arXiv:2603.03480v2 Announce Type: replace
Abstract: We study reinforcement learning with delayed state observation, where the agent observes the current state after some random number of time steps. We propose an algorithm that combines the augmentation method and the upper confidence bound approach. For tabular Markov decision processes (MDPs), we derive a regret bound of $\tilde{\mathcal{O}}(H \sqrt{D_{\max} SAK})$, where $S$ and $A$ are the cardinalities of the state and action spaces, $H$ is the time horizon, $K$ is the number of episodes, and $D_{\max}$ is the maximum length of the delay. We also provide a matching lower bound up to logarithmic factors, showing the optimality of our approach. Our analytical framework formulates this problem as a special case of a broader class of MDPs, where their transition dynamics decompose into a known component and an unknown but structured component. We establish general results for this abstract setting, which may be of independent interest.
Science Journals
arXiv:2603.24761v2 Announce Type: replace
Abstract: Spreading and storing erasure-coded data effectively in distributed systems is challenging in practical settings. The dissemination of erasure-coded information is typically designed to complete only after receiving messages from $(N-F)$ nodes, thereby preparing for the worst-case, but rare, scenario of $F$ failures. In steady state, the remaining $F$ nodes may in fact be healthy, but their resources are not counted. This leads to over-provisioning of storage for encoded data.
This paper introduces eAID, a novel elastic information dispersal algorithm that addresses this conundrum through a two-stage approach.
First, the core protocol estimates the actual number $f$ of faulty nodes, rather than assuming the worst-case bound $F$. Dissemination completes quickly when messages are received from $(N-f)$ nodes, and more gradually when fewer nodes respond. Second, after initial dissemination completes, eAID continues monitoring for additional responses. As responses arrive from up to $N$ nodes, the system prunes the information stored at responding nodes accordingly.
A key technique enabling this seamless elasticity is an agile encoding scheme that varies the number of disseminated fragments while keeping both fragment size and the recovery threshold $(F+1)$ fixed. Not only does this enable varying the number of disseminated fragments on the fly, it also allows nodes to discard encoded fragments autonomously. Crucially, this is achieved without maintaining complex metadata, without requiring nodes to reconstruct or re-encode information, and without global coordination for storage decisions.
We demonstrate the practicality of eAID by integrating it with a replicated key-value store, and evaluating it in network environments with unpredictable latencies. The results show that eAID improves overall performance while significantly reducing long-term storage consumption.
arXiv:2602.07075v5 Announce Type: replace
Abstract: Current chemical large language models (LLMs) predominantly rely on explicit Chain-of-Thought (CoT) to solve complex reasoning problems. However, forcing nonverbal tacit chemical logic into discrete natural language imposes a fundamental ``modality mismatch,'' creating an artificial bottleneck for reasoning. We introduce LatentChem, a reasoning interface that decouples chemical logic from linguistic generation, enabling the model to process information via continuous thought vectors and dynamic perception. Our investigation reveals a pivotal emergent behavior: spontaneous internalization, defined here as self-selected under outcome-only optimization. When optimized for task success, the model abandons verbose textual derivations in favor of implicit latent computation, suggesting that it identifies the continuous manifold as a more native substrate for chemical logic. This paradigm shift also proves to be a superior computational strategy: LatentChem achieves a 59.88\% non-tie win rate against the strong CoT baseline on the rigorous ChemCoTBench, while delivering a broad 10.84$\times$ average reduction in reasoning step overhead (5.96$\times$ wall-clock speedup) across all evaluated benchmarks. Our results provide empirical evidence that chemical reasoning is more naturally and effectively realized as continuous latent dynamics rather than discretized linguistic trajectories.
arXiv:2604.12176v2 Announce Type: replace
Abstract: Relational reasoning is the ability to infer relations that jointly bind multiple entities, attributes, or variables. This ability is central to scientific reasoning, but existing evaluations of relational reasoning in large language models often focus on structured inputs such as tables, graphs, or synthetic tasks, and do not isolate the difficulty introduced by higher-arity relational binding. We study this problem through the lens of Relational Complexity (RC), which we define as the minimum number of independent entities or operands that must be simultaneously bound to apply a relation. RC provides a principled way to vary reasoning difficulty while controlling for confounders such as input size, vocabulary, and representational choices. Building on RC, we introduce REL, a generative benchmark framework spanning algebra, chemistry, and biology that varies RC within each domain. Across frontier LLMs, performance degrades consistently and monotonically as RC increases, even when the total number of entities is held fixed. This failure mode persists with increased test-time compute and in-context learning, suggesting a limitation tied to the arity of the required relational binding rather than to insufficient inference steps or lack of exposure to examples. Our results identify a regime of higher-arity reasoning in which current models struggle, and motivate re-examining benchmarks through the lens of relational complexity.
arXiv:2602.07639v2 Announce Type: replace
Abstract: With the emergence of large language models (LLMs) as a powerful class of generative artificial intelligence (AI), their use in tutoring has become increasingly prominent. Prior works on LLM-based tutoring typically learn a single tutor policy and do not capture the diversity of tutoring styles. In real-world tutor-student interactions, pedagogical intent is realized through adaptive instructional strategies, with tutors varying the level of scaffolding, instructional directiveness, feedback, and affective support in response to learners' needs. These differences can all impact dialogue dynamics and student engagement. In this paper, we explore how tutor personas embedded in human tutor-student dialogues can be used to guide LLM behavior without relying on explicitly prompted instructions. We train a steering vector using preference optimization: an activation-space direction that guides model responses toward specific tutor personas. We find that this steering vector captures tutor-specific variation across dialogue contexts, improving semantic alignment with ground-truth tutor utterances and increasing preference-based evaluations, while largely preserving lexical similarity. Analysis of the learned scaling coefficients further reveals interpretable structure across tutors, corresponding to consistent differences in tutoring behavior. These results demonstrate that activation steering offers an effective and interpretable way for controlling tutor-specific variation in LLMs using signals derived directly from human dialogue data.
arXiv:2603.20508v2 Announce Type: replace
Abstract: Reasoning language models (RLMs) and the intermediate chains of thought they emit play an increasingly central role in multi-agent setups such as inter-model monitoring or distillation into smaller models. When agents at different capability tiers must cooperate, strong models need to produce traces digestible by weaker ones. We refer to this goal as "weak-to-strong legibility". Trustworthiness of large models depends in part on this legibility property. For safety oversight in particular, adoption of weak monitors may become a standard for reliability scaffolds on a healthy budget. Legibility requires that the shape of these decision-making traces takes some form accessible to weaker monitors. Existing efficiency-based metrics for legibility fail to capture "thoroughness", instead focusing on conciseness.
arXiv:2602.09708v2 Announce Type: replace
Abstract: We propose physics-informed spectral diffusion (PISD), a methodology that combines generative latent diffusion models with physics-informed machine learning to generate solutions of partial differential equations (PDEs) conditioned on partial observations, which includes, in particular, forward and inverse PDE problems. We learn the joint distribution of PDE parameters and solutions via a diffusion process in a latent space of scaled spectral representations, where Gaussian noise corresponds to functions with controlled regularity. This spectral formulation enables significant dimensionality reduction compared to grid-based diffusion models and ensures that the induced process in function space remains within a class of functions for which the PDE operators are well defined. Building on diffusion posterior sampling, we enforce physics-informed constraints and measurement conditions during inference, applying Adam-based updates at each diffusion step. We evaluate the proposed approach on Poisson, Helmholtz, and incompressible Navier-Stokes equations, demonstrating improved accuracy and computational efficiency compared with existing diffusion-based PDE solvers, which are state of the art for sparse observations. Code is available at https://github.com/deeplearningmethods/PISD.
arXiv:2602.10352v2 Announce Type: replace
Abstract: Self-interpretation methods prompt language models to describe their own internal states, but remain unreliable due to hyperparameter sensitivity. We show that training lightweight adapters on interpretability artifacts, while keeping the LM entirely frozen, yields reliable self-interpretation across tasks and model families. A scalar affine adapter with just $d_\text{model}+1$ parameters suffices: trained adapters generate sparse autoencoder feature labels that outperform the training labels themselves (70% vs 50% generation scoring at 70B scale), identify topics with 94% recall@1 versus 1% for untrained baselines, and decode bridge entities in multi-hop reasoning that appear in neither prompt nor response, surfacing implicit reasoning without chain-of-thought. The learned bias vector alone accounts for 85% of improvement, and simpler adapters generalize better than more expressive alternatives. Controlling for model knowledge via prompted descriptions, we find self-interpretation gains outpace capability gains from 7B to 72B parameters. Our results demonstrate that self-interpretation improves with scale, without modifying the model being interpreted.
arXiv:2602.10387v2 Announce Type: replace
Abstract: Traditional query optimization relies on cost-based optimizers that estimate execution cost (e.g., runtime, memory, and I/O) using predefined heuristics and statistical models. Improving these requires substantial engineering effort, yet they often cannot exploit semantic correlations in queries and schemas that could enable better physical plans. Large language models (LLMs), however, can reason about column semantics, value distributions, and broader domain context that classical statistics miss. We introduce DBPlanBench, a harness for the DataFusion engine that exposes physical plans through a compact serialized representation and applies LLM-proposed edits as JSON patches. On this harness, we instantiate a test-time optimization workflow where an LLM examines physical query plans, proposes localized edits based on semantic reasoning, and an evolutionary search refines the candidates across iterations. We target OLAP queries, where heavy, repeated execution turns even small efficiency gains into substantial cumulative savings. We specifically focus our evaluation on join reordering and join-side selection, where cardinality-estimation errors compound multiplicatively. Median speedups reach $1.10$-$1.12\times$ on TPC-H and $1.05$-$1.07\times$ on TPC-DS, with some achieving up to $4.78\times$. We also demonstrate that optimizations discovered at small scale factors transfer effectively to larger ones, supporting a low-cost small-to-large workflow.
arXiv:2606.00567v2 Announce Type: replace
Abstract: Recent diffusion accelerators exploit activation sparsity by skipping near-zero GELU outputs, reporting 52--85% element-level sparsity. However, systolic-array hardware processes activations at column granularity, where a single non-zero element forces the entire column to be computed. We present the first systematic column-level sparsity characterization across seven diffusion workloads spanning three workload groups and four modalities. Our measurements reveal that element-level sparsity overstates hardware-exploitable sparsity by up to 78 percentage points and exposes a three-way taxonomy. UNet+transformer workloads exhibit activation concentration with workload-dependent cycle reductions up to 30.6%. Pure-transformer DiT shows dispersion, yielding 12.4%. Motion/dance transformer workloads range from modest reductions to 50.8% for MLD, driven by its extreme token dimension and expansion ratio. Cycle-level simulation on a GDDR6-based accelerator confirms that memory stalls account for up to 84--89% of total cycles and that layout sensitivity tracks the profiling-based taxonomy. A full accuracy sweep across five thresholds reveals that UNet+transformer workloads degrade gracefully, while motion models exhibit an accuracy cliff between the primary operating point and the next threshold. Our characterization shows that workload group and model dimensions jointly determine whether column-level memory layout optimization is beneficial, and element-level sparsity alone is insufficient for that prediction.
arXiv:2602.12221v2 Announce Type: replace
Abstract: We propose UniDFlow, a unified discrete flow-matching framework for multimodal understanding, generation, and editing. It decouples understanding and generation via task-specific low-rank adapters, avoiding objective interference and representation entanglement, while a novel reference-based multimodal preference alignment optimizes relative outcomes under identical conditioning, improving faithfulness and controllability without large-scale retraining. UniDFlpw achieves SOTA performance across eight benchmarks and exhibits strong zero-shot generalization to tasks including inpainting, in-context image generation, reference-based editing, and compositional generation, despite no explicit task-specific training.
arXiv:2602.14279v2 Announce Type: replace
Abstract: Eliciting information to reduce uncertainty about latent group-level properties from surveys and other collective assessments requires allocating limited questioning effort under real costs and missing data. Although large language models enable adaptive, multi-turn interactions in natural language, most existing elicitation methods optimize what to ask with a fixed respondent pool, and do not adapt respondent selection or leverage population structure when responses are partial or incomplete. To address this gap, we study adaptive group elicitation, a multi-round setting where an agent adaptively selects both questions and respondents under explicit query and participation budgets. We propose a theoretically grounded framework that combines (i) an LLM-based expected information gain objective for scoring candidate questions with (ii) heterogeneous graph neural network propagation that aggregates observed responses and participant attributes to impute missing responses and guide per-round respondent selection. This closed-loop procedure queries a small, informative subset of individuals while inferring population-level responses via structured similarity. Across three real-world opinion datasets, our method consistently improves population-level response prediction under constrained budgets, including a >12% relative gain on CES at a 10% respondent budget.
arXiv:2602.16666v3 Announce Type: replace
Abstract: AI agents are increasingly deployed to execute important tasks. While rising accuracy scores on standard benchmarks suggest rapid progress, many agents still continue to fail in practice. This discrepancy highlights a fundamental limitation of current evaluations: compressing agent behavior into a single success metric obscures critical operational flaws. Notably, it ignores whether agents behave consistently across runs, withstand perturbations, fail predictably, or have bounded error severity. Grounded in safety-critical engineering, we provide a holistic performance profile by proposing twelve concrete metrics that decompose agent reliability along four key dimensions: consistency, robustness, predictability, and safety. Evaluating 15 models across two complementary benchmarks, we find that recent capability gains have only yielded small improvements in reliability. By exposing these persistent limitations, our metrics complement traditional evaluations while offering tools for reasoning about how agents perform, degrade, and fail.
arXiv:2602.17149v2 Announce Type: replace
Abstract: Recent time series modeling faces a sharp divide between numerical generation and semantic understanding, with research showing that generation models often rely on superficial pattern matching, while understanding-oriented models struggle with high-fidelity numerical output. Although unified multimodal models (UMMs) have bridged this gap in vision, their potential for time series remains untapped. We propose TimeOmni-VL, the first vision-centric framework that unifies time series understanding and generation through two key innovations: (1) Fidelity-preserving bidirectional mapping between time series and images (Bi-TSI), which advances Time Series-to-Image (TS2I) and Image-to-Time Series (I2TS) conversions to ensure near-lossless transformations. (2) Understanding-guided generation. We introduce TSUMM-Suite, a novel dataset consisting of six understanding tasks rooted in time series analytics and coupled with two generation tasks. With a calibrated Chain-of-Thought, TimeOmni-VL is the first to leverage time series understanding as an explicit control signal for high-fidelity generation. Experiments confirm that this unified approach significantly improves semantic understanding and numerical precision, establishing a new frontier for multimodal time series modeling.
arXiv:2602.22480v4 Announce Type: replace
Abstract: An important emerging application of coding agents is agent harness optimization: the iterative improvement of a target agent by editing and evaluating its code. Despite its relevance, the community lacks a systematic understanding of coding agent performance on this task. Harness optimization differs from conventional software engineering: agent harnesses interleave deterministic code with stochastic LLM completions, requiring structured capture of both intermediate execution traces and downstream outcomes. To address these challenges, we introduce (1) VeRO (Versioning, Rewards, and Observations), an outer harness that provides versioned snapshots, budget-controlled evaluation, and structured execution traces of target harnesses, and (2) VeRO-Bench, a benchmark suite of target agents and tasks with reference evaluation procedures. Using VeRO, we conduct an empirical study comparing optimizers across tasks and analyzing which modifications reliably improve target agent harnesses. We release VeRO to support research on agent optimization as a core capability for coding agents. Code is available at https://github.com/scaleapi/vero.
arXiv:2603.00029v3 Announce Type: replace
Abstract: Large Language Models (LLMs) exhibit highly anisotropic internal representations, often characterized by massive activations, a phenomenon where a small subset of feature dimensions possesses magnitudes significantly larger than the rest. While prior works view these extreme dimensions primarily as artifacts to be managed, we propose a distinct perspective: these dimensions serve as intrinsic interpretable functional units arising from domain specialization. Specifically, we propose a simple magnitude-based criterion to identify Domain-Critical Dimensions in a training-free manner. Our analyses reveal that such dimensions behave as interpretable semantic detectors for symbolic/quantitative patterns or domain-specific terms. In addition, we introduce Critical Dimension Steering, which applies activation steering exclusively to the identified dimensions. Empirical results show that this approach outperforms conventional whole-dimension steering in domain adaptation and jailbreaking scenarios.
arXiv:2603.01372v2 Announce Type: replace
Abstract: Concept Bottleneck Models (CBMs) enhance the interpretability of end-to-end neural networks by introducing a layer of concepts and predicting the class label from the concept predictions. A key property of CBMs is that they support interventions, i.e., domain experts can correct mispredicted concept values at test time to improve the final accuracy. However, typical CBMs apply interventions by overwriting only the corrected concept while leaving other concept predictions unchanged, which ignores causal dependencies among concepts. To address this, we propose the Causal Neural Probabilistic Circuit (CNPC), which combines a neural attribute predictor with a causal probabilistic circuit compiled from a causal graph. This circuit supports exact, tractable causal inference that inherently respects causal dependencies. Under interventions, CNPC models the class distribution based on a Product of Experts (PoE) that fuses the attribute predictor's predictive distribution with the interventional marginals computed by the circuit. We theoretically characterize the compositional interventional error of CNPC w.r.t. its modules and identify conditions under which CNPC closely matches the ground-truth interventional class distribution. Experiments on five benchmark datasets in both in-distribution and out-of-distribution settings show that, compared with five baseline models, CNPC achieves higher task accuracy across different numbers of intervened attributes.
arXiv:2603.01915v2 Announce Type: replace
Abstract: We present a novel, practical approach to speed up sparse matrix-vector multiplication (SpMVM) on GPUs. The novel key idea is to apply lossless entropy coding to further compress the sparse matrix when stored in one of the commonly supported formats. Our method is based on dtANS, our new lossless compression method that improves the entropy coding technique of asymmetric numeral systems (ANS) specifically for fast parallel GPU decoding when used in tandem with SpMVM. We apply dtANS on the widely used CSR format and present extensive benchmarks on the SuiteSparse collection of matrices against the state-of-the-art cuSPARSE library. On matrices with at least 2^(15) entries and at least 10 entries per row on average, our compression reduces the matrix size over the smallest cuSPARSE format (CSR, COO and SELL) in almost all cases and up to 11.77 times. Further, we achieve an SpMVM speedup for the majority of matrices with at least 2^(25) nonzero entries. The best speedup is 3.48x. We also show that we can improve over the AI-based multi-format AlphaSparse in an experiment that is limited due to its extreme computation overhead. We provide our code as an open source C++/CUDA header library, which includes both compression and multiplication kernels.
arXiv:2603.03612v3 Announce Type: replace
Abstract: The community is increasingly exploring linear RNNs (LRNNs) as language models, motivated by their expressive power and parallelizability. While prior work establishes the expressivity benefits of LRNNs over transformers, it is unclear what makes LRNNs -- but not traditional, nonlinear RNNs -- as easy to parallelize in practice as transformers. We answer this question by providing a tight connection between types of RNNs and standard complexity classes. We show that LRNNs can be viewed as log-depth (bounded fan-in) arithmetic circuits, which represents only a slight depth overhead relative to log-depth boolean circuits that transformers admit. Furthermore, we show that nonlinear RNNs can solve $\mathsf{L}$-complete problems (and even $\mathsf{P}$-complete ones, under polynomial precision), revealing a fundamental barrier to parallelizing them as efficiently as transformers. Our theory also identifies fine-grained expressivity differences between recent popular LRNN variants: permutation-diagonal LRNNs are $\mathsf{NC}^1$-complete whereas diagonal-plus-low-rank LRNNs are more expressive ($\mathsf{PNC}^1$-complete). We provide further insight by associating each type of RNN with a corresponding automata-theoretic model that it can simulate. Together, our results reveal fundamental tradeoffs between nonlinear RNNs and different variants of LRNNs, providing a foundation for designing LLM architectures that achieve an optimal balance between expressivity and parallelism.
arXiv:2603.05207v2 Announce Type: replace
Abstract: Retrieval-Augmented Generation (RAG) enhances large language models by incorporating external knowledge. However, existing vector-based methods often fail on global sensemaking tasks that require reasoning across many documents. GraphRAG addresses this by organizing documents into a knowledge graph with hierarchical communities that can be recursively summarized. Current GraphRAG approaches rely on Leiden clustering for community detection, but we prove that on sparse knowledge graphs, where average degree is constant and most nodes have low degree, modularity optimization admits exponentially many near-optimal partitions, making Leiden-based communities inherently non-reproducible. To address this, we propose replacing Leiden with k-core decomposition, which yields a deterministic, density-aware hierarchy in linear time. We introduce a set of lightweight heuristics that leverage the k-core hierarchy to construct size-bounded, connectivity-preserving communities for retrieval and summarization, along with a token-budget-aware sampling strategy that reduces LLM costs. We evaluate our methods on real-world datasets including financial earnings transcripts, news articles, and podcasts, using three LLMs for answer generation and five independent LLM judges for head-to-head evaluation. Across datasets and models, our approach consistently improves answer comprehensiveness and diversity while reducing token usage, demonstrating that k-core-based GraphRAG is an effective and efficient framework for global sensemaking.
arXiv:2603.14377v2 Announce Type: replace
Abstract: Prevailing High Dynamic Range (HDR) video reconstruction methods are fundamentally trapped in a fragile alignment-and-fusion paradigm. While explicit spatial alignment can successfully recover fine details in controlled environments, it becomes a severe bottleneck in unconstrained dynamic scenes. By forcing rigid alignment across unpredictable motions and varying exposures, these methods inevitably translate registration errors into severe ghosting artifacts and temporal flickering. In this paper, we rethink this conventional prerequisite. Recognizing that explicit alignment is inherently vulnerable to real-world complexities, we propose LoCAtion, a Long-time Collaborative Attention framework that reformulates HDR video generation from a fragile spatial warping task into a robust, alignment-free collaborative feature routing problem. Guided by this new formulation, our architecture explicitly decouples the highly entangled reconstruction task. Rather than struggling to rigidly warp neighboring frames, we anchor the scene on a continuous medium-exposure backbone and utilize collaborative attention to dynamically harvest and inject reliable irradiance cues from unaligned exposures. Furthermore, we introduce a learned global sequence solver. By leveraging bidirectional context and long-range temporal modeling, it propagates corrective signals and structural features across the entire sequence, inherently enforcing whole-video coherence and eliminating jitter. Extensive experiments demonstrate that LoCAtion achieves state-of-the-art visual quality and temporal stability, offering a highly competitive balance between accuracy and computational efficiency.
arXiv:2603.20831v2 Announce Type: replace
Abstract: We study local verification of graph properties in distributed networks under the framework of \emph{locally checkable proofs} (LCPs). In an LCP, a prover assigns proof labels to nodes, and a distributed verifier must make all nodes accept if the graph satisfies the property, while at least one node rejects otherwise. Each node bases its decision on a local neighborhood, called its \emph{view distance}.
Our focus is twofold. First, we study cycle existence, i.e., whether a graph contains a cycle (as opposed to cycle-freeness). We show that cycle existence admits verification with only $3$ proof labels and view distance $1$, and establish a matching lower bound. More importantly, inspired by direction-encoding techniques based on BFS distances, we introduce a novel gadget that encodes direction using only $2$ labels and view distance $3$ through repeated occurrences of the string $001101$. Although developed for cycle existence, this gadget may be useful for other verification tasks.
Second, we introduce an \emph{erroneous proof} model in which an adversary may corrupt proof labels of at most $i$ nodes within the $(2i+1)$-hop neighborhood of each node. We present an algorithmic framework, called \textbf{\texttt{refix}}, that transforms an error-free verifier into one that tolerates such errors at the cost of a view distance of $2i+1$. We demonstrate the framework on cycle existence, cycle-freeness, and bipartiteness, and establish lower bounds relating the number of errors to the required view distance. Finally, we show that our $2$-label, view-distance-$3$ verifier for cycle existence admits a $3$-round implementation in the \textsc{CONGEST} model, providing a first step toward implementing LCPs under communication constraints.
arXiv:2604.17860v2 Announce Type: replace
Abstract: Software vulnerabilities remain one of the most persistent threats to modern digital infrastructure. While static application security testing (SAST) tools have long served as the first line of defense, they suffer from high false-positive rates. This article presents TitanCA, a collaborative project between Singapore Management University and GovTech Singapore that orchestrates multiple large language model (LLM)-powered agents into a unified vulnerability discovery pipeline. Applied in open-source software, TitanCA has discovered 203 confirmed zero-day vulnerabilities and yielded 118 CVEs. We describe the four-module architecture, i.e., matching, filtering, inspection, and adaptation, and share key lessons from building and deploying an LLM-based vulnerability discovery solution in practice.
arXiv:2603.26738v3 Announce Type: replace
Abstract: While automated sleep staging has achieved expert-level accuracy, its clinical adoption is hindered by a lack of auditable reasoning. We introduce SleepVLM, a rule-grounded vision-language model (VLM) that stages sleep from multi-channel polysomnography (PSG) waveform images and generates clinician-readable rationales based on American Academy of Sleep Medicine (AASM) scoring criteria. Utilizing waveform-perceptual pre-training and rule-grounded supervised fine-tuning, SleepVLM achieved Cohen's kappa of 0.767 on a held-out test set (MASS-SS1) and 0.743 on an external cohort (ZUAMHCS), matching state-of-the-art performance. Independent expert evaluation by two trained sleep technologists further validated the model's reasoning quality, with mean scores of 3.75-3.96 out of 5 across factual accuracy, evidence comprehensiveness, and logical coherence on both datasets. By coupling competitive performance with transparent, rule-based explanations, SleepVLM may improve the trustworthiness and auditability of automated sleep staging in clinical workflows. To facilitate further research in interpretable sleep medicine, we release MASS-EX, a novel expert-annotated dataset.
arXiv:2604.08782v3 Announce Type: replace
Abstract: Large language models (LLMs) suffer significant performance degradation when user instructions and context are distributed over multiple conversational turns, yet multi-turn (MT) interactions dominate chat interfaces. The routine approach of appending full chat history to prompts rapidly exhausts context windows, leading to increased latency, higher computational costs, and diminishing returns as conversations extend. We introduce MT-OSC, a One-off Sequential Condensation framework that efficiently and automatically condenses chat history in the background without disrupting the user experience. MT-OSC employs a Condenser Agent that uses a few-shot inference-based Condenser and a lightweight Decider to selectively retain essential information, reducing token counts by up to 72% in 10-turn dialogues. Evaluated across 13 state-of-the-art LLMs and diverse multi-turn benchmarks, MT-OSC consistently narrows the multi-turn performance gap - yielding improved or preserved accuracy across datasets while remaining robust to distractors and irrelevant turns. Our results establish MT-OSC as a scalable solution for multi-turn chats, enabling richer context within constrained input spaces, reducing latency and operational cost, while balancing performance.