, Yue Liu, Zengfeng Huang
Published: 2026-02-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.00030v1

LLM-based function calling enables intelligent agents to interact with external tools and environments, yet autoregressive decoding imposes a fundamental latency bottleneck that limits real-time applications such as embodied intelligence, game AI, and interactive avatars (e.g., 10 Hz control frequency). We observe that function calling differs fundamentally from free-form text generation: structured outputs exhibit substantial token redundancy (delimiters, parameter names), and arguments exhibit weak causal dependencies. Crucially, these two properties must be exploited jointly to achieve real-time performance. We present SimpleTool, which introduces special tokens that serve a dual role: compressing low-entropy tokens (4-6x reduction) while acting as mode selectors that enable independent parallel generation of function name and arguments. This synergistic design achieves 3-6x end-to-end speedup (up to 9.6x) with only +8.2% parallelization overhead. Experiments on five benchmarks across Qwen-series models (0.5B-14B) demonstrate substantial speedup while maintaining competitive or improved accuracy. On Mobile Actions, ST-Qwen-0.5B outperforms Google's FunctionGemma in both accuracy and latency consistency. With quantization on consumer-grade GPU, SimpleTool achieves 61.2ms P50 latency, enabling 16 Hz real-time control at 4B model scale, bridging the gap between LLM function calling and latency-critical real-world deployment.

---

### Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models

Authors: Youngji Roh, Hyunjin Cho, Jaehyung Kim
Published: 2026-02-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.00029v3

Large Language Models (LLMs) exhibit highly anisotropic internal representations, often characterized by massive activations, a phenomenon where a small subset of feature dimensions possesses magnitudes significantly larger than the rest. While prior works view these extreme dimensions primarily as artifacts to be managed, we propose a distinct perspective: these dimensions serve as intrinsic interpretable functional units arising from domain specialization. Specifically, we propose a simple magnitude-based criterion to identify Domain-Critical Dimensions in a training-free manner. Our analyses reveal that such dimensions behave as interpretable semantic detectors for symbolic/quantitative patterns or domain-specific terms. In addition, we introduce Critical Dimension Steering, which applies activation steering exclusively to the identified dimensions. Empirical results show that this approach outperforms conventional whole-dimension steering in domain adaptation and jailbreaking scenarios.

---

### EPPCMinerBen: A Novel Benchmark for Evaluating Large Language Models on Electronic Patient-Provider Communication via the Patient Portal

Authors: Samah Fodeh, Yan Wang, Linhai Ma, Srivani Talakokkul, Jordan M. Alpert, Sarah Schellhorn
Published: 2026-02-04
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.00028v1

Effective communication in health care is critical for treatment outcomes and adherence. With patient-provider exchanges shifting to secure messaging, analyzing electronic patient-communication (EPPC) data is both essential and challenging. We introduce EPPCMinerBen, a benchmark for evaluating LLMs in detecting communication patterns and extracting insights from electronic patient-provider messages. EPPCMinerBen includes three sub-tasks: Code Classification, Subcode Classification, and Evidence Extraction. Using 1,933 expert annotated sentences from 752 secure messages of the patient portal at Yale New Haven Hospital, it evaluates LLMs on identifying communicative intent and supportive text. Benchmarks span various LLMs under zero-shot and few-shot settings, with data to be released via the NCI Cancer Data Service. Model performance varied across tasks and settings. Llama-3.1-70B led in evidence extraction (F1: 82.84%) and performed well in classification. Llama-3.3-70b-Instruct outperformed all models in code classification (F1: 67.03%). DeepSeek-R1-Distill-Qwen-32B excelled in subcode classification (F1: 48.25%), while sdoh-llama-3-70B showed consistent performance. Smaller models underperformed, especially in subcode classification (>30% F1). Few-shot prompting improved most tasks. Our results show that large, instruction-tuned models generally perform better in EPPCMinerBen tasks, particularly evidence extraction while smaller models struggle with fine-grained reasoning. EPPCMinerBen provides a benchmark for discourse-level understanding, supporting future work on model generalization and patient-provider communication analysis. Keywords: Electronic Patient-Provider Communication, Large language models, Data collection, Prompt engineering

---

### Bilevel Optimization with Lower-Level Uniform Convexity: Theory and Algorithm

Authors: Yuman Wu, Xiaochuan Gong, Jie Hao, Mingrui Liu
Published: 2026-02-04
Categories: math.OC, cs.LG
arXiv: https://arxiv.org/abs/2603.00027v1

Bilevel optimization is a hierarchical framework where an upper-level optimization problem is constrained by a lower-level problem, commonly used in machine learning applications such as hyperparameter optimization. Existing bilevel optimization methods typically assume strong convexity or Polyak-Łojasiewicz (PL) conditions for the lower-level function to establish non-asymptotic convergence to a solution with small hypergradient. However, these assumptions may not hold in practice, and recent work~\citep{chen2024finding} has shown that bilevel optimization is inherently intractable for general convex lower-level functions with the goal of finding small hypergradients. In this paper, we identify a tractable class of bilevel optimization problems that interpolates between lower-level strong convexity and general convexity via \emph{lower-level uniform convexity}. For uniformly convex lower-level functions with exponent $p\geq 2$, we establish a novel implicit differentiation theorem characterizing the hyperobjective's smoothness property. Building on this, we design a new stochastic algorithm, termed UniBiO, with provable convergence guarantees, based on an oracle that provides stochastic gradient and Hessian-vector product information for the bilevel problems. Our algorithm achieves $\widetilde{O}(ε^{-5p+6})$ oracle complexity bound for finding $ε$-stationary points. Notably, our complexity bounds match the optimal rates in terms of the $ε$ dependency for strongly convex lower-level functions ($p=2$), up to logarithmic factors. Our theoretical findings are validated through experiments on synthetic tasks and data hyper-cleaning, demonstrating the effectiveness of our proposed algorithm.

---

### ActMem: Bridging the Gap Between Memory Retrieval and Reasoning in LLM Agents

Authors: Xiaohui Zhang, Zequn Sun, Chengyuan Yang, Yaqin Jin, Yazhong Zhang, Wei Hu
Published: 2026-02-04
Categories: cs.CL, cs.AI, cs.IR
arXiv: https://arxiv.org/abs/2603.00026v2

Memory management is essential for LLM agents in long-term interactions. Current memory frameworks typically treat agents as passive ``recorders'' and retrieve information without understanding its deeper implications. They may fail in scenarios requiring reasoning and complex decision-making. To bridge this critical gap, we propose a novel actionable memory framework called ActMem that integrates memory retrieval with active causal reasoning. ActMem transforms unstructured dialogue history into a structured causal and semantic graph. By leveraging counterfactual reasoning and commonsense completion, it enables agents to deduce implicit constraints and resolve potential conflicts between past states and current intentions. Furthermore, we introduce a comprehensive dataset ActMemEval to evaluate agent reasoning capabilities in logic-driven scenarios, moving beyond the fact-retrieval focus of existing memory benchmarks. Experiments demonstrate that ActMem significantly outperforms baselines in handling complex, memory-dependent tasks, paving the way for more consistent and reliable intelligent assistants.

---