# Research Papers — 2026-03 · bundle page 71 of 243

> Full-text markdown of papers #3501–3550 of 12123 published in 2026-03, newest first. AI-readable mirrors of the full 12123-paper 2026-03 archive (50 papers per file).

Canonical page: https://aipod.dev/papers/archive/2026-03/page/71
[← Previous bundle](https://aipod.dev/papers/archive/2026-03/page/70.md)
[Next bundle →](https://aipod.dev/papers/archive/2026-03/page/72.md)

### Knowledge Packs: Zero-Token Knowledge Delivery via KV Cache Injection

Authors: Andrey Pustovit
Published: 2026-03-22
Categories: cs.CL
arXiv: https://arxiv.org/abs/2604.03270v1

RAG wastes tokens. We propose Knowledge Packs: pre-computed KV caches that deliver the same knowledge at zero token cost. For causal transformers, the KV cache from a forward pass on text F is identical to what a joint pass on F+q would produce - this follows directly from the causal mask. The equivalence is exact but fragile: wrong chat template formatting causes 6-7pp degradation, which we believe explains prior claims of KV outperforming RAG. With correct formatting: zero divergences across 700 questions on Qwen3-8B and Llama-3.1-8B, up to 95% token savings. The KV interface also enables behavioral steering that RAG cannot do. Because RoPE rotates keys but leaves values untouched, contrastive deltas on cached values can nudge model behavior while key arithmetic destroys coherence. The effect sits in mid-layer values (33-66%), independent directions are nearly orthogonal (cos~0) and compose, and both channels - knowledge and steering - run simultaneously at alpha<=0.7 without interference. No training, no weight modification.

---

### CIPHER: Conformer-based Inference of Phonemes from High-density EEG

Authors: Varshith Madishetty
Published: 2026-03-22
Categories: cs.CL, cs.AI, cs.SD
arXiv: https://arxiv.org/abs/2604.02362v1

Decoding speech information from scalp EEG remains difficult due to low SNR and spatial blurring. We present CIPHER (Conformer-based Inference of Phonemes from High-density EEG Representations), a dual-pathway model using (i) ERP features and (ii) broadband DDA coefficients. On OpenNeuro ds006104 (24 participants, two studies with concurrent TMS), binary articulatory tasks reach near-ceiling performance but are highly confound-vulnerable (acoustic onset separability and TMS-target blocking). On the primary 11-class CVC phoneme task under full Study 2 LOSO (16 held-out subjects), performance is substantially lower (real-word WER: ERP 0.671 +/- 0.080, DDA 0.688 +/- 0.096, indicating limited fine-grained discriminability. We therefore position this work as a benchmark and feature-comparison study rather than an EEG-to-text system, and we constrain neural-representation claims to confound-controlled evidence.

---

### When and Where: A Model Hippocampal Network Unifies Formation of Time Cells and Place Cells

Authors: Qiaorong S. Yu, Zhaoze Wang, Vijay Balasubramanian
Published: 2026-03-22
Categories: q-bio.NC, cs.AI, cs.LG, cs.NE, physics.bio-ph
arXiv: https://arxiv.org/abs/2604.00036v1

Hippocampal place and time cells encode spatial and temporal aspects of experience. Both have the same neural substrate, but have been modeled as having different functions and mechanistic origins, place cells as continuous attractors, and time cells as leaky integrators. Here, we show that both types emerge from two dynamical regimes of a single recurrent network (RNN) modeling hippocampal CA3 as a predictive autoencoder. The network receives simulated, partially occluded ``experience vectors" containing spatial patterns (location-specific activity sampled during environmental traversal) and/or temporal patterns (correlated activity pairs separated by ``void" intervals), and is trained to reconstruct missing input. During spatial navigation, the network generates stable attractor-like place fields. But trained on temporally structured inputs, the network produces sequentially broadened fields, recapitulating time cells. By varying spatio-temporal input patterning, we observe hidden units transition smoothly between time cell-like and place cell-like representations. These results suggest a shared origin, but task-driven difference, between place and time cells.

---

### Quantum Fuzzy Sets Revisited: Density Matrices, Decoherence, and the Q-Matrix Framework

Authors: Mirco A. Mannucci
Published: 2026-03-22
Categories: quant-ph, cs.AI
arXiv: https://arxiv.org/abs/2603.26739v1

In 2006 we proposed Quantum Fuzzy Sets, observing that states of a quantum register could serve as characteristic functions of fuzzy subsets, embedding Zadeh's unit interval into the Bloch sphere. That paper was deliberately preliminary. In the two decades since, the idea has been taken up by researchers working on quantum annealers, intuitionistic fuzzy connectives, and quantum machine learning, while parallel developments in categorical quantum mechanics have reshaped the theoretical landscape. The present paper revisits that programme and introduces two main extensions. First, we move from pure states to density matrices, so that truth values occupy the entire Bloch ball rather than its surface; this captures the phenomenon of semantic decoherence that pure-state semantics cannot express. Second, we introduce the Q-Matrix, a global density matrix from which individual quantum fuzzy sets emerge as local sections via partial trace. We define a category QFS of quantum fuzzy sets, establish basic structural properties (monoidal structure, fibration over Set), characterize the classical limit as simultaneous diagonalizability, and exhibit an obstruction to a fully internal Frobenius-algebra treatment.

---

### SleepVLM: A Rule-Grounded Vision-Language Model for Auditable Sleep Staging

Authors: Guifeng Deng, Pan Wang, Mengfan Niu, Jiquan Wang, Shuying Rao, Junyi Xie, Xi'ang Chen, Sha Zhao, Gang Pan, Wanjun Guo, Tao Li, Haiteng Jiang
Published: 2026-03-22
Categories: cs.CV, cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2603.26738v4

Sleep staging is essential for sleep assessment and disorder diagnosis. In recent years, automatic sleep staging systems have achieved accuracy approaching that of human experts, but the black-box nature of their predictions hinders clinical adoption. Existing interpretability methods offer partial insight into model behavior, but their outputs still require expert reinterpretation and do not provide a direct basis for auditing individual predictions. To improve trustworthiness, we propose the task of auditable sleep staging. To solve this task, we present SleepVLM, a vision-language model that casts sleep staging as visual reasoning over rendered polysomnography (PSG) waveform images. For each epoch, SleepVLM outputs a stage together with the applicable American Academy of Sleep Medicine (AASM) rules and an auditable rationale. The model is trained using a two-stage framework: Waveform-Perceptual Pre-training followed by Rule-Grounded Supervised Fine-tuning over a mixture of fine-grained and coarse annotations. Experiments on four datasets show that SleepVLM outperforms state-of-the-art methods on average. An automated AASM-feature audit shows broad coverage of stage-defining evidence in the rationales, and independent experts validate their reasoning quality. To facilitate further research, we construct and release MASS-EX, an expert-annotated dataset for rule-grounded sleep staging with AASM rule annotations and expert-written rationales.

---

### Relational graph-driven differential denoising and diffusion attention fusion for multimodal conversation emotion recognition

Authors: Ying Liu, Yuntao Shou, Wei Ai, Tao Meng, Keqin Li
Published: 2026-03-22
Categories: cs.CL, cs.SD, eess.AS
arXiv: https://arxiv.org/abs/2603.25752v1

In real-world scenarios, audio and video signals are often subject to environmental noise and limited acquisition conditions, resulting in extracted features containing excessive noise. Furthermore, there is an imbalance in data quality and information carrying capacity between different modalities. These two issues together lead to information distortion and weight bias during the fusion phase, impairing overall recognition performance. Most existing methods neglect the impact of noisy modalities and rely on implicit weighting to model modality importance, thereby failing to explicitly account for the predominant contribution of the textual modality in emotion understanding. To address these issues, we propose a relation-aware denoising and diffusion attention fusion model for MCER. Specifically, we first design a differential Transformer that explicitly computes the differences between two attention maps, thereby enhancing temporally consistent information while suppressing time-irrelevant noise, which leads to effective denoising in both audio and video modalities. Second, we construct modality-specific and cross-modality relation subgraphs to capture speaker-dependent emotional dependencies, enabling fine-grained modeling of intra- and inter-modal relationships. Finally, we introduce a text-guided cross-modal diffusion mechanism that leverages self-attention to model intra-modal dependencies and adaptively diffuses audiovisual information into the textual stream, ensuring more robust and semantically aligned multimodal fusion.

---

### STEM Agent: A Self-Adapting, Tool-Enabled, Extensible Architecture for Multi-Protocol AI Agent Systems

Authors: Alfred Shen, Aaron Shen
Published: 2026-03-22
Categories: cs.AI
arXiv: https://arxiv.org/abs/2603.22359v1

Current AI agent frameworks commit early to a single interaction protocol, a fixed tool integration strategy, and static user models, limiting their deployment across diverse interaction paradigms. To address these constraints, we introduce STEM Agent (Self-adapting, Tool-enabled, Extensible, Multi-agent), a modular architecture inspired by biological pluripotency in which an undifferentiated agent core differentiates into specialized protocol handlers, tool bindings, and memory subsystems that compose into a fully functioning AI system. The framework unifies five interoperability protocols (A2A, AG-UI, A2UI, UCP, and AP2) behind a single gateway, introduces a Caller Profiler that continuously learns user preferences across more than twenty behavioral dimensions, externalizes all domain capabilities through the Model Context Protocol (MCP), and implements a biologically inspired skills acquisition system in which recurring interaction patterns crystallize into reusable agent skills through a maturation lifecycle analogous to cell differentiation. Complementing these capabilities, the memory system incorporates consolidation mechanisms, including episodic pruning, semantic deduplication, and pattern extraction, designed for sub-linear growth under sustained interaction. A comprehensive 413-test suite validates protocol handler behavior and component integration across all five architectural layers, completing in under three seconds.

---

### Bridging neuroscience and AI: adaptive, culturally sensitive technologies transforming aphasia rehabilitation

Authors: Andreea I. Niculescu, Jochen Ehnes, Minghui Dong
Published: 2026-03-22
Categories: q-bio.NC, cs.AI
arXiv: https://arxiv.org/abs/2603.22357v1

Aphasia, a language impairment primarily resulting from stroke or brain injury, profoundly disrupts communication and everyday functioning. Despite advances in speech therapy, barriers such as limited therapist availability and the scarcity of personalized, culturally relevant tools continue to hinder optimal rehabilitation outcomes. This paper reviews recent developments in neurocognitive research and language technologies that contribute to the diagnosis and therapy of aphasia. Drawing on findings from our ethnographic field study, we introduce two digital therapy prototypes designed to reflect local linguistic diversity and enhance patient engagement. We also show how insights from neuroscience and the local context guided the design of these tools to better meet patient and therapist needs. Our work highlights the potential of adaptive, AI-enhanced assistive technologies to complement conventional therapy and broaden access to therapy. We conclude by outlining future research directions for advancing personalized and scalable aphasia rehabilitation.

---

### Demystifying Low-Rank Knowledge Distillation in Large Language Models: Convergence, Generalization, and Information-Theoretic Guarantees

Authors: Alberlucia Rafael Soarez, Daniel Kim, Mariana Costa, Alejandro Torre
Published: 2026-03-22
Categories: stat.ML, cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2603.22355v1

Knowledge distillation has emerged as a powerful technique for compressing large language models (LLMs) into efficient, deployable architectures while preserving their advanced capabilities. Recent advances in low-rank knowledge distillation, particularly methods like Low-Rank Clone (LRC), have demonstrated remarkable empirical success, achieving comparable performance to full-parameter distillation with significantly reduced training data and computational overhead. However, the theoretical foundations underlying these methods remain poorly understood. In this paper, we establish a rigorous theoretical framework for low-rank knowledge distillation in language models. We prove that under mild assumptions, low-rank projection preserves the optimization dynamics, yielding explicit convergence rates of $O(1/\sqrt{T})$. We derive generalization bounds that characterize the fundamental trade-off between model compression and generalization capability, showing that the generalization error scales with the rank parameter as $O(r(m+n)/\sqrt{n})$. Furthermore, we provide an information-theoretic analysis of the activation cloning mechanism, revealing its role in maximizing the mutual information between the teacher's and student's intermediate representations. Our theoretical results offer principled guidelines for rank selection, mathematically suggesting an optimal rank $r^* = O(\sqrt{n})$ where $n$ is the sample size. Experimental validation on standard language modeling benchmarks confirms our theoretical predictions, demonstrating that the empirical convergence, rank scaling, and generalization behaviors align closely with our bounds.

---

### WIST: Web-Grounded Iterative Self-Play Tree for Domain-Targeted Reasoning Improvement

Authors: Fangyuan Li, Pengfei Li, Shijie Wang, Junqi Gao, Jianxing Liu, Biqing Qi, Yuqiang Li
Published: 2026-03-22
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.22352v1

Recent progress in reinforcement learning with verifiable rewards (RLVR) offers a practical path to self-improvement of language models, but existing methods face a key trade-off: endogenous self-play can drift over iterations, while corpus-grounded approaches rely on curated data environments. We present \textbf{WIST}, a \textbf{W}eb-grounded \textbf{I}terative \textbf{S}elf-play \textbf{T}ree framework for domain-targeted reasoning improvement that learns directly from the open web without requiring any pre-arranged domain corpus. WIST incrementally expands a domain tree for exploration, and retrieves and cleans path-consistent web corpus to construct a controllable training environment. It then performs Challenger--Solver self-play with verifiable rewards, and feeds learnability signals back to update node posteriors and guide subsequent exploration through an adaptive curriculum. Across four backbones, WIST consistently improves over the base models and typically outperforms both purely endogenous self-evolution and corpus-grounded self-play baselines, with the Overall gains reaching \textbf{+9.8} (\textit{Qwen3-4B-Base}) and \textbf{+9.7} (\textit{OctoThinker-8B}). WIST is also domain-steerable, improving \textit{Qwen3-8B-Base} by \textbf{+14.79} in medicine and \textit{Qwen3-4B-Base} by \textbf{+5.28} on PhyBench. Ablations further confirm the importance of WIST's key components for stable open-web learning. Our Code is available at https://github.com/lfy-123/WIST.

---

### Session Risk Memory (SRM): Temporal Authorization for Deterministic Pre-Execution Safety Gates

Authors: Florin Adrian Chitan
Published: 2026-03-22
Categories: cs.AI, cs.CR
arXiv: https://arxiv.org/abs/2603.22350v1

Deterministic pre-execution safety gates evaluate whether individual agent actions are compatible with their assigned roles. While effective at per-action authorization, these systems are structurally blind to distributed attacks that decompose harmful intent across multiple individually-compliant steps. This paper introduces Session Risk Memory (SRM), a lightweight deterministic module that extends stateless execution gates with trajectory-level authorization. SRM maintains a compact semantic centroid representing the evolving behavioral profile of an agent session and accumulates a risk signal through exponential moving average over baseline-subtracted gate outputs. It operates on the same semantic vector representation as the underlying gate, requiring no additional model components, training, or probabilistic inference. We evaluate SRM on a multi-turn benchmark of 80 sessions containing slow-burn exfiltration, gradual privilege escalation, and compliance drift scenarios. Results show that ILION+SRM achieves F1 = 1.0000 with 0% false positive rate, compared to stateless ILION at F1 = 0.9756 with 5% FPR, while maintaining 100% detection rate for both systems. Critically, SRM eliminates all false positives with a per-turn overhead under 250 microseconds. The framework introduces a conceptual distinction between spatial authorization consistency (evaluated per action) and temporal authorization consistency (evaluated over trajectory), providing a principled basis for session-level safety in agentic systems.

---

### Learning Safely Without Knowing the World:COMPASS-Hedge

Authors: Ting Hu, Luanda Cai, Emmanouil-Vasileios Vlatakis-Gkaragkounis
Published: 2026-03-22
Categories: cs.LG, cs.GT
arXiv: https://arxiv.org/abs/2603.22348v4

Online learning algorithms often face a fundamental trilemma: balancing regret guarantees between adversarial and stochastic settings and providing baseline safety against a fixed comparator. While existing methods excel in one or two of these regimes, they typically fail to unify all three without sacrificing optimal rates or requiring oracle access to problem-dependent parameters. In this work, we bridge this gap by introducing COMPASS-Hedge. To the best of our knowledge, our algorithm is the first full-information anytime method to simultaneously achieve, up to logarithmic factors: i) minimax-optimal regret in adversarial environments; ii) instance-optimal, gap-dependent regret in stochastic environments; and iii) $\tilde{\mathcal{O}}(1)$ regret relative to a designated baseline policy. Crucially, COMPASS-Hedge is parameter-free and requires no prior knowledge of the environment's nature or the magnitude of the stochastic suboptimality gaps. Our approach hinges on a novel integration of adaptive pseudo-regret scaling and phase-based aggression, coupled with a comparator-aware mixing strategy. To the best of our knowledge, this provides the first "best-of-three-world" guarantee in the full-information setting, establishing that baseline safety does not have to come at the cost of worst-case robustness or stochastic efficiency.

---

### Intelligence Inertia: Physical Isomorphism and Applications

Authors: Jipeng Han
Published: 2026-03-22
Categories: cs.AI, cond-mat.stat-mech, cs.LG
arXiv: https://arxiv.org/abs/2603.22347v2

Classical frameworks like Fisher Information approximate the cost of neural adaptation only in low-density regimes, failing to explain the explosive computational overhead incurred during deep structural reconfiguration. To address this, we introduce \textbf{Intelligence Inertia}, a property derived from the fundamental non-commutativity between rules and states ($(\hat{S}, \hat{R}) = i\mathcal{D}$). Rather than claiming a new fundamental physical law, we establish a \textbf{heuristic mathematical isomorphism} between deep learning dynamics and Minkowski spacetime. Acting as an \textit{effective theory} for high-dimensional tensor evolution, we derive a non-linear cost formula mirroring the Lorentz factor, predicting a relativistic $J$-shaped inflation curve -- a computational wall where classical approximations fail. We validate this framework via three experiments: (1) adjudicating the $J$-curve divergence under high-entropy noise, (2) mapping the optimal geodesic for architecture evolution, and (3) deploying an \textbf{inertia-aware scheduler wrapper} that prevents catastrophic forgetting. Adopting this isomorphism yields an exact quantitative metric for structural resistance, advancing the stability and efficiency of intelligent agents.

---

### First-Mover Bias in Gradient Boosting Explanations: Mechanism, Detection, and Resolution

Authors: Drake Caraker, Bryan Arnold, David Rhoads
Published: 2026-03-22
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.22346v2

We identify first-mover bias -- path-dependent concentration of SHAP feature importance from sequential residual fitting in gradient boosting -- as a mechanistic contributor to attribution instability under multicollinearity. Scaling up a single model amplifies this effect: a Large Single Model matching our method's total tree count produces the poorest attribution reproducibility of any approach tested. We show that model independence largely neutralizes first-mover bias. Both DASH (Diversified Aggregation of SHAP) and simple seed-averaging (Stochastic Retrain) restore stability by breaking the sequential dependency chain. At rho=0.9, both achieve stability ~0.977, while Single Best degrades to 0.958 and LSM to 0.938. On Breast Cancer, DASH improves stability from 0.376 to 0.925 (+0.549), outperforming Stochastic Retrain by +0.063. Under nonlinear DGPs, the advantage emerges at rho>=0.7. DASH provides two diagnostic tools -- the Feature Stability Index and Importance-Stability Plot -- that detect first-mover bias without ground truth. A crossed ANOVA with formal F-statistics confirms the mechanism: DASH shifts variance from model-dominated (40.6%) to data-dominated (73.6%). Software at https://github.com/DrakeCaraker/dash-shap

---

### Dynamic Fusion-Aware Graph Convolutional Neural Network for Multimodal Emotion Recognition in Conversations

Authors: Tao Meng, Weilun Tang, Yuntao Shou, Yilong Tan, Jun Zhou, Wei Ai, Keqin Li
Published: 2026-03-22
Categories: cs.AI
arXiv: https://arxiv.org/abs/2603.22345v1

Multimodal emotion recognition in conversations (MERC) aims to identify and understand the emotions expressed by speakers during utterance interaction from multiple modalities (e.g., text, audio, images, etc.). Existing studies have shown that GCN can improve the performance of MERC by modeling dependencies between speakers. However, existing methods usually use fixed parameters to process multimodal features for different emotion types, ignoring the dynamics of fusion between different modalities, which forces the model to balance performance between multiple emotion categories, thus limiting the model's performance on some specific emotions. To this end, we propose a dynamic fusion-aware graph convolutional neural network (DF-GCN) for robust recognition of multimodal emotion features in conversations. Specifically, DF-GCN integrates ordinary differential equations into graph convolutional networks (GCNs) to {capture} the dynamic nature of emotional dependencies within utterance interaction networks and leverages the prompts generated by the global information vector (GIV) of the utterance to guide the dynamic fusion of multimodal features. This allows our model to dynamically change parameters when processing each utterance feature, so that different network parameters can be equipped for different emotion categories in the inference stage, thereby achieving more flexible emotion classification and enhancing the generalization ability of the model. Comprehensive experiments conducted on two public multimodal conversational datasets {confirm} that the proposed DF-GCN model delivers superior performance, benefiting significantly from the dynamic fusion mechanism introduced.

---

### Safety as Computation: Certified Answer Reuse via Capability Closure in Task-Oriented Dialogue

Authors: Cosimo Spera
Published: 2026-03-22
Categories: cs.AI
arXiv: https://arxiv.org/abs/2603.21448v1

We introduce a new paradigm for task-oriented dialogue systems: safety certification as a computational primitive for answer reuse. Current systems treat each turn independently, recomputing answers via retrieval or generation even when they are already derivable from prior state. We show that in capability-based systems, the safety certification step computes a fixed-point closure cl(At) that already contains every answer reachable from the current configuration. We operationalize this insight with a Certified Answer Store (CAS) augmented by Pre-Answer Blocks (PAB): at each certified turn, the system materializes all derivable follow-up answers together with minimal provenance witnesses. Subsequent queries are answered in sub-millisecond time via formal containment checks, eliminating redundant retrieval and generation.

---

### KG-Hopper: Empowering Compact Open LLMs with Knowledge Graph Reasoning via Reinforcement Learning

Authors: Shuai Wang, Yinan Yu
Published: 2026-03-22
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2603.21440v4

Large Language Models (LLMs) demonstrate impressive natural language capabilities but often struggle with knowledge-intensive reasoning tasks. Knowledge Base Question Answering (KBQA), which leverages structured Knowledge Graphs (KGs) exemplifies this challenge due to the need for accurate multi-hop reasoning. Existing approaches typically perform sequential reasoning steps guided by predefined pipelines, restricting flexibility and causing error cascades due to isolated reasoning at each step. To address these limitations, we propose KG-Hopper, a novel Reinforcement Learning (RL) framework that empowers compact open LLMs with the ability to perform integrated multi-hop KG reasoning within a single inference round. Rather than reasoning step-by-step, we train a Reasoning LLM that embeds the entire KG traversal and decision process into a unified ``thinking'' stage, enabling global reasoning over cross-step dependencies and dynamic path exploration with backtracking. Experimental results on eight KG reasoning benchmarks show that KG-Hopper, based on a 7B-parameter LLM, consistently outperforms larger multi-step systems (up to 70B) and achieves competitive performance with proprietary models such as GPT-3.5-Turbo and GPT-4o-mini, while remaining compact, open, and data-efficient. The code is publicly available at: https://github.com/Wangshuaiia/KG-Hopper.

---

### LLM-Powered Workflow Optimization for Multidisciplinary Software Development: An Automotive Industry Case Study

Authors: Shuai Wang, Yinan Yu, Earl Barr, Dhasarathy Parthasarathy
Published: 2026-03-22
Categories: cs.SE, cs.AI
arXiv: https://arxiv.org/abs/2603.21439v4

Multidisciplinary Software Development (MSD) requires domain experts and developers to collaborate across incompatible formalisms and separate artifact sets. Today, even with AI coding assistants like GitHub Copilot, this process remains inefficient; individual coding tasks are semi-automated, but the workflow connecting domain knowledge to implementation is not. Developers and experts still lack a shared view, resulting in repeated coordination, clarification rounds, and error-prone handoffs. We address this gap through a graph-based workflow optimization approach that progressively replaces manual coordination with LLM-powered services, enabling incremental adoption without disrupting established practices. We evaluate our approach on \texttt{spapi}, a production in-vehicle API system at Volvo Group involving 192 endpoints, 420 properties, and 776 CAN signals across six functional domains. The automated workflow achieves 93.7\% F1 score while reducing per-API development time from approximately 5 hours to under 7 minutes, saving an estimated 979 engineering hours. In production, the system received high satisfaction from both domain experts and developers, with all participants reporting full satisfaction with communication efficiency.

---

### PROMPT2BOX:Improving LLM Weakness Discovery and Specificity Estimation by Uncovering Entailment Structure among Prompts

Authors: Neeladri Bhuiya, Shib Sankar Dasgupta, Andrew McCallum, Haw-Shiuan Chang
Published: 2026-03-22
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.21438v3

To discover the weaknesses of LLMs, researchers often embed prompts into a vector space and cluster them to extract insightful patterns. However, vector embeddings primarily capture topical similarity; as a result, prompts that share a topic but differ in specificity, and consequently in difficulty, are often represented similarly, making fine-grained weakness analysis difficult. To address this limitation, we propose Prompt2Box, which embeds prompts into a box embedding space using a trained encoder. The encoder, trained on existing and synthesized datasets, outputs box embeddings that capture not only semantic similarity but also specificity relations between prompts (e.g., "writing an adventure story" is more specific than "writing a story"). We further develop a novel dimension reduction technique for box embeddings to facilitate dataset visualization and comparison. Our experiments demonstrate that box embeddings consistently capture prompt specificity better than vector baselines and achieve 45% error reduction on average in predicting specificity compared to the prompt length baseline. On the downstream task of creating hierarchical clustering trees for 17 LLMs from the UltraFeedback dataset, Prompt2Box can identify 13.5% more LLM weaknesses than vector baselines and achieves an approximately 33% stronger correlation between hierarchical depth and instruction specificity. The code is available at https://github.com/zawedcvg/box_embeddings.

---

### PROMPT2BOX: Uncovering Entailment Structure among LLM Prompts

Authors: Neeladri Bhuiya, Shib Sankar Dasgupta, Andrew McCallum, Haw-Shiuan Chang
Published: 2026-03-22
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.21438v2

To discover the weaknesses of LLMs, researchers often embed prompts into a vector space and cluster them to extract insightful patterns. However, vector embeddings primarily capture topical similarity. As a result, prompts that share a topic but differ in specificity, and consequently in difficulty, are often represented similarly, making fine-grained weakness analysis difficult. To address this limitation, we propose PROMPT2BOX, which embeds prompts into a box embedding space using a trained encoder. The encoder, trained on existing and synthesized datasets, outputs box embeddings that capture not only semantic similarity but also specificity relations between prompts (e.g., "writing an adventure story" is more specific than "writing a story"). We further develop a novel dimension reduction technique for box embeddings to facilitate dataset visualization and comparison. Our experiments demonstrate that box embeddings consistently capture prompt specificity better than vector baselines. On the downstream task of creating hierarchical clustering trees for 17 LLMs from the UltraFeedback dataset, PROMPT2BOX can identify 8.9\% more LLM weaknesses than vector baselines and achieves an approximately 33\% stronger correlation between hierarchical depth and instruction specificity.

---

### Pooling and Semantic Shift: The Fundamental Challenges in Long Text Embedding and Retrieval

Authors: Hang Gao, Wujiang Xu, Kai Mei, Dimitris N. Metaxas
Published: 2026-03-22
Categories: cs.CL, cs.IR
arXiv: https://arxiv.org/abs/2603.21437v2

Transformer-based embedding models frequently exhibit geometric pathologies, such as anisotropy and length-induced representation collapse, which can degrade downstream retrieval performance. While prior work often attributes these issues directly to text length or attention mechanisms, we argue that the fundamental drivers are instead the inherent pooling operations coupled with internal semantic shift. In this paper, we establish a unified theoretical framework proving that contextual pooling intrinsically causes embedding collapse. Specifically, we mathematically prove that pooling semantically diverse sentences inevitably leads to micro-level semantic dilution, and strictly reduces the Mean Pairwise Distance of the vector space, guaranteeing macro-level spatial concentration. Grounded in these geometric insights, we formally define semantic shift to capture the natural semantic evolution and dispersion within a text. Through carefully controlled experiments across diverse models and corpora, we disentangle text length from semantic content. We demonstrate that semantic shift is the primary predictor of severe embedding concentration. Crucially, our retrieval evaluations reveal that anisotropy is fundamentally harmful only when induced by strong semantic shifts, reconciling conflicting observations in prior literature and offering a principled explanation for the long-context challenges faced by modern embedding models.

---

### PAS3R: Pose-Adaptive Streaming 3D Reconstruction for Long Video Sequences

Authors: Lanbo Xu, Liang Guo, Caigui Jiang, Cheng Wang
Published: 2026-03-22
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.21436v1

Online monocular 3D reconstruction enables dense scene recovery from streaming video but remains fundamentally limited by the stability-adaptation dilemma: the reconstruction model must rapidly incorporate novel viewpoints while preserving previously accumulated scene structure. Existing streaming approaches rely on uniform or attention-based update mechanisms that often fail to account for abrupt viewpoint transitions, leading to trajectory drift and geometric inconsistencies over long sequences. We introduce PAS3R, a pose-adaptive streaming reconstruction framework that dynamically modulates state updates according to camera motion and scene structure. Our key insight is that frames contributing significant geometric novelty should exert stronger influence on the reconstruction state, while frames with minor viewpoint variation should prioritize preserving historical context. PAS3R operationalizes this principle through a motion-aware update mechanism that jointly leverages inter-frame pose variation and image frequency cues to estimate frame importance. To further stabilize long-horizon reconstruction, we introduce trajectory-consistent training objectives that incorporate relative pose constraints and acceleration regularization. A lightweight online stabilization module further suppresses high-frequency trajectory jitter and geometric artifacts without increasing memory consumption. Extensive experiments across multiple benchmarks demonstrate that PAS3R significantly improves trajectory accuracy, depth estimation, and point cloud reconstruction quality in long video sequences while maintaining competitive performance on shorter sequences.

---

### Behavioural feasible set: Value alignment constraints on AI decision support

Authors: Taejin Park
Published: 2026-03-22
Categories: cs.AI, econ.GN
arXiv: https://arxiv.org/abs/2603.21435v1

When organisations adopt commercial AI systems for decision support, they inherit value judgements embedded by vendors that are neither transparent nor renegotiable. The governance puzzle is not whether AI can support decisions but which recommendations the system can actually produce given how its vendor has configured it. I formalise this as a behavioural feasible set, the range of recommendations reachable under vendor-imposed alignment constraints, and characterise diagnostic thresholds for when organisational requirements exceed the system's flexibility. In scenario-based experiments using binary decision scenarios and multi-stakeholder ranking tasks, I show that alignment materially compresses this set. Comparing pre- and post-alignment variants of an open-weight model isolates the mechanism: alignment makes the system substantially less able to shift its recommendation even under legitimate contextual pressure. Leading commercial models exhibit comparable or greater rigidity. In multi-stakeholder tasks, alignment shifts implied stakeholder priorities rather than neutralising them, meaning organisations adopt embedded value orientations set upstream by the vendor. Organisations thus face a governance problem that better prompting cannot resolve: selecting a vendor partially determines which trade-offs remain negotiable and which stakeholder priorities are structurally embedded.

---

### Image-Based Structural Analysis Using Computer Vision and LLMs: PhotoBeamSolver

Authors: Altamirano-Muñiz Emilio Fernando
Published: 2026-03-22
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.21432v1

This paper presents the development of a documented program capable of solving idealized beam models, such as those commonly used in textbooks and academic exercises, from drawings made by a person. The system is based on computer vision and statistical learning techniques for the detection and visual interpretation of structural elements. Likewise, the main challenges and limitations associated with the integration of computer vision into structural analysis are analyzed, as well as the requirements necessary for its reliable application in the field of civil engineering. In this context, the implementation of the PhotoBeamSolver program is explored, and the current state of computer vision in civil engineering is discussed, particularly in relation to structural analysis, infrastructure inspection, and engineering decision-support systems.

---

### DomAgent: Leveraging Knowledge Graphs and Case-Based Reasoning for Domain-Specific Code Generation

Authors: Shuai Wang, Dhasarathy Parthasarathy, Robert Feldt, Yinan Yu
Published: 2026-03-22
Categories: cs.AI, cs.SE
arXiv: https://arxiv.org/abs/2603.21430v2

Large language models (LLMs) have shown impressive capabilities in code generation. However, because most LLMs are trained on public domain corpora, directly applying them to real-world software development often yields low success rates, as these scenarios frequently require domain-specific knowledge. In particular, domain-specific tasks usually demand highly specialized solutions, which are often underrepresented or entirely absent in the training data of generic LLMs. To address this challenge, we propose DomAgent, an autonomous coding agent that bridges this gap by enabling LLMs to generate domain-adapted code through structured reasoning and targeted retrieval. A core component of DomAgent is DomRetriever, a novel retrieval module that emulates how humans learn domain-specific knowledge, by combining conceptual understanding with experiential examples. It dynamically integrates top-down knowledge-graph reasoning with bottom-up case-based reasoning, enabling iterative retrieval and synthesis of structured knowledge and representative cases to ensure contextual relevance and broad task coverage. DomRetriever can operate as part of DomAgent or independently with any LLM for flexible domain adaptation. We evaluate DomAgent on an open benchmark dataset in the data science domain (DS-1000) and further apply it to real-world truck software development tasks. Experimental results show that DomAgent significantly enhances domain-specific code generation, enabling small open-source models to close much of the performance gap with large proprietary LLMs in complex, real-world applications. The code is available at: https://github.com/Wangshuaiia/DomAgent.

---

### Uncertainty-Aware Knowledge Distillation for Multimodal Large Language Models

Authors: Jingchen Sun, Shaobo Han, Deep Patel, Wataru Kohno, Can Jin, Changyou Chen
Published: 2026-03-22
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.21426v1

Knowledge distillation establishes a learning paradigm that leverages both data supervision and teacher guidance. However, determining the optimal balance between learning from data and learning from the teacher is challenging, as some samples may be noisy while others are subject to teacher uncertainty. This motivates the need for adaptively balancing data and teacher supervision. We propose Beta-weighted Knowledge Distillation (Beta-KD), an uncertainty-aware distillation framework that adaptively modulates how much the student relies on teacher guidance. Specifically, we formulate teacher--student learning from a unified Bayesian perspective and interpret teacher supervision as a Gibbs prior over student activations. This yields a closed-form, uncertainty-aware weighting mechanism and supports arbitrary distillation objectives and their combinations. Extensive experiments on multimodal VQA benchmarks demonstrate that distilling student Vision-Language Models from a large teacher VLM consistently improves performance. The results show that Beta-KD outperforms existing knowledge distillation methods. The code is available at https://github.com/Jingchensun/beta-kd.

---

### HyReach: Vision-Guided Hybrid Manipulator Reaching in Unseen Cluttered Environments

Authors: Shivani Kamtikar, Kendall Koe, Justin Wasserman, Samhita Marri, Benjamin Walt, Naveen Kumar Uppalapati, Girish Krishnan, Girish Chowdhary
Published: 2026-03-22
Categories: cs.RO, cs.AI
arXiv: https://arxiv.org/abs/2603.21421v1

As robotic systems increasingly operate in unstructured, cluttered, and previously unseen environments, there is a growing need for manipulators that combine compliance, adaptability, and precise control. This work presents a real-time hybrid rigid-soft continuum manipulator system designed for robust open-world object reaching in such challenging environments. The system integrates vision-based perception and 3D scene reconstruction with shape-aware motion planning to generate safe trajectories. A learning-based controller drives the hybrid arm to arbitrary target poses, leveraging the flexibility of the soft segment while maintaining the precision of the rigid segment. The system operates without environment-specific retraining, enabling direct generalization to new scenes. Extensive real-world experiments demonstrate consistent reaching performance with errors below 2 cm across diverse cluttered setups, highlighting the potential of hybrid manipulators for adaptive and reliable operation in unstructured environments.

---

### Is the future of AI green? What can innovation diffusion models say about generative AI's environmental impact?

Authors: Robert Viseur, Nicolas Jullien
Published: 2026-03-22
Categories: cs.AI
arXiv: https://arxiv.org/abs/2603.21419v1

The rise of generative artificial intelligence (GAI) has led to alarming predictions about its environmental impact. However, these predictions often overlook the fact that the diffusion of innovation is accompanied by the evolution of products and the optimization of their performance, primarily for economic reasons. This can also reduce their environmental impact. By analyzing the GAI ecosystem using the classic A-U innovation diffusion model, we can forecast this industry's structure and how its environmental impact will evolve. While GAI will never be green, its impact may not be as problematic as is sometimes claimed. However, this depends on which business model becomes dominant.

---

### Efficient Fine-Tuning Methods for Portuguese Question Answering: A Comparative Study of PEFT on BERTimbau and Exploratory Evaluation of Generative LLMs

Authors: Mariela M. Nina, Caio Veloso Costa, Lilian Berton, Didier A. Vega-Oliveros
Published: 2026-03-22
Categories: cs.CL, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2603.21418v1

Although large language models have transformed natural language processing, their computational costs create accessibility barriers for low-resource languages such as Brazilian Portuguese. This work presents a systematic evaluation of Parameter-Efficient Fine-Tuning (PEFT) and quantization techniques applied to BERTimbau for Question Answering on SQuAD-BR, the Brazilian Portuguese translation of SQuAD v1. We evaluate 40 configurations combining four PEFT methods (LoRA, DoRA, QLoRA, QDoRA) across two model sizes (Base: 110M, Large: 335M parameters). Our findings reveal three critical insights: (1) LoRA achieves 95.8\% of baseline performance on BERTimbau-Large while reducing training time by 73.5\% (F1=81.32 vs 84.86); (2) higher learning rates (2e-4) substantially improve PEFT performance, with F1 gains of up to +19.71 points over standard rates; and (3) larger models show twice the quantization resilience (loss of 4.83 vs 9.56 F1 points). These results demonstrate that encoder-based models can be efficiently fine-tuned for extractive Brazilian Portuguese QA with substantially lower computational cost than large generative LLMs, promoting more sustainable approaches aligned with \textit{Green AI} principles. An exploratory evaluation of Tucano and Sabiá on the same extractive QA benchmark shows that while generative models can reach competitive F1 scores with LoRA fine-tuning, they require up to 4.2$\times$ more GPU memory and 3$\times$ more training time than BERTimbau-Base, reinforcing the efficiency advantage of smaller encoder-based architectures for this task.

---

### Silent Commitment Failure in Instruction-Tuned Language Models: Evidence of Governability Divergence Across Architectures

Authors: Gregory M. Ruddell
Published: 2026-03-22
Categories: cs.AI, cs.CR, cs.LG
arXiv: https://arxiv.org/abs/2603.21415v1

As large language models are deployed as autonomous agents with tool execution privileges, a critical assumption underpins their security architecture: that model errors are detectable at runtime. We present empirical evidence that this assumption fails for two of three instruction-following models evaluable for conflict detection. We introduce governability -- the degree to which a model's errors are detectable before output commitment and correctable once detected -- and demonstrate it varies dramatically across models. In six models across twelve reasoning domains, two of three instruction-following models exhibited silent commitment failure: confident, fluent, incorrect output with zero warning signal. The remaining model produced a detectable conflict signal 57 tokens before commitment under greedy decoding. We show benchmark accuracy does not predict governability, correction capacity varies independently of detection, and identical governance scaffolds produce opposite effects across models. A 2x2 experiment shows a 52x difference in spike ratio between architectures but only +/-0.32x variation from fine-tuning, suggesting governability is fixed at pretraining. We propose a Detection and Correction Matrix classifying model-task combinations into four regimes: Governable, Monitor Only, Steer Blind, and Ungovernable.

---

### Fingerprinting Deep Neural Networks for Ownership Protection: An Analytical Approach

Authors: Guang Yang, Ziye Geng, Yihang Chen, Changqing Luo
Published: 2026-03-22
Categories: cs.CR, cs.AI
arXiv: https://arxiv.org/abs/2603.21411v1

Adversarial-example-based fingerprinting approaches, which leverage the decision boundary characteristics of deep neural networks (DNNs) to craft fingerprints, have proven effective for model ownership protection. However, a fundamental challenge remains unresolved: how far a fingerprint should be placed from the decision boundary to simultaneously satisfy two essential properties, i.e., robustness and uniqueness, for effective and reliable ownership protection. Despite the importance of the fingerprint-to-boundary distance, existing works lack a theoretical solution and instead rely on empirical heuristics, which may violate either robustness or uniqueness properties. We propose AnaFP, an analytical fingerprinting scheme that constructs fingerprints under theoretical guidance. Specifically, we formulate fingerprint generation as controlling the fingerprint-to-boundary distance through a tunable stretch factor. To ensure both robustness and uniqueness, we mathematically formalize these properties that determine the lower and upper bounds of the stretch factor. These bounds jointly define an admissible interval within which the stretch factor must lie, thereby establishing a theoretical connection between the two constraints and the fingerprint-to-boundary distance. To enable practical fingerprint generation, we approximate the original (infinite) sets of pirated and independently trained models using two finite surrogate model pools and employ a quantile-based relaxation strategy to relax the derived bounds. Due to the circular dependency between the lower bound and the stretch factor, we apply grid search over the admissible interval to determine the most feasible stretch factor. Extensive experimental results show that AnaFP consistently outperforms prior methods, achieving effective ownership verification across diverse model architectures and model modification attacks.

---

### Multi-Perspective LLM Annotations for Valid Analyses in Subjective Tasks

Authors: Navya Mehrotra, Adam Visokay, Kristina Gligorić
Published: 2026-03-22
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.21404v1

Large language models are increasingly used to annotate texts, but their outputs reflect some human perspectives better than others. Existing methods for correcting LLM annotation error assume a single ground truth. However, this assumption fails in subjective tasks where disagreement across demographic groups is meaningful. Here we introduce Perspective-Driven Inference, a method that treats the distribution of annotations across groups as the quantity of interest, and estimates it using a small human annotation budget. We contribute an adaptive sampling strategy that concentrates human annotation effort on groups where LLM proxies are least accurate. We evaluate on politeness and offensiveness rating tasks, showing targeted improvements for harder-to-model demographic groups relative to uniform sampling baselines, while maintaining coverage.

---

### The Myhill-Nerode Theorem for Bounded Interaction: Canonical Abstractions via Agent-Bounded Indistinguishability

Authors: Anthony T. Nixon
Published: 2026-03-22
Categories: cs.AI
arXiv: https://arxiv.org/abs/2603.21399v1

Any capacity-limited observer induces a canonical quotient on its environment: two situations that no bounded agent can distinguish are, for that agent, the same. We formalise this for finite POMDPs. A fixed probe family of finite-state controllers induces a closed-loop Wasserstein pseudometric on observation histories and a probe-exact quotient merging histories that no controller in the family can distinguish. The quotient is canonical, minimal, and unique-a bounded-interaction analogue of the Myhill-Nerode theorem. For clock-aware probes, it is exactly decision-sufficient for objectives that depend only on the agent's observations and actions; for latent-state rewards, we use an observation-Lipschitz approximation bound. The main theorem object is the clock-aware quotient; scalable deterministic-stationary experiments study a tractable coarsening with gap measured on small exact cases and explored empirically at larger scale. We validate theorem-level claims on Tiger and GridWorld. We also report operational case studies on Tiger, GridWorld, and RockSample as exploratory diagnostics of approximation behavior and runtime, not as theorem-facing evidence when no exact cross-family certificate is available; heavier stress tests are archived in the appendix and artifact package.

---

### Persona Vectors in Games: Measuring and Steering Strategies via Activation Vectors

Authors: Johnathan Sun, Andrew Zhang
Published: 2026-03-22
Categories: cs.AI, cs.GT
arXiv: https://arxiv.org/abs/2603.21398v1

Large language models (LLMs) are increasingly deployed as autonomous decision-makers in strategic settings, yet we have limited tools for understanding their high-level behavioral traits. We use activation steering methods in game-theoretic settings, constructing persona vectors for altruism, forgiveness, and expectations of others by contrastive activation addition. Evaluating on canonical games, we find that activation steering systematically shifts both quantitative strategic choices and natural-language justifications. However, we also observe that rhetoric and strategy can diverge under steering. In addition, vectors for self-behavior and expectations of others are partially distinct. Our results suggest that persona vectors offer a promising mechanistic handle on high-level traits in strategic environments.

---

### Mechanisms of Introspective Awareness

Authors: Uzay Macar, Li Yang, Atticus Wang, Peter Wallich, Emmanuel Ameisen, Jack Lindsey
Published: 2026-03-22
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.21396v5

Recent work has shown that LLMs can sometimes detect when steering vectors are injected into their residual stream and identify the injected concept -- a phenomenon termed "introspective awareness." We investigate the mechanisms underlying this capability in open-weights models. First, we find that it is behaviorally robust: models detect injected steering vectors at moderate rates with 0% false positives across diverse prompts and dialogue formats. Notably, this capability emerges specifically from post-training; we show that preference optimization algorithms like DPO can elicit it, but standard supervised finetuning does not. We provide evidence that detection cannot be explained by simple linear association between certain steering vectors and directions promoting affirmative responses. We trace the detection mechanism to a two-stage circuit in which "evidence carrier" features in early post-injection layers detect perturbations monotonically along diverse directions, suppressing downstream "gate" features that implement a default negative response. This circuit is absent in base models and robust to refusal ablation. Identification of injected concepts relies on largely distinct later-layer mechanisms that only weakly overlap with those involved in detection. Finally, we show that introspective capability is substantially underelicited: ablating refusal directions improves detection by +53%, and a trained bias vector improves it by +75% on held-out concepts, both without meaningfully increasing false positives. Our results suggest that this introspective awareness of injected concepts is robust and mechanistically nontrivial, and could be substantially amplified in future models. Code: https://github.com/safety-research/introspection-mechanisms.

---

### A Generalised Exponentiated Gradient Approach to Enhance Fairness in Binary and Multi-class Classification Tasks

Authors: Maryam Boubekraoui, Giordano d'Aloisio, Antinisca Di Marco
Published: 2026-03-22
Categories: cs.LG, stat.ML
arXiv: https://arxiv.org/abs/2603.21393v2

The widespread use of AI and ML models in sensitive areas raises significant concerns about fairness. While the research community has introduced various methods for bias mitigation in binary classification tasks, the issue remains under-explored in multi-class classification settings. To address this limitation, in this paper, we first formulate the problem of fair learning in multi-class classification as a multi-objective problem between effectiveness (i.e., prediction correctness) and multiple linear fairness constraints. Next, we propose a Generalised Exponentiated Gradient (GEG) algorithm to solve this task. GEG is an in-processing algorithm that enhances fairness in binary and multi-class classification settings under multiple fairness definitions. We conduct an extensive empirical evaluation of GEG against six baselines across seven multi-class and three binary datasets, using four widely adopted effectiveness metrics and three fairness definitions.

---

### Task-Specific Efficiency Analysis: When Small Language Models Outperform Large Language Models

Authors: Jinghan Cao, Yu Ma, Xinjin Li, Qingyang Ren, Xiangyun Chen
Published: 2026-03-22
Categories: cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2603.21389v1

Large Language Models achieve remarkable performance but incur substantial computational costs unsuitable for resource-constrained deployments. This paper presents the first comprehensive task-specific efficiency analysis comparing 16 language models across five diverse NLP tasks. We introduce the Performance-Efficiency Ratio (PER), a novel metric integrating accuracy, throughput, memory, and latency through geometric mean normalization. Our systematic evaluation reveals that small models (0.5--3B parameters) achieve superior PER scores across all given tasks. These findings establish quantitative foundations for deploying small models in production environments prioritizing inference efficiency over marginal accuracy gains.

---

### Knowledge Priors for Identity-Disentangled Open-Set Privacy-Preserving Video FER

Authors: Feng Xu, Xun Li, Lars Petersson, Yulei Sui, David Ahmedt-Aristizabal, Dadong Wang
Published: 2026-03-22
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.21387v2

Facial expression recognition relies on facial data that inherently expose identity and thus raise significant privacy concerns. Current privacy-preserving methods typically fail in realistic open-set video settings where identities are unknown, and identity labels are unavailable. We propose a two-stage framework for video-based privacy-preserving FER in challenging open-set settings that requires no identity labels at any stage. To decouple privacy and utility, we first train an identity-suppression network using intra- and inter-video knowledge priors derived from real-world videos without identity labels. This network anonymizes identity while preserving expressive cues. A subsequent denoising module restores expression-related information and helps recover FER performance. Furthermore, we introduce a falsification-based validation method that uses recognition priors to rigorously evaluate privacy robustness without requiring annotated identity labels. Experiments on three video datasets demonstrate that our method effectively protects privacy while maintaining FER accuracy comparable to identity-supervised baselines.

---

### Mitigating Objectness Bias and Region-to-Text Misalignment for Open-Vocabulary Panoptic Segmentation

Authors: Nikolay Kormushev, Josip Šarić, Matej Kristan
Published: 2026-03-22
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.21386v1

Open-vocabulary panoptic segmentation remains hindered by two coupled issues: (i) mask selection bias, where objectness heads trained on closed vocabularies suppress masks of categories not observed in training, and (ii) limited regional understanding in vision-language models such as CLIP, which were optimized for global image classification rather than localized segmentation. We introduce OVRCOAT, a simple, modular framework that tackles both. First, a CLIP-conditioned objectness adjustment (COAT) updates background/foreground probabilities, preserving high-quality masks for out-of-vocabulary objects. Second, an open-vocabulary mask-to-text refinement (OVR) strengthens CLIP's region-level alignment to improve classification of both seen and unseen classes with markedly lower memory cost than prior fine-tuning schemes. The two components combine to jointly improve objectness estimation and mask recognition, yielding consistent panoptic gains. Despite its simplicity, OVRCOAT sets a new state of the art on ADE20K (+5.5% PQ) and delivers clear gains on Mapillary Vistas and Cityscapes (+7.1% and +3% PQ, respectively). The code is available at: https://github.com/nickormushev/OVRCOAT

---

### PivotRL: High Accuracy Agentic Post-Training at Low Compute Cost

Authors: Junkeun Yi, Damon Mosk-Aoyama, Baihe Huang, Ritu Gala, Charles Wang, Sugam Dipak Devare, Khushi Bhardwaj, Abhibha Gupta, Oleksii Kuchaiev, Jiantao Jiao, Jian Zhang, Venkat Srinivasan
Published: 2026-03-22
Categories: cs.AI
arXiv: https://arxiv.org/abs/2603.21383v1

Post-training for long-horizon agentic tasks has a tension between compute efficiency and generalization. While supervised fine-tuning (SFT) is compute efficient, it often suffers from out-of-domain (OOD) degradation. Conversely, end-to-end reinforcement learning (E2E RL) preserves OOD capabilities, but incurs high compute costs due to many turns of on-policy rollout. We introduce PivotRL, a novel framework that operates on existing SFT trajectories to combine the compute efficiency of SFT with the OOD accuracy of E2E RL. PivotRL relies on two key mechanisms: first, it executes local, on-policy rollouts and filters for pivots: informative intermediate turns where sampled actions exhibit high variance in outcomes; second, it utilizes rewards for functional-equivalent actions rather than demanding strict string matching with the SFT data demonstration. We theoretically show that these mechanisms incentivize strong learning signals with high natural gradient norm, while maximally preserving policy probability ordering on actions unrelated to training tasks. In comparison to standard SFT on identical data, we demonstrate that PivotRL achieves +4.17% higher in-domain accuracy on average across four agentic domains, and +10.04% higher OOD accuracy in non-agentic tasks. Notably, on agentic coding tasks, PivotRL achieves competitive accuracy with E2E RL with 4x fewer rollout turns. PivotRL is adopted by NVIDIA's Nemotron-3-Super-120B-A12B, acting as the workhorse in production-scale agentic post-training.

---

### An InSAR Phase Unwrapping Framework for Large-scale and Complex Events

Authors: Yijia Song, Juliet Biggs, Alin Achim, Robert Popescu, Simon Orrego, Nantheera Anantrasirichai
Published: 2026-03-22
Categories: cs.CV, cs.AI, physics.geo-ph
arXiv: https://arxiv.org/abs/2603.21378v2

Phase unwrapping remains a critical and challenging problem in InSAR processing, particularly in scenarios involving complex deformation patterns. In earthquake-related deformation, shallow sources can generate surface-breaking faults and abrupt displacement discontinuities, which severely disrupt phase continuity and often cause conventional unwrapping algorithms to fail. Another limitation of existing learning-based unwrapping methods is their reliance on fixed and relatively small input sizes, while real InSAR interferograms are typically large-scale and spatially heterogeneous. This mismatch restricts the applicability of many neural network approaches to real-world data. In this work, we present a phase unwrapping framework based on a diffusion model, developed to process large-scale interferograms and to address phase discontinuities caused by deformation. By leveraging a diffusion model architecture, the proposed method can recover physically consistent unwrapped phase fields even in the presence of fault-related phase jumps. Experimental results on both synthetic and real datasets demonstrate that the method effectively addresses discontinuities associated with near-surface deformation and scales well to large InSAR images, offering a practical alternative to manual unwrapping in challenging scenarios.

---

### An InSAR Phase Unwrapping Framework for Large-scale and Complex Events

Authors: Yijia Song, Juliet Biggs, Alin Achim, Robert Popescu, Simon Orrego, Nantheera Anantrasirichai
Published: 2026-03-22
Categories: cs.CV, cs.AI, physics.geo-ph
arXiv: https://arxiv.org/abs/2603.21378v1

Phase unwrapping remains a critical and challenging problem in InSAR processing, particularly in scenarios involving complex deformation patterns. In earthquake-related deformation, shallow sources can generate surface-breaking faults and abrupt displacement discontinuities, which severely disrupt phase continuity and often cause conventional unwrapping algorithms to fail. Another limitation of existing learning-based unwrapping methods is their reliance on fixed and relatively small input sizes, while real InSAR interferograms are typically large-scale and spatially heterogeneous. This mismatch restricts the applicability of many neural network approaches to real-world data. In this work, we present a phase unwrapping framework based on a diffusion model, developed to process large-scale interferograms and to address phase discontinuities caused by deformation. By leveraging a diffusion model architecture, the proposed method can recover physically consistent unwrapped phase fields even in the presence of fault-related phase jumps. Experimental results on both synthetic and real datasets demonstrate that the method effectively addresses discontinuities associated with near-surface deformation and scales well to large InSAR images, offering a practical alternative to manual unwrapping in challenging scenarios.

---

### HamVision: Hamiltonian Dynamics as Inductive Bias for Medical Image Analysis

Authors: Mohamed A Mabrok
Published: 2026-03-22
Categories: cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2603.21377v1

We present HamVision, a framework for medical image analysis that uses the damped harmonic oscillator, a fundamental building block of signal processing, as a structured inductive bias for both segmentation and classification tasks. The oscillator's phase-space decomposition yields three functionally distinct representations: position~$q$ (feature content), momentum~$p$ (spatial gradients that encode boundary and texture information), and energy $H = \tfrac{1}{2}|z|^2$ (a parameter-free saliency map). These representations emerge from the dynamics, not from supervision, and can be exploited by different task-specific heads without any modification to the oscillator itself. For segmentation, energy gates the skip connections while momentum injects boundary information at every decoder level (HamSeg). For classification, the three representations are globally pooled and concatenated into a phase-space feature vector (HamCls). We evaluate HamVision across ten medical imaging benchmarks spanning five imaging modalities. On segmentation, HamSeg achieves state-of-the-art Dice scores on ISIC\,2018 (89.38\%), ISIC\,2017 (88.40\%), TN3K (87.05\%), and ACDC (92.40\%), outperforming most baselines with only 8.57M parameters. On classification, HamCls achieves state-of-the-art accuracy on BloodMNIST (98.85\%) and PathMNIST (96.65\%), and competitive results on the remaining MedMNIST datasets against MedMamba and MedViT. Diagnostic analysis confirms that the oscillator's momentum consistently encodes an interior$\,{>}\,$boundary$\,{>}\,$exterior gradient for segmentation and that the energy map correlates with discriminative regions for classification, properties that emerge entirely from the Hamiltonian dynamics. Code is available at https://github.com/Minds-R-Lab/hamvision.

---

### A transformer architecture alteration to incentivise externalised reasoning

Authors: Elizabeth Pavlova, Mariia Koroliuk, Karthik Viswanathan, Cameron Tice, Edward James Young, Puria Radmard
Published: 2026-03-22
Categories: cs.AI
arXiv: https://arxiv.org/abs/2603.21376v2

We propose a new architectural change, and post-training pipeline, for making LLMs more verbose reasoners by teaching a model to truncate forward passes early. We augment an existing transformer architecture with an early-exit mechanism at intermediate layers and train the model to exit at shallower layers when the next token can be predicted without deep computation. After a calibration stage, we incentivise the model to exit as early as possible while maintaining task performance using reinforcement learning. We provide preliminary results to this effect for small reasoning models, showing that they learn to adaptively reduce computations across tokens. We predict that, applied at the right scale, our approach can minimise the amount of excess computation that reasoning models have at their disposal to perform non-myopic planning using their internal activations, reserving this only for difficult-to-predict tokens.

---

### Constrained Online Convex Optimization with Memory and Predictions

Authors: Mohammed Abdullah, George Iosifidis, Salah Eddine Elayoubi, Tijani Chahed
Published: 2026-03-22
Categories: cs.LG, stat.ML
arXiv: https://arxiv.org/abs/2603.21375v1

We study Constrained Online Convex Optimization with Memory (COCO-M), where both the loss and the constraints depend on a finite window of past decisions made by the learner. This setting extends the previously studied unconstrained online optimization with memory framework and captures practical problems such as the control of constrained dynamical systems and scheduling with reconfiguration budgets. For this problem, we propose the first algorithms that achieve sublinear regret and sublinear cumulative constraint violation under time-varying constraints, both with and without predictions of future loss and constraint functions. Without predictions, we introduce an adaptive penalty approach that guarantees sublinear regret and constraint violation. When short-horizon and potentially unreliable predictions are available, we reinterpret the problem as online learning with delayed feedback and design an optimistic algorithm whose performance improves as prediction accuracy improves, while remaining robust when predictions are inaccurate. Our results bridge the gap between classical constrained online convex optimization and memory-dependent settings, and provide a versatile learning toolbox with diverse applications.

---

### PLR: Plackett-Luce for Reordering In-Context Learning Examples

Authors: Pawel Batorski, Paul Swoboda
Published: 2026-03-22
Categories: cs.LG, cs.CL
arXiv: https://arxiv.org/abs/2603.21373v2

In-context learning (ICL) adapts large language models by conditioning on a small set of ICL examples, avoiding costly parameter updates. Among other factors, performance is often highly sensitive to the ordering of the examples. However, exhaustive search over the $n!$ possible orderings is infeasible. Therefore more efficient ordering methods use model confidence measures (e.g., label-probability entropy) over label sets or take a direct approach to finding the best ordering. We propose PLR, a probabilistic approach to in-context example ordering that replaces discrete ordering search with learning a probability distribution over orderings with the Plackett-Luce model. PLR models orderings using a Plackett-Luce distribution and iteratively updates its parameters to concentrate probability mass on high-performing orderings under a task-level metric. Candidate orderings are sampled efficiently via a Gumbel perturb-and-sort procedure. Experiments on multiple classification benchmarks show that PLR consistently improves few-shot accuracy for $k \in \{4, 8, 16, 32\}$ examples, and we further demonstrate gains on mathematical reasoning tasks where label-based ordering methods are not applicable. Our code is available at https://github.com/Batorskq/PLR.

---

### Conspiracy Frame: a Semiotically-Driven Approach for Conspiracy Theories Detection

Authors: Heidi Campana Piva, Shaina Ashraf, Maziar Kianimoghadam Jouneghani, Arianna Longo, Rossana Damiano, Lucie Flek, Marco Antonio Stranisci
Published: 2026-03-22
Categories: cs.CL
arXiv: https://arxiv.org/abs/2603.21368v1

Conspiracy theories are anti-authoritarian narratives that lead to social conflict, impacting how people perceive political information. To help in understanding this issue, we introduce the Conspiracy Frame: a fine-grained semantic representation of conspiratorial narratives derived from frame-semantics and semiotics, which spawned the Conspiracy Frames (Con.Fra.) dataset: a corpus of Telegram messages annotated at span-level. The Conspiracy Frame and Con.Fra. dataset contribute to the implementation of a more generalizable understanding and recognition of conspiracy theories. We observe the ability of LLMs to recognize this phenomenon in-domain and out-of-domain, investigating the role that frames may have in supporting this task. Results show that, while the injection of frames in an in-context approach does not lead to clear increase of performance, it has potential; the mapping of annotated spans with FrameNet shows abstract semantic patterns (e.g., `Kinship', `Ingest\_substance') that potentially pave the way for a more semantically- and semiotically-aware detection of conspiratorial narratives.

---

### Relax Forcing: Relaxed KV-Memory for Consistent Long Video Generation

Authors: Zengqun Zhao, Yanzuo Lu, Ziquan Liu, Jifei Song, Jiankang Deng, Ioannis Patras
Published: 2026-03-22
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.21366v2

Autoregressive video diffusion has recently emerged as a promising paradigm for long-video generation, enabling causal synthesis beyond the temporal limits of bidirectional models. Existing forcing-based training strategies reduce exposure bias by conditioning models on their own predictions during rollout, yet minute-scale generation remains challenging due to progressive temporal degradation and constrained motion evolution. In this work, we study the role of temporal KV memory during long-horizon autoregressive inference. Our analysis shows that simply retaining more historical frames does not consistently improve generation quality; instead, both the quantity and temporal placement of memory strongly affect motion dynamics. These findings suggest that temporal memory should be treated as structured context rather than a homogeneous chronological buffer. Motivated by this observation, we introduce Relax Forcing, a training-free memory mechanism for autoregressive video diffusion. Relax Forcing decomposes temporal context into three functional components: Sink frames that provide global stability, Tail frames that preserve short-term continuity, and dynamically selected History frames that supply mid-range motion structure. History frames are selected using a relaxation-based criterion that encourages alignment with global anchors while suppressing redundancy with recent context. This role-aware sparse memory design mitigates error accumulation during long-horizon rollout while preserving motion evolution and reducing attention overhead. Experiments on VBench-Long show that Relax Forcing improves long-video generation, achieving stronger motion dynamics and higher overall scores than existing autoregressive baselines. These results indicate that structured temporal memory is an effective and complementary direction for scalable long-video generation.

---

### Relax Forcing: Relaxed KV-Memory for Consistent Long Video Generation

Authors: Zengqun Zhao, Yanzuo Lu, Ziquan Liu, Jifei Song, Jiankang Deng, Ioannis Patras
Published: 2026-03-22
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.21366v1

Autoregressive (AR) video diffusion has recently emerged as a promising paradigm for long video generation, enabling causal synthesis beyond the limits of bidirectional models. To address training-inference mismatch, a series of self-forcing strategies have been proposed to improve rollout stability by conditioning the model on its own predictions during training. While these approaches substantially mitigate exposure bias, extending generation to minute-scale horizons remains challenging due to progressive temporal degradation. In this work, we show that this limitation is not primarily caused by insufficient memory, but by how temporal memory is utilised during inference. Through empirical analysis, we find that increasing memory does not consistently improve long-horizon generation, and that the temporal placement of historical context significantly influences motion dynamics while leaving visual quality largely unchanged. These findings suggest that temporal memory should not be treated as a homogeneous buffer. Motivated by this insight, we introduce Relax Forcing, a structured temporal memory mechanism for AR diffusion. Instead of attending to the dense generated history, Relax Forcing decomposes temporal context into three functional roles: Sink for global stability, Tail for short-term continuity, and dynamically selected History for structural motion guidance, and selectively incorporates only the most relevant past information. This design mitigates error accumulation during extrapolation while preserving motion evolution. Experiments on VBench-Long demonstrate that Relax Forcing improves motion dynamics and overall temporal consistency while reducing attention overhead. Our results suggest that structured temporal memory is essential for scalable long video generation, complementing existing forcing-based training strategies.

---

### TIDE: Token-Informed Depth Execution for Per-Token Early Exit in LLM Inference

Authors: Jaber Jaber, Osama Jaber
Published: 2026-03-22
Categories: cs.LG, cs.CL
arXiv: https://arxiv.org/abs/2603.21365v1

Large language models run every token through every layer, regardless of difficulty. We present TIDE, a post-training system that attaches tiny learned routers at periodic checkpoint layers and, at inference time, selects the earliest layer whose hidden state has converged for each token. TIDE requires no model retraining, works with any HuggingFace causal LM, auto-detects GPU architecture, and supports float32, float16, and bfloat16 through fused CUDA kernels. On an NVIDIA A100 with DeepSeek R1 Distill 8B, TIDE achieves 100% prefill exit rate (5% of tokens exit at layer 11, the remaining at layer 31), reduces prefill latency by 7.2%, and increases single-batch throughput by 6.6%. During autoregressive decoding, 98-99% of tokens exit early while the model correctly solves a multi-step math problem with 95 unique output tokens. On Qwen3 8B (36 layers), throughput improves by 8.1% at batch size 8. Calibration on 2,000 WikiText samples takes under 3 minutes and produces a ~4 MB router checkpoint. The system comprises 1,308 lines of Python and 1,081 lines of CUDA/C++ with 74 passing tests. Code: https://github.com/RightNow-AI/TIDE

---