Papers

Low-rank decomposition is a promising compression paradigm for large language models (LLMs), yet its effectiveness hinges on rank budget allocation across weight matrices: uniform or hand-crafted rules ignore module-wise importance, while learning-based allocation incurs substantial training overhead. We formulate rank allocation as a global sorting-and-truncation pipeline that scores every singular component by combining local singular energy with global functional importance, estimated via layer-wise input–output cosine similarity on a tiny calibration set. We show, both geometrically and empirically, that high input–output cosine similarity implies low effective rank. We further propose rank-preserving fine-tuning (RPFT), which adapts only a small subset of retained singular components so that the allocated rank stays bounded without re-decomposition. Experimental results show that UniRank cuts zero-shot perplexity by up to 50%, improves average reasoning accuracy by 3.0% over LoRAP at 25% sparsity, and boosts four SVD-based decomposition methods as a plug-and-play module.

AI agents in long-context applications drift from their specified identity. Current methods detect this only after qualitative degradation is visible. We present a geometric framework for measuring identity structure using $\sqrt{\mathrm{JSD}}$ metric spaces and magnitude homology from enriched category theory, where identity is non-geodesic structure and drift is its relaxation toward the geodesic. Validated on a persistent AI agent, the framework's strongest empirical finding is a two-mechanism conditioning structure: cross-condition distances reveal an identity-vacuum cluster where the identity specification fills a behavioral void, and a safety-basin cluster where it displaces from post-training attractors. An equilateral probe baseline confirms that the identity specification creates measurable behavioral richness (55 unique response patterns vs. 1 for the base model) at maximum probe separation. A first-order perturbation theory for equilateral configurations predicts magnitude changes from perimeter changes alone, with shape perturbations first-order cancelled by the $S_n$ symmetry; the formula is self-consistent at the observed perturbation amplitudes. A drift experiment measuring magnitude decrease under context pressure was subsequently found to reflect repetitive-padding artifacts rather than genuine context-length drift; diverse padding produces no measurable deformation through 150K tokens. The magnitude homology framework's full diagnostic promise – detecting anisotropic contraction and structural collapse via homological simplification – is architecturally grounded in the perturbation theory and selection rules but remains empirically unconfirmed.

Traditional architectural design space exploration (DSE) is highly inefficient, typically requiring tens of thousands of simulator evaluations across various optimization methods. This inefficiency arises because conventional methods treat the simulator as a black-box oracle. In contrast, human architects effectively guide exploration by reasoning through physical constraints, performance bottlenecks, data reuse, and workload structures. To bridge this gap, we introduce AgentDSE, a simulator-in-the-loop methodology driven by a general-purpose large language model (LLM) coding agent. AgentDSE automates this architectural-reasoning loop without requiring model fine-tuning, precomputed design databases, or domain-specific optimizer code. Across deep neural network (DNN) accelerator mapping, hardware/software co-design, and CPU cache-hierarchy optimization, AgentDSE achieves competitive or better design quality with up to two orders of magnitude fewer evaluations. AgentDSE also produces inspectable traces that surface architectural hypotheses, performance cliffs, implicit priors, and simulator artifacts, making every search decision traceable rather than buried in optimizer state.

Computerized Adaptive Testing (CAT), as a key technology for personalized education, aims to accurately assess examinee proficiency by retrieving exercises dynamically matching current ability estimates. However, existing CAT research is constrained by limitations of static offline data and isolated component optimization. Restricted by partial labels in offline logs, researchers degrade the dynamic assessment process into static sequence prediction. Current research focuses on isolated perspectives, e.g., selection or diagnosis, neglecting the overall CAT interaction process. To address this, we propose AgentCAT, a Large Language Model-based multi-agent simulation system, to construct a high-fidelity benchmarking environment for dynamic testing. This framework comprises three modules: (1) The examinee agent with memory retrieval and Chain-of-Thought reasoning simulates responses based on cognitive profiles; (2) The selection agent uses coarse-to-fine bucketing and knowledge graph exploration to balance local difficulty and global coverage; (3) The supervisor uses dual-auditing and robust update to ensure convergence and validity. To validate the framework, we evaluated on two real-world datasets across three dimensions: macro-level ability convergence, micro-level interaction logic, and data sparsity resilience. Results show AgentCAT achieves effective ability estimation, and its selection strategy balances difficulty adaptation and instructional coherence, aligning with human pedagogical intuition.

Memory safety errors account for a large proportion of security bugs in systems written in C; modern languages such as Java and Rust prevent such bugs because they are memory-safe by design. To migrate systems to safer languages or identify memory errors, we must first determine how legacy code manipulates memory. This information is only represented implicitly in such code. In many cases, memory usage patterns are merely tedious for humans to figure out, rather than truly difficult. In this work, we ask if large language models (LLMs) can perform this task by having them synthesize annotations representing memory usage as specifications in CN, a hybrid testing/verification tool. Our tool, CNnotator, uses LLMs to automatically generate and test CN specifications. We find that current models are able to generate CN specifications for small-to-medium C programs, with the OpenAI o3 reasoning model achieving a 90% success rate on first attempts and 97% overall success, while the chat model GPT-4o correctly annotates 65% of first attempts. These results suggest AI-assisted annotation is becoming practical for real-world C codebases.

Epidemiological models often rely on survey data to represent how individuals make health-related decisions, such as whether to vaccinate or adopt protective behaviors. However, repeated large-scale surveys are costly, time-consuming, and limited in the range of scenarios they can capture. In this work, we investigate whether large language models (LLMs) can generate synthetic survey responses that reproduce patterns observed in real populations. Using longitudinal data from the FluPaths surveys, we first identify groups associated with broadly positive or negative attitudes toward vaccination through clustering analysis. We then evaluate several LLMs using a cluster-informed prompting approach to generate synthetic survey responses across multiple epidemic waves. Across models, the synthetic data generally reproduce the distributions of demographic characteristics, vaccination-related beliefs, risk perceptions, and health behaviors observed in the survey data. However, they are less successful at capturing how these factors vary together within respondents. Some models reproduce group-level vaccination trends more reliably than others, although performance varies across waves. We also trained a classifier to distinguish real from synthetic records and found that the generated responses remained identifiable as synthetic. Overall, our findings suggest that LLM-generated survey data may provide a useful tool for exploratory data augmentation and we hope that it could support agent-based epidemic modeling approaches. However, the generated data should not be treated as a substitute for human survey data without further methodological improvements and validation.

Coding tasks are typically complicated and require multiple capabilities, ranging from high-level planning to low-level implementation. While coding agents are optimized for the joint capabilities, individual capabilities such as high-level planning may have different optima and remain a major bottleneck. To address this challenge, we train a separate critic model that is specialized in high-level planning to steer the coding agent in inference. We construct SFT and DPO data to train the critic model to identify errors made by the coding agent and provide correct and clear high-level guidance without generating concrete actions. Experiments show that our fine-tuned 4B and 8B critic models significantly improve the performance of 6 larger coding agents (e.g., improving the resolved rates of GLM-4.7-Flash-30B-A3B and GPT-OSS-120B by 16.0% and 14.4% on SWE-Bench Verified). The critic model also reduces the total inference costs for some coding agents by solving tasks in fewer steps (e.g., reducing the per-example inference cost for GPT-OSS-20B from \$0.07 to \$0.03). Code: https://github.com/shubhamrgandhi/critic-training

A common claim is that zero-shot large language models (LLMs) can replace fine-tuned NLU classifiers for intent detection. We test this claim head-to-head and find that the honest answer is: it depends on the intent space. On full ATIS and CLINC150 we compare a fine-tuned RoBERTa, a TF-IDF+logistic-regression baseline, sentence-embedding kNN, and Claude Haiku zero-shot, reporting bootstrap 95% confidence intervals and paired significance tests. When abundant in-domain labels exist, fine-tuned RoBERTa is as good or better and three orders of magnitude cheaper and faster: on ATIS it beats Claude zero-shot by 11.8 points (95.9 vs. 84.1, p<0.001). On the broad 150-intent CLINC150 schema the two are statistically tied (89.1 vs. 88.5, p=0.24): the LLM matches a fully supervised model with no training data. The LLM's advantages appear in three production-relevant regimes: out-of-scope detection (OOS recall 85.6 vs. 58.1 for RoBERTa); robustness to realistic ASR noise via a controlled text-to-speech to noise to Whisper pipeline (92.5 vs. 80.0 at 0 dB); and dynamic per-deployment schemas, where a classifier trained on one app's intents scores 0% on a new app's intents while the schema-prompted LLM serves both at ~94% with zero retraining. We distill these findings into a decision framework for practitioners.

Structured reporting converts free-text radiology narratives into queryable data keys, facilitating cohort assembly, longitudinal tracking, and training label generation for medical AI. The prevailing paradigm follows a two-stage pipeline: (1) constructing a reporting template, (2) extracting information to populate it. While the extraction stage has benefited from advances in large language models (LLMs), template construction remains a manual bottleneck relying on labor-intensive expert consensus that is static, difficult to scale, and may fail to capture real-world reporting diversity. We address this limitation with \textbf{ASTAR}, an LLM-based framework for Automated induction of STAndardized radiology Reporting templates from large-scale clinical free-text corpora. Extensive experiments on 4,215 fetal brain MRI reports from multiple centers demonstrate that the \textbf{ASTAR}-induced template surpasses two expert-curated templates across template coverage, information fidelity, diagnostic fidelity, and expert-rated usability, reducing template development from weeks of committee deliberation to hours of automated processing. Code: https://github.com/birthlab/ASTAR

Background: Recent advances in information retrieval (IR) leverage both dense and sparse representations, large language models (LLMs), and specialized retrieval models to improve ranking accuracy, relevance, and cross-lingual performance. Complementary techniques such as passage indexing, document layout analysis, and semantic knowledge representation further enhance retrieval effectiveness by capturing fine-grained contextual and structural information. Emerging agentic LLM frameworks extend these capabilities by enabling planning, iterative reasoning, tool use, and multi-agent collaboration, thereby broadening applications across diverse domains. These frameworks also emphasize rigorous evaluation, ethical considerations, and trustworthiness, ensuring responsible deployment in real-world settings. We propose a modular, agent-based pipeline for botanical trait extraction. Optical character recognition (OCR) converts PDFs into machine-readable text, while segmentation and indexing organize content by genus and species. Rule-based parsers extract structured botanical traits, and ensembles of large language models (LLMs) expand trait vocabularies and resolve ambiguities. This approach ensures accurate species recognition, scalable annotation, and explainable integration of textual botanical descriptions, enabling robust and interpretable data extraction across large botanical corpora. Results: Using three regional botanical datasets, our system extracted 55,737 trait annotations across 4,961 species, averaging 9.1 traits per species. Integration of LLM-based enrichment improved coverage for 75% of traits, increasing total annotations by 59%. While the choice of OCR engine had a minor effect on species recognition, overall annotation counts remained stable, demonstrating the robustness, scalability, and reliability of the pipeline for large-scale botanical trait extraction.

Vision-Language-Action (VLA) models are becoming a promising paradigm for autonomous driving, but their deployment on existing vehicle platforms remains difficult because they introduce both high inference latency and strong GPU-side resource pressure. In a full autonomous driving stack, this problem is even more pronounced: legacy vehicle platforms were provisioned for modular pipelines, yet after several planning-related functions are absorbed into a unified VLA model, part of the original CPU budget becomes underutilized, while the visual encoder and the main reasoning path still concentrate most computation and memory demand on the GPU. As a result, directly deploying VLA together with the rest of the onboard system can be hard under realistic GPU memory constraints. To address this issue, we present a hybrid CPU–GPU inference framework with flexible resource scheduling for autonomous driving. Our design partitions the VLA backbone at the block-layer granularity, executes the visual encoder and LLM prefix on the GPU, and offloads the LLM suffix to the CPU through a cross-frame asynchronous pipeline, thereby exposing a schedulable boundary for redistributing compute and memory pressure across heterogeneous processors. We evaluate the proposed framework on two representative driving VLA models, Orion and MindDrive. On Bench2Drive, our method reduces average latency from 521ms to 408.0ms for Orion and from 443ms to 306.2ms for MindDrive, corresponding to 21.7% and 30.9% reduction, respectively. For Orion, the estimated peak GPU memory is further reduced from 45GB to 29GB. In real-vehicle deployment under coexistence with Autoware.Universe, native Orion cannot run because the onboard GPU memory budget is insufficient, whereas the hybrid version runs successfully together with the full vehicle stack.

The development and benchmarking of Large Language Model (LLM)-based Motivational Interviewing (MI) counsellors now often rely on LLM-based simulated clients. Prior work on simulated clients, however, has not aligned with the specific tasks fundamental to the MI therapy approach. A key task is evoking, in which the counsellor first elicits the client's ambivalence and then strengthens the client's motivation for change. We present Evoke-Sim, a task-aware, multi-stage LLM-based client simulation framework for evaluating MI counsellors in smoking cessation, designed specifically for the evoking MI task. Evoke-Sim employs structured client profiles, an evoking-specific three-stage conversation flow, and a reveal policy that regulates which client profile information might be disclosed at each stage. We show that compared to existing profile-grounded simulated clients, Evoke-Sim is better at differentiating levels of MI quality using task-aware evaluation metrics, while reducing non-grounded client statements and premature disclosure of client information, setting a higher standard for the evaluation of LLM-based MI counsellors.

Autonomous AI agents in social media present concrete risks to democratic discourse and platform governance, while also offering tools for pre-deployment recommender system testing. A central open question is whether persona-prompted LLMs can simulate individual-level social media reactions with sufficient accuracy to support either application, and how accuracy depends on profile completeness, model selection, and the generalization challenge posed by novel post content. This study benchmarks twelve LLM configurations on binary like/dislike prediction across 296 survey-based agent profiles and 26 ground-truth-mapped posts under three profile conditions, with leave-post-out machine learning classifiers as baselines. Across full-profile conditions, accuracy ranges from 75.54% to 96.68%, with a 30-point spread attributable primarily to model selection and confirmed by paired McNemar tests with agent-level bootstrap intervals. GPT-5.5 Pro accuracy degrades monotonically from 96.68% under a full profile to 62.32% under a reduced profile and to 51.00% with demographics alone, the last indistinguishable from the majority-class baseline, which confirms that demographic inference provides negligible predictive signal. Supervised classifiers collapse to 15.4% under leave-post-out, while LLMs sustain genuine zero-shot generalization unavailable to trained methods. Adaptive reasoning improves accuracy substantially for some models. Inter-model agreement is nearly double for posts with direct profile anchors (mean ąppa} = 0.44) than for posts without them (ąppa} = 0.23), and the least heterogeneous configuration homogenizes 34% of simulated population reactions. Results validate LLM-based simulation for recommender system stress-testing while documenting the behavioral accuracy that makes large-scale synthetic agent swarms a credible threat to public opinion.

Current text-to-speech systems face a trade-off: autoregres- sive codec language models produce highly intelligible speech but require large-scale models and training data and decode tokens sequentially, while non-autoregressive approaches im- prove speed at the cost of linguistic accuracy. We present DLLM-TTS, a framework that formulates TTS as conditional block discrete diffusion over X-Codec2 neural audio codec to- kens. The model decomposes sequences into blocks and applies masked diffusion within each block while processing blocks se- quentially, learning both local acoustic coherence and global text-speech alignment. During inference, parallel token pre- diction within blocks enables efficient generation with a real- time factor (RTF) of 0.15. A 0.6B-parameter model trained on 20K hours achieves competitive performance on the Seed- TTS-eval benchmark, demonstrating that block discrete diffu- sion language models enable practical and data-efficient speech synthesis with parallel generation.

Early-life monitoring in laying hens remains constrained by fragmented single-modality sensing and the absence of formal system-level state representations. HenTwin, a multimodal digital twin framework implemented as a five-layer IoT architecture, formalizes flock-level multimodal biological state dynamics from hatch through 25 weeks of age. A four-dimensional biological state vector integrating body surface temperature, acoustic energy entropy, band energy ratio, and optical-flow-based motion is defined, with the temperature-humidity index treated as an exogenous environmental input to preserve intervention capability. A discrete-time state transition model is estimated from 25 weeks of longitudinal multimodal data collected from 150 Lohmann LSL-Lite hens across five controlled rooms at the Atlantic Poultry Research Centre, Dalhousie University. The estimated transition matrix exhibits modality-specific persistence while remaining asymptotically stable. Perturbation analysis demonstrates that a sustained +2.0 THI increase produces a stable long-run acoustic entropy elevation of 0.54 nats, approximately one-quarter of the entire 1.87-nat developmental decline observed across the study period. Pettitt change-point detection identifies coordinated multimodal developmental state transitions at Weeks 12-14. Cross-room validation suggests that structural transition parameters are partially transferable across rooms, whereas environmental input sensitivity requires room-specific calibration, supporting a two-tier IoT deployment architecture. Leave-one-out cross-validation demonstrates consistent out-of-sample model performance. HenTwin takes a first step toward formal, state-aware digital twin inference in precision livestock farming.

RAG systems rely on chunking, which destroys structural information in documents. Existing heading-based retrieval (Jeong et al., 2025) requires multiple LLM calls per document and returns sub-chunks within matched sections. We introduce ToC-guided page retrieval, which infers headings from visual formatting without LLM calls, embeds them as a parallel index, and loads full page sections. Across 1,280 conditions on 8 enterprise documents (5 to 195 pages), we find: (1) ToC is a significant main effect on answer quality (d = +0.41, p = 0.031), with the largest gains in completeness (+0.40) and usefulness (+0.40); (2) combined with answer-side verification, it outperforms query-side decomposition + verification (d = +0.32, p = 0.036); (3) ToC contributes 20% of citations despite adding only 2.9 pages per query; (4) gains are directionally larger on longer documents (up to +1.50 on 118 pages), though the trend does not reach significance with 8 documents; and (5) a 480-condition sensitivity analysis finds no significant parameter effects (all p > 0.38), confirming defaults are near-optimal. The contribution is both methodological (a new zero-LLM-cost retrieval algorithm) and empirical: factorial evidence that document-side, query-side, and answer-side enhancements are complementary, a three-way interaction not previously studied.

Fixed benchmarks are costly to renew and cannot adapt their questions to model-specific failures. We ask whether LLMs can instead discover one another's weaknesses and turn those observations into an evaluation process. To study this question, we introduce LivingArena, an automated peer-probing framework in which models take turns testing one another. Using the interaction history, each questioner identifies potential weaknesses of its opponent and constructs targeted, verifiable questions to probe them. A 3,600-round tournament of ten models reveals a clear role asymmetry: strong answerers are not always reliable questioners, because they may generate internally inconsistent tests or fail to verify their own reference answers. After a questioner exposes an answerer's failure, it is more likely to pursue the same capability domain, while the answerer's weakness recurs on independently generated questions, including questions written by different models. These findings show that peer probing can reveal persistent model-specific weaknesses while separately evaluating answering and reliable test construction. By automating this process and allowing test difficulty to evolve with model capabilities, LivingArena provides a "living" benchmark for model development, red-teaming, and capability-aware multi-agent coordination. We publicly release our code: https://github.com/galaxyChen/LivingArena

The rapid rise of vlogs as a personalized storytelling medium has created a demand for automated systems to evaluate and refine vlog editing plans. However, vlog assessment is highly subjective and remains challenging due to a lack of standardized criteria, dataset and benchmark, and effective reward models. To address these challenges, we define a comprehensive vlog evaluation framework guided by professional vlog creators and product managers, establishing a taxonomy of six key dimensions, i.e., Creativity, Consistency, Concept Design, Cinematography, Narration, and Pacing. Subsequently, we curate a large-scale dataset of 100k vlog edits and a dedicated benchmark, VRMBench, to evaluate the vlog rewarding capabilities of Multimodal Large Language Models (MLLMs). Finally, we present VlogReward, a robust vlog reward model that can provide both fine-grained multi-dimensional scores and actionable feedback for iterative refinement. Technically, we enhance the Group Relative Policy Optimization (GRPO) framework by introducing an adjustable inter-group comparison reward, which mitigates the "direction blindness" issue of standard GRPO and enables the model to better distinguish varied-quality edits. VlogReward achieves state-of-the-art results that significantly outperform existing MLLMs, including GPT-5 and Gemini-3-Pro. We hope that our study can help vlog creators and foster automated vlog evaluation and refinement systems.

An international agreement to limit AI development could be crucial to mitigate risks from AI. However, it remains unclear which conditions should determine when the limiting measures are relaxed. We survey existing international agreements, outline what properties appropriate conditions should satisfy, list possible conditions, and finally give a recommendation in an example scenario. We recommend a fixed time period after which a new organization established at the start of the period specifies conditions that address when AI development can be safely conducted, with a possibility of withdrawal in extraordinary circumstances. We hope to illustrate the considerations that would likely go into an international agreement to limit AI.

Representative clutter height (RCH) is a key parameter in radio propagation and interference analysis because it captures the dominant height of local obstructions that drive terminal clutter loss. Current practice often relies on fixed clutter heights assigned to land use classes in Recommendation ITU-R P.452-18, but this misses within class variation and can lead to conservative exclusion zones and poor site ranking for low Earth orbit ground station siting and spectrum coordination. We present an interpretable, globally deployable machine learning framework for predicting RCH from open geospatial data. The model is trained using LiDAR derived labels from the U.S. Geological Survey 3D Elevation Program and inference time features from global land-cover, terrain, demographic, thermal, and optical remote sensing products. We define RCH using a robust 75th percentile clutter height statistic, evaluate multiple regressors, and select LightGBM for its accuracy, efficiency, and compatibility with feature attribution analysis. The final model achieves a mean absolute error of 1.79m and an R^2=0.765, reducing absolute error by more than 60% relative to the ITU baseline. Beyond aggregate fit, we evaluate domain facing criteria relevant to RF planning, including meter scale error, tolerance band accuracy, over and under estimation tails, agreement with ITU clutter height regimes, and SHAP-based physical plausibility. SHAP identifies tree canopy cover, land-cover semantics, and spectral reflectance as the most influential predictors. Studies on segmentation derived features, non-forest ablations, and land-cover matched international validation show that open geospatial data can improve clutter modeling at scale without sacrificing interpretability or deployability.

Type 1 Diabetes (T1D) is a chronic, life-threatening autoimmune condition characterized by the complete destruction of insulin-producing pancreatic beta cells. While Artificial Pancreas Systems (APS) powered by Reinforcement Learning (RL) have shown promise in automating insulin delivery, their “black-box” nature makes it hard for patients and doctors to trust them fully. This paper presents LLM-T1D, a promising approach that combines the precision of RL with the clear, human-like reasoning of Large Language Models (LLMs) to create a more transparent and reliable insulin pump controller. By training an expert RL system and distilling its knowledge into fine-tuned LLaMA 3.1 8B and Qwen3 8B models, we developed a controller that not only surpasses the RL system's performance but also explains its decisions in plain, understandable language. Tested on the FDA-approved UVA/Padova T1D simulator, the LLM controllers deliver excellent blood sugar control (73.5% Time in Range) while maintaining strict formal safety verification against hallucinations.

Reinforcement learning (RL) has become the dominant paradigm for improving the reasoning capabilities of large language models, but it requires expensive training, curated data, and reward signals. Recent work shows that sampling from sharpened base-model distributions at test time recovers much of the RL gain, yet existing methods rely solely on output-layer likelihoods and ignore the transformer's internal forward-pass dynamics. We introduce Depth-Entropy Guided Sampling (DEGS), a training-free, test-time method that exploits layer-wise entropy collapse as an intrinsic quality signal. We observe that stronger reasoners – including RL-posttrained variants – exhibit a distinctive "late collapse": logit-lens decoded entropy stays elevated until deeper layers before converging. We define a per-sequence collapse depth $D(\mathbf{x})$ and a joint objective $\pi(\mathbf{x}) \propto p(\mathbf{x})^\alpha \exp(\beta D(\mathbf{x}))$ that combines sequence likelihood with this depth-entropy structure, instantiated inside an MCMC power-sampling framework (DEGS-MCMC). Across three open-weight models and four reasoning benchmarks, this near-chance per-candidate signal compounds over the sampling trajectory into state-of-the-art training-free accuracy, with gains largest out of domain and on the harder splits – exactly where likelihood alone falls short – at single-digit-percent wall-clock overhead. DEGS narrowly trails an in-house GRPO reference on the math splits GRPO was trained for, yet surpasses it out of domain on GPQA for all three models, without any training, reward model, or labeled data.

A modern coding agent can hold an entire repository in its context window. Most of its reading is wasted – and the interesting question is not how much context an agent can use, but what it actually needs. We study that question at the moment it matters most: when the agent must edit code. Separating finding the work site from acting on it, we hold localization fixed with an oracle, vary only how the code is represented, and score context against real issue resolution on SWE-bench Verified. The answer is starkly minimal. The signal lives in the code being edited itself: natural-language summaries of it answer almost none of the behavioral questions that the source answers ($4/45$ vs.\ $27/45$, held-out repositories, independent judge), and the gap belongs to the representation, not the summarizer – a frontier model's summaries score exactly as poorly as a 3B model's. The surrounding context hardly matters either: across every multi-file instance in Verified, under a protocol frozen before any data, rendering a file's remainder as UML skeletons and signatures resolves no more issues than deleting that remainder outright ($N{=}70$, exact McNemar $p{=}0.75$). That was our registered hypothesis, and it failed. Compressed context, meanwhile, matches whole files at a third of the tokens: a resolved issue costs $19$K context tokens, not $94$K. The instrument also yielded a finding the field should keep: temperature-0 API inference flips ${\sim}9\%$ of per-instance outcomes between byte-identical runs. That is a noise floor under every small effect reported on this benchmark, including ours. We release the instrument – gold-validated environments, per-instance proof that every reference edit is expressible from every arm's context, deterministic patch construction, and pre-registered hypotheses whose nulls we publish.

Dense sentence embeddings are fundamental to modern Retrieval-Augmented Generation (RAG) systems but suffer from a lack of interpretability due to feature superposition. This opacity hinders the alignment of retrieval processes with human intent, as the entangled representations are difficult to analyze or control. In this work, we propose a method to disentangle the dense representations of sentence transformers (e.g., E5) into human-interpretable concepts using Top-k Sparse Autoencoders (SAEs). We demonstrate that these disentangled features align with specific semantic, syntactic, and pragmatic categories. Furthermore, we introduce an activation steering mechanism that allows for precise intervention in the retrieval process. By clamping specific latent features, we show that it is possible to re-rank search results to better align with user constraints without retraining the backbone model. Our findings suggest that SAE-based decomposition offers a viable path toward transparent and steerable neural information retrieval.

We introduce LEDGER to tackle the novel context engineering challenge of agentic document editing, where localized edits to long, structured documents must be applied efficiently without breaking cross-references or semantic consistency. LEDGER constructs a lightweight dependency graph that explicitly models document structure, including hierarchical organization, explicit references, implicit dependencies, and semantic relationships. For each edit, graph-guided retrieval selects only the necessary context, avoiding full-document processing while preserving consistency. We evaluate LEDGER on a curated benchmark of 1.9k test cases with various document types and lengths, spanning six state-of-the-art models: LEDGER improves consistency from 56% to 76% across all six models and test scenarios while reducing token usage. Notably, LEDGER with low reasoning effort matches baseline performance at high reasoning effort using fewer tokens, showing that explicit dependency representations can partially substitute for expensive internal reasoning in agentic document editing.

HAR using Inertial Measurement Unit (IMU) sensors is vital for healthcare monitoring and rehabilitation. Despite deep learning advancements, major challenges remain: reliance on labeled data, multi-sensor fusion complexity, and the limited ability of unsupervised methods to capture spatiotemporal dependencies. These issues are pronounced in real-world scenarios with noisy data, overlapping activities, and missing labels. We propose a fully unsupervised spatiotemporal feature fusion framework using a memory-augmented autoencoder. It enhances activity representations via short temporal windows of multi-sensor IMU data, enabling real-time applications. Our framework extracts hierarchical static features via a Stacked Autoencoder, fusing them within and across sensors. A sequence-to-sequence LSTM Autoencoder then temporally refines these features, incorporating historical motion patterns without labels. We analyze key hyperparameters to identify configurations that maximize feature separability under short-window constraints. Evaluated on DaLiAc and PAMAP2 using realistic inter-class window segmentation, our method achieves 96.6% and 98.4% accuracy, respectively, surpassing supervised baselines and unsupervised approaches. Our method improves feature separability by up to 9% despite shorter temporal windows. While our realistic inter-class segmentation reduces accuracy by ~7%, it was intentionally adopted to better reflect real-world activity transitions and practical relevance.

Long-horizon large language model (LLM) agents accumulate interaction trajectories that quickly exceed any practical prompt budget, and existing memory methods either truncate aggressively and lose non-local evidence or retain boilerplate that degrades decision quality. We ask a mechanism question rather than claiming a better general-purpose memory system: when does organizing trajectory memory into overlapping semantic units (OSUs) – groups of related steps in which one step may belong to several units – help retrieval over flat or disjoint alternatives? We instantiate this in OSU-Mem, which retrieves from an overlapping OSU pool via budgeted coarse-to-fine expansion, and show its benefit is conditional: overlapping memory helps when the evidence steps a query needs share tool calls or entities, but hurts when those steps are fully heterogeneous and share neither. On a synthetic benchmark where evidence carries such shared structure by construction, OSU-Mem improves over the strongest baseline as the theory predicts; yet on a concatenated, constructed unaugmented $\tau$-bench setting its aggregate advantage over flat retrieval vanishes. Splitting queries by whether their evidence shares tools and entities shows this near-tie to be an artifact of mixing query types rather than a property of either method, and ToolBench, a controlled probe built to carry shared structure by design, corroborates the same mechanism via an overlap-vs.-disjoint construction contrast (under a coverage-guided variant), isolating the construction principle rather than validating the full default system. Because the relevant sharing is cheaply estimable from metadata, the analysis yields a metadata-based heuristic for predicting when overlap is likely to improve retrieval. We deliberately isolate the retrieval layer, assessed by retrieval quality and an LLM-mediated evidence-selection stage.

This paper develops a multiscale model of coalition formation in which strategic exit-and-join decisions are coupled with tactical consensus dynamics inside coalitions. Coalition value is generated endogenously from within-coalition information aggregation, while Aumann-Dreze payoffs, switching frictions, and acceptance rules govern strategic reconfiguration. The framework introduces a fast-slow architecture in which transferable coalition value emerges from DeGroot-style consensus processes, while coalition structures evolve through incentive-driven exit-and-join dynamics. The analysis characterizes joint tactical-strategic equilibria, conditions for tactical and strategic unanimity, segregation, polarization, and cognitive barriers that sustain stable coalition structures. A fixed-point characterization and existence results are established for the coupled dynamics. Numerical experiments reveal an instability-consensus paradox: low or negative switching barriers may prevent strategic convergence while simultaneously promoting temporal mixing sufficient to achieve global tactical consensus. The results provide a unified perspective on coalition formation, consensus dynamics, information aggregation, and strategic stability in multi-agent systems.

End-to-end automatic speech recognition systems frequently hallucinate rare entities and domain-specific terms, especially in low-resource languages. While retrieval-augmented generation frameworks can mitigate these errors using large language models, current architectures face significant challenges. They either rely on standard sparse retrieval that ignores phonetic misrecognitions or utilize heavyweight cross-modal embeddings that introduce high latency. This letter proposes a highly efficient, purely lexical error-aware framework designed to explicitly resolve phonetic and loop hallucinations. Our approach integrates a symmetric text normalization module with a novel error-aware term frequency-inverse document frequency algorithm. By constructing a sparse diagonal penalty matrix based on historical errors, the retriever mathematically prioritizes corrective documents containing specific high-risk misrecognitions. Evaluated on the Persian subset of the FLEURS dataset, our method increased the error-aware hit rate from 53.7% to 90.9%. In end-to-end evaluations, the integrated framework reduced the final word error rate from 23.06% to 18.83%, achieving significant accuracy gains with near-zero inference latency.

Automatic Music Transcription (AMT) models have achieved a high level of success in polyphonic transcription of various instruments. Velocity, typically a measure of note intensity, is less commonly predicted in these models due to the absence of velocity labels in available datasets and lack of a proper definition for instruments other than piano. We present a methodology and model for velocity prediction in Automatic Guitar Transcription (AGT) which uses virtual instruments to generate synthetic training data with velocity labels. We first pretrain a model on this synthetic data. These weights are then transferred to a different model and trained on real guitar audio, allowing the model to retain the working velocity prediction while also achieving high performance and generalisability from the real training data. The velocity prediction is shown to outperform a baseline model which does not use the pretrained velocity weights, when evaluated on synthetic data. In addition, using the pretrained velocity weights offers a small improvement in note transcription, though the magnitude of this improvement is limited and not always significant depending on the testing data. Overall the model achieves results comparable to the state of the art in guitar transcription, while also successfully predicting velocity.