Named after the hundred-eyed watchman of Greek myth, Argus watches the education landscape: spotting new opportunities, pressure-testing the ventures we're building, and tracing every read back to the real-world signals behind it.
The evidence library: the raw signals the pipeline is watching across the education ecosystem. Every idea is built from these.
arXiv:2606.05121v2 Announce Type: replace-cross Abstract: Audio is continuous and interactive, yet most Large Audio Language Models (LALMs) remain offline and streaming systems usually specialize in ASR or spoken dialogue. We formalize the Audio Interaction Model, an always-on perceive--decide--respond paradigm that tracks context, decides whether intervention is warranted, and responds without stopping listening. We instantiate it with Audio-Interaction and introduce SoundFlow, coupling streaming-native data construction, comprehension-aware silence/response supervision, dual-loss training, and asynchronous FIFO inference. We also construct textsc{StreamAudio-2M, a 2.6M-item, 302k-hour corpus spanning 7 capability families and 28 sub-tasks, together with Proactive-Sound-Bench. Across 8 benchmarks, Audio-Interaction remains competitive on mainstream audio tasks while enabling spoken-instruction robustness, long-stream interaction, and proactive intervention.
arXiv:2606.02530v2 Announce Type: replace-cross Abstract: Aligning Large Language Models (LLMs) with human values often degrades their general capabilities, termed the alignment tax. Existing methods mitigate this by balancing dual objectives, which heavily rely on massive general-purpose data or auxiliary reward models. In this paper, we argue that, because safety features are inherently sparse within the output distribution, alignment requires localized modifications rather than global trade-offs. To this end, we propose SafeSteer, which performs on-policy distillation confined to safety tokens. First, we construct a safety teacher via activation steering. Based on this teacher, we develop a safety token selection algorithm. Consequently, SafeSteer restricts the reverse KL penalty to these tokens during training to preserve general capabilities. Experimental results across diverse models show that our SafeSteer achieves a superior trade-off between safety and general capability compa
arXiv:2606.01444v2 Announce Type: replace-cross Abstract: Scientific discovery is not only answer generation but revision of the representational regime in which evidence, artifacts, operations, and verifiers are typed. We develop a category-theoretic account of agentic discovery for materials science. In a fixed regime b with schema category S_b, the system state is a copresheaf I_t: S_b -> Set, and provenance is the category of elements \int_{S_b} I_t. Fixed-regime operation is an update on such states, endofunctorial only when provenance-preserving refinements are specified and preserved. Discovery is instead a verified regime transition u: S_b -> S_b': old artifacts are preserved, transported by the left Kan extension Lan_u I_t, and compared with the post-transition state to identify residual content beyond functorial transport. This separates retrieval, search, and discovery without subjective novelty. We instantiate the framework in two systems. In Builder/Breaker, a protein-mech
arXiv:2605.24817v2 Announce Type: replace-cross Abstract: As Mixture-of-Experts (MoE) architectures are increasingly adopted for scaling Large Language Models (LLMs), safety auditing becomes necessary to verify whether these models produce or facilitate harmful behaviors during operation. However, existing content-based auditing methods typically require access to user prompts, model internals, or outputs, potentially exposing sensitive user information and creating a tension between LLM safety and user privacy. On the other hand, we observe that, in MoE models, different inputs induce different sparse expert-routing patterns, which produce measurable footprints in low-level GPU execution telemetry. We refer to these hardware-observable signals induced by expert-routing decisions as expert routing telemetry; they are derived from GPU execution rather than from router logits or token-level routing assignments. Inspired by this observation, we propose RouteScan, a non-intrusive auditing
arXiv:2605.15508v3 Announce Type: replace-cross Abstract: The quadratic complexity of attention imposes severe memory and computational bottlenecks on Large Language Model (LLM) inference. This challenge is particularly acute for emerging agentic applications that require processing multi-million token sequences. We propose STS, a sparse attention mechanism that requires no model retraining. STS leverages the key insight that tokens identified as important by a smaller draft model are highly predictive of important tokens for a larger target model. By integrating into speculative decoding frameworks, STS repurposes the draft model's attention scores to dynamically construct a token-and-head-wise sparsity mask. This mask effectively prunes the expensive attention computation in the target LLM. Our evaluation shows that STS achieves a 2.67x speedup operating at approximately 90% sparsity on representative benchmark NarrativeQA, maintaining negligible accuracy degradation compared to dens
arXiv:2605.01913v2 Announce Type: replace-cross Abstract: Fine-tuning safety-aligned language models for downstream tasks often leads to substantial degradation of refusal behavior, making models vulnerable to adversarial misuse. While prior work has shown that safety-relevant features are encoded in structured representations within the model's activation space, how these representations change during fine-tuning and why alignment degrades remains poorly understood. In this work, we investigate the representation-level mechanisms underlying alignment degradation. Our analysis shows that standard fine-tuning induces systematic drift in safety-relevant representations, distorts their geometric structure, and introduces interference between task optimization and safety features. These effects collectively lead to increased harmful compliance. Motivated by these findings, we introduce REFUSALGUARD, a representation-level fine-tuning framework that preserves safety-relevant structure durin
arXiv:2601.00894v2 Announce Type: replace-cross Abstract: Large language models apply uniform computation to all inputs, regardless of difficulty. We propose PonderTTT, a gating strategy using the TTT layer's self-supervised reconstruction loss to selectively trigger Test-Time Training (TTT) updates. The gating decision itself is training-free--requiring no learned classifier or auxiliary networks; only a single scalar threshold is initially calibrated on unlabeled data and continuously adapted via EMA to maintain target update rates. Our experiments with GPT-2 models (124M to 1.5B) on code language modeling (The Stack v2, teacher-forced perplexity) demonstrate that this signal is inference-compatible, requiring no ground-truth labels. Our Reconstruction Gating achieves 82-89% Oracle Recovery while being fully training-free, significantly outperforming Random Skip baselines (up to 16% lower loss on OOD languages).
arXiv:2512.03262v3 Announce Type: replace-cross Abstract: Vibe coding is a new software development paradigm in which human engineers prompt a large language model (LLM) agent to complete complex coding tasks with little supervision. Although vibe coding is increasingly adopted, is the generated code really safe to deploy in production? To investigate this question, we propose SUSVIBES, a benchmark consisting of 186 feature-request software engineering tasks from real-world open-source projects, for which, human programmers committed vulnerable implementations. We evaluate 12 widely used coding agentic settings with frontier models on the benchmark. Disturbingly, all agents perform poorly in terms of software security. Although 57% of the solutions from SWE-Agent with Claude 4 Sonnet are functionally correct, only 11.8% are secure. Further experiments demonstrate that preliminary security strategies, such as augmenting the feature request with vulnerability hints, cannot mitigate these
arXiv:2511.17886v2 Announce Type: replace-cross Abstract: Vision-language models (VLMs) have achieved remarkable success across multimodal tasks, yet their substantial computational demands hinder efficient deployment. Knowledge distillation (KD) has emerged as a powerful approach for building lightweight but competitive models, with strong evidence from both language and vision domains. However, its application to VLMs, particularly CLIP-style models, remains limited, often constrained to small-scale teachers and narrow evaluation tasks such as classification or retrieval. In this work, we present the first systematic study of distillation across a range of CLIP-style teacher models, ranging from standard baselines to large-scale state-of-the-art models. Contrary to trends observed in NLP and vision, we find that stronger teachers do not consistently yield better students; in fact, existing distillation frameworks often fail to scale, leading to degraded performance in downstream mult
arXiv:2510.13922v2 Announce Type: replace-cross Abstract: Clinical notes contain unstructured text provided by clinicians during patient encounters. These notes are usually accompanied by a sequence of diagnostic codes following the International Classification of Diseases (ICD). Correctly assigning and ordering ICD codes is essential for medical diagnosis and reimbursement. However, automating this task remains challenging. State-of-the-art methods treated this problem as a classification task, leading to ignoring the order of ICD codes that is essential for different purposes. In this work, as a first attempt, we approach this task from a retrieval system perspective to consider the order of codes, thus formulating this problem as a classification and ranking task. Our results and analysis show that the proposed framework has a superior ability to identify high-priority codes compared to other methods. For instance, our model's accuracy in correctly ranking primary diagnosis codes is
arXiv:2509.22202v4 Announce Type: replace-cross Abstract: Large language models (LLMs) now play a central role in code generation, yet they continue to hallucinate, frequently inventing non-existent libraries. Such library hallucinations are not just benign errors: they can mislead developers, break builds, and expose systems to supply chain threats such as slopsquatting. Despite growing awareness of these risks, there is limited understanding of how library hallucinations manifest under realistic usage conditions. To fill this gap, we present the first systematic study of how user-level prompt variations influence library hallucinations in LLM-generated code. Across seven diverse LLMs, we analyse library name hallucinations (invalid imports) and library member hallucinations (invalid calls from valid libraries), examining the effects of realistic developer language and controlled user mistakes, including misspellings and fabricated libraries or members. Our findings expose systemic vu
arXiv:2409.16322v4 Announce Type: replace-cross Abstract: Alzheimer's Disease (AD) detection commonly employs machine learning classification models to distinguish between individuals with AD and those without. Different from conventional classification tasks, AD detection involves substantial within-class variation, as individuals sharing the same diagnosis may exhibit different degrees of cognitive impairment. We formulate two aspects of this issue: within-class heterogeneity and instance-level imbalance. To model such variation under binary supervision, we estimate sample-specific AD class probabilities as sample scores and develop two corresponding methods: Soft Target Distillation (SoTD) and Instance-level Re-balancing (InRe). Experiments on the ADReSS and CU-MARVEL corpora show that the estimated scores align with independent cognitive assessments and that the proposed approaches improve AD detection performance. These findings provide insights for modeling within-class variation
arXiv:2608.01347v4 Announce Type: replace Abstract: Coding-agent efficiency cannot be characterized by token count or model price alone. We study how end-to-end cost and task success depend jointly on prompt semantics, inference effort, harness policy, model, task difficulty, tool use, context management, and provider accounting. Controlled prompt experiments show that wording can change reasoning and verification behavior without changing the task. A separate SWE-bench Verified study shows that additional inference effort can improve difficult tasks for some models but can also add cost without benefit. A DeepSeek Harness extension shows that the effect of an effort-control intervention changes substantially when the harness changes, even when the model, tasks, prompts, and controller logic are held fixed. These results show that prompt, effort, and harness are interacting experimental factors rather than independent efficiency controls. We model efficiency as cost per successful task
arXiv:2607.23740v2 Announce Type: replace Abstract: As large language models move from isolated task solving toward long-term service in human environments, they require social intelligence: the ability to infer mental states, track social relations, reason over norms, and adapt behavior under context. This report presents ZenGen, an integrated framework for measuring, internalizing, and grounding social intelligence. For measurement, we introduce SoMBench, a psychology-grounded benchmark spanning 3 primary dimensions, 17 secondary dimensions, and 71 task paradigms. It controls question format, narrative perspective, and context length across 284 shared scenarios and 3,481 expert-verified instances. Evaluation of 20 representative LLMs reveals substantial headroom: the best model achieves only 72.08% overall accuracy, and none of the 17 secondary dimensions reaches the 90% near-ceiling band. For internalization, we develop ZenGen, a diagnosis-driven training recipe combining supervised
arXiv:2606.21008v3 Announce Type: replace Abstract: We present evidence that analogy is at the core of LLM intelligence. In our benchmark, LLMs compete in generating sets of analogous statements and rate each other's sets on their own understandings of factual correctness, beauty, intelligence, distinctness, length, and structural diversity. Nothing enters from outside: the only given is the game rules; every item is generated in play; the scores come from the players' ratings alone. Ground truth is replaced by the SVD of the factual rating matrix, which scores players as generators and judges at once -- to our knowledge the first eigen-equation that judges the judges for an LLM council-of-peers. For subjective criteria like beauty, judges are weighted by their rating consistency. The best generators turn out to be middling judges. GPQA Diamond -- difficult multiple-choice questions written by human experts -- could not be more different in method, yet the two benchmarks correlate at P
arXiv:2605.27268v2 Announce Type: replace Abstract: Modern Large Language Models (LLMs) are often criticized for producing repetitive and homogeneous text, despite possessing vast latent vocabularies. While previous research has focused on model knowledge and training data, we investigate the role of decoding mechanics in suppressing linguistic diversity. We introduce the Word Coverage Score (WCS), a metric that quantifies the extent to which contextually appropriate human vocabulary is mathematically pruned by standard sampling filters (e.g., Top-$p$, Top-$k$, and Min-$p$). Rather than assessing static knowledge, the WCS measures the lexical survival rate of low-frequency, high-information human words as a function of sampling parameters. By auditing open-weight models on human-authored corpus fragments, we identify which logical lexical choices are rendered unreachable by the decoder, even when they reside within the probability space. Our results provide quantitative evidence that i
arXiv:2605.01048v2 Announce Type: replace Abstract: Counterfactual prompting (i.e., perturbing a single factor and measuring output change) is widely used to evaluate things like LLM bias and CoT faithfulness. But in this work we argue that observed effects cannot be attributed to the targeted factor without accounting for baseline "meaning-preserving" modifications to text that establish general model sensitivity. This is because every counterfactual edit is a compound treatment that bundles the variable of interest with incidental surface-form variation; this violates treatment variation irrelevance. We observe prediction flip rates on MedQA of 14.9% when we surgically change patient gender. However, this is statistically indistinguishable from the flip rates induced by simply paraphrasing inputs (14.1%). In this case, it would therefore be unwarranted to conclude that the LLM is especially sensitive to patient gender. To account for this and robustly measure the effects of targeted
arXiv:2603.02760v2 Announce Type: replace Abstract: Diffusion large language models (dLLMs) have recently attracted significant attention for their ability to enhance diversity, controllability, and parallelism. However, their non-sequential, bidirectionally masked generation makes quality assessment difficult, underscoring the need for effective self-evaluation. In this work, we propose DiSE, a simple yet effective self-evaluation confidence quantification method for dLLMs. DiSE quantifies confidence by computing the probability of regenerating the tokens in the entire generated sequence, given the full context. This method enables more efficient and reliable quality assessment by leveraging token regeneration probabilities, facilitating both likelihood estimation and robust uncertainty quantification. Building upon DiSE, we further introduce a flexible-length generation framework, which adaptively controls the sequence length based on the model's self-assessment of its own output. We
arXiv:2601.20476v2 Announce Type: replace Abstract: GenAI is widespread in educational applications; however, it is known to generate content with intrinsic and extrinsic hallucination. We introduce a novel method for ICL diagram generation based on Rhetorical Structure Theory, which improves diagram faithfulness to its source text context. We find that ICL performance depends on task distribution and models' reasoning ability, with higher reasoning allowing better quality and performance for an out-of-distribution task. We perform an expert evaluation of 150 generated diagrams and analyze our findings using Bayesian GLMMs. Additionally, we use our evaluation rubric and samples from the data set for automated diagram evaluation, achieving statistically significant agreement with human evaluation.
arXiv:2601.09487v2 Announce Type: replace Abstract: The rapid evolution of Large Language Models (LLMs) has fostered diverse paradigms for automated slide generation, ranging from code-driven layouts to image-centric synthesis. However, evaluating these heterogeneous systems remains challenging, as existing protocols often struggle to provide comparable scores across architectures or rely on uncalibrated judgments. In this paper, we introduce SlidesGen-Bench, a benchmark designed to evaluate slide generation through a lens of three core principles: universality, quantification, and reliability. First, to establish a unified evaluation framework, we ground our analysis in the visual domain, treating terminal outputs as renderings to remain agnostic to the underlying generation method. Second, we propose a computational approach that quantitatively assesses slides across three distinct dimensions - Content, Aesthetics, and Editability - offering reproducible metrics where prior works rel
arXiv:2601.06519v2 Announce Type: replace Abstract: Biomedical retrieval-augmented generation (RAG) can ground LLM answers in medical literature, yet long-form outputs often contain isolated unsupported or contradictory claims with safety implications. We introduce MedRAGChecker, a claim-level verification and diagnostic framework for biomedical RAG. Given a question, retrieved evidence, and a generated answer, MedRAGChecker decomposes the answer into atomic claims and estimates claim support by combining evidence-grounded natural language inference (NLI) with biomedical knowledge-graph (KG) consistency signals. Aggregating claim decisions yields answer-level diagnostics that help disentangle retrieval and generation failures, including faithfulness, under-evidence, contradiction, and safety-critical error rates. To enable scalable evaluation, we distill the pipeline into compact biomedical models and use an ensemble verifier with class-specific reliability weighting. Experiments on fo
arXiv:2512.12613v2 Announce Type: replace Abstract: Sparse Knowledge Graphs (KGs) are commonly encountered in real-world applications, where knowledge is often incomplete or limited. Sparse KG reasoning, the task of inferring missing knowledge over sparse KGs, is inherently challenging due to the scarcity of knowledge and the difficulty of capturing relational patterns in sparse scenarios. Among all sparse KG reasoning methods, path-based ones have attracted plenty of attention due to their interpretability. Existing path-based methods typically rely on computationally intensive random walks to collect paths, producing paths of variable quality. Additionally, these methods fail to leverage the structured nature of graphs by treating paths independently. To address these shortcomings, we propose a Structural and Probabilistic framework named StruProKGR, tailored for efficient and interpretable reasoning on sparse KGs. StruProKGR utilizes a distance-guided path collection mechanism to si
arXiv:2509.20377v2 Announce Type: replace Abstract: Retrieval-Augmented Generation (RAG) has significantly improved the performance of large language models (LLMs) on knowledge-intensive tasks in recent years. However, since retrieval systems may return irrelevant content, incorporating such information into the model often leads to hallucinations. Thus, identifying and filtering out unhelpful retrieved content is a key challenge for improving RAG performance.To better integrate the internal knowledge of the model with external knowledge from retrieval, it is essential to understand what the model "knows" and "does not know" (which is also called "self-knowledge"). Based on this insight, we propose SKILL-RAG (Self-Knowledge Induced Learning and Filtering for RAG), a novel method that leverages the model's self-knowledge to determine which retrieved documents are beneficial for answering a given query. We design a reinforcement learning-based training framework to explicitly elicit self
arXiv:2508.15813v2 Announce Type: replace Abstract: A big issue in modern LLM applications is they tend to feed long context to LLM, which results in high inference cost and latency, and may exceed the context limit. Prompt compression addresses this issue by reducing the length of input context with minimum loss of generation quality, i.e, the goal of prompt compression is to shorten the LLM input while maintaining a high generation quality. To overcome these limitations, we propose SCOPE, a training-free generative prompt compression framework based on chunk-level rewriting. Unlike the existing token removal methods, our method centers at a chunking-and-summarization mechanism. Specifically, SCOPE splits a prompt into semantically coherent chunks and rewrites the chunks to be more concise. Then the chunks are reconstructed into a meaningful prompt. Additionally, we design several optimization techniques for SCOPE, effectively preserving critical information and text coherence in comp
arXiv:2506.16633v3 Announce Type: replace Abstract: Multimodal reasoning is a process of understanding, integrating and inferring information across different data modalities. It has recently attracted surging academic attention. Although there are various tasks for evaluating multimodal reasoning ability, they still have limitations. Reasoning on hierarchical visual clues at different levels of granularity, i.e., local details and global context, is of little discussion, despite its frequent involvement in human reasoning. To bridge the gap, we introduce a challenging dataset, namely GeoExplain, which evaluates explainable geo-localization. Given a street view image, the task is to predict its location and provide a detailed explanation. GeoExplain consists of 40350 panoramas-location-explanation tuples. Each instance contains a set of street-view panoramas, a location on street level, and human-expert explanations describing how the location can be inferred from the visual content of
arXiv:2506.10903v2 Announce Type: replace Abstract: Statement autoformalization plays a crucial role in formal mathematical reasoning by enabling the automatic translation of natural language statements into formal languages. While recent advances using large language models (LLMs) have shown promising capability of autoformalization, methods for automatically evaluating autoformalization remain underexplored. LLM-as-a-judge presents a promising approach for automating such evaluation, however, existing methods typically employ coarse-grained and generic evaluation criteria, which limit their effectiveness for advanced formal mathematical reasoning, where quality hinges on nuanced, multi-granular dimensions. In this work, we take a step toward addressing this gap by introducing a systematic, automatic method to evaluate autoformalization tasks. The proposed method is based on an epistemically and formally grounded ensemble (EFG) of LLM judges, defined on criteria encompassing logical p
arXiv:2505.16222v2 Announce Type: replace Abstract: With the growing use of large language models(LLMs) as evaluators, their application has expanded to code evaluation tasks, where they assess the correctness of generated code without relying on reference implementations. While this offers scalability and flexibility, it also raises a critical, unresolved question: Can LLM judges fairly and robustly evaluate semantically equivalent code with superficial variations? Functionally correct code often exhibits variations-such as differences in variable names, comments, or formatting-that should not influence its correctness. Yet, whether LLM judges can reliably handle these variations remains unclear. We present the first comprehensive study of this issue, defining six types of potential bias in code evaluation and revealing their systematic impact on LLM judges. Across five programming languages and multiple LLMs, we empirically demonstrate that all tested LLM judges are susceptible to bo
arXiv:2505.11924v4 Announce Type: replace Abstract: Intrinsic moral self-correction refers to the phenomenon where a language model refines its ethical judgments or aligns its outputs purely through prompting. While effective across diverse tasks, its mechanism remains unclear. We hypothesize intrinsic moral self-correction functions by steering hidden representations along interpretable latent directions. Evaluating six LLMs across four morality-related tasks, we demonstrate that the representation shifts induced by self-correction prompts align with contrastive steering vectors. This alignment transfers even when the steering vectors are constructed from a disjoint corpus. Notably, when applied via activation addition, these prompt-induced shifts can alter model behavior more effectively than the self-correction prompts and the steering vectors. Our findings suggest representation steering is the mechanistic driver of intrinsic moral self-correction.
arXiv:2501.10573v2 Announce Type: replace Abstract: We study the geometry of token representations at the prompt level in large language models through the lens of intrinsic dimension. Viewing transformers as mean-field particle systems, we estimate the intrinsic dimension of the empirical measure at each layer and demonstrate that it correlates with next-token uncertainty. Across models and intrinsic dimension estimators, we find that intrinsic dimension peaks in early to middle layers and increases under syntactic and semantic disruption (by shuffling tokens), and that it is strongly correlated with average surprisal, with a simple analysis linking logits geometry to entropy via softmax. As a case study in practical interpretability and safety, we train a linear probe on the per-layer intrinsic dimension profile to distinguish malicious from benign prompts before generation. This probe achieves accuracy of 90 to 95\% in different datasets, outperforming widely used guardrails such as
arXiv:2608.21343v1 Announce Type: cross Abstract: Contextualization is essential for production automatic speech recognition (ASR) systems, where user-provided phrases must be recognized accurately under strict latency constraints. Although many context-biasing methods improve recognition accuracy, they often do not address the practical requirements of modern production ASR systems: streaming inference, efficient batched decoding, user-specific context lists, and low runtime overhead. We propose TurboBias 2.0, a production-oriented framework for efficient phrase boosting in Transducer-based ASR systems. The framework extends GPU-accelerated TurboBias with a case-insensitive boosting graph and per-stream batched decoding, allowing each utterance in a batch to use an independent context-biasing configuration. This enables personalized context biasing for multiple simultaneous users without sharing or mixing their context lists. The proposed framework supports both offline and streaming
arXiv:2608.21218v1 Announce Type: cross Abstract: Predictive political question answering (QA), such as predicting how a political actor will vote, goes beyond factual lookup. External political resources offer rich historical evidence, but rarely contain the answer itself. Existing LLM augmentation methods, including actor-profile-based simulation and knowledge graph evidence injection, improve political reasoning but largely treat external resources as knowledge-based evidence, leaving prediction-relevant signals under-modeled. We identify two complementary signals for predictive political QA: actor stances that capture issue-specific preferences, and high-order structure signals that capture indirect dependencies among political actors. We propose PSL, a dual-view framework that converts semi-structured political records into inference-oriented evidence for LLMs. PSL extracts stance signals from question-relevant actor records in a semantic view, and learns structure-aware actor rep
arXiv:2608.21209v1 Announce Type: cross Abstract: The rise of agentic AI enables LLMs to access diverse user data, raising critical privacy concerns. Prior work on contextual privacy studies whether LLMs regulate information disclosure according to context-dependent norms. However, acceptable disclosure boundaries may vary across users even within the same context. To address this limitation, we introduce \textit{personalized privacy}, which incorporates user-specific disclosure preferences into privacy control. We further present P3Bench~(\textbf{P}ersonalized \textbf{P}rivacy \textbf{P}reservation \textbf{Bench}mark), a novel benchmark extending contextual privacy policies with personalized disclosure policies. Experiments show that prompt-based policies fail to reliably enforce personalized privacy policies, with Qwen2.5-7B and Gemma3-4B showing average policy ignorance ratios of 51.25\% and 74.28\%, respectively. Finally, to address this problem, we propose \textsc{Repair}, a robus
arXiv:2608.21095v1 Announce Type: cross Abstract: Retrieval-Augmented Generation (RAG) grounds Large Language Model (LLM) outputs in external knowledge, but RAG systems usually trust whatever they retrieve, creating a Security-Reliability Gap: high semantic relevance does not guarantee factual truth. Adversaries exploit this through knowledge poisoning, inserting malicious documents to cause targeted misinformation. We propose an Evaluation Agent, middleware that combines Natural Language Inference (NLI) factual verification, a five-signal poison detector with relevance-weighted aggregation, and a Trust Index T = 0.4 F + 0.35 C + 0.25 (1 - P ) with a non-linear dampener for high-contamination contexts. On TruthfulQA with Llama 3.3 70B, the agent reaches 91% accuracy and 100% precision, with 100% recall on instruction injection, while in-place edits, such as entity swaps, remain hard to detect. Across three LLMs the Trust Index stays discriminative, with a Receiver Operating Characteris
arXiv:2608.21030v1 Announce Type: cross Abstract: Video multimodal large language models have advanced significantly, yet fine-grained motion-temporal understanding remains fragile. The core bottleneck is not only sparse frame sampling, but also the lack of a complete temporal modeling pipeline for explicitly representing frame-to-frame change, enabling appearance-motion interaction, and optimizing temporal direction sensitivity. We propose COMET, a temporally grounded framework that systematically strengthens video MLLMs through explicit temporal representation, appearance-motion fusion, and direction-aware optimization. Architecturally, COMET introduces a temporal motion branch built on Taylor frame differences and injects its motion evidence into the appearance stream via temporal attention bias-enhanced cross-attention. For optimization, COMET combines temporal prior distillation with a forward-reverse TC-GRPO stage that turns temporal order into a direct learning signal and streng
arXiv:2608.20961v1 Announce Type: cross Abstract: Modern open models are hybrids: most layers are linear-attention (Gated DeltaNet, GDN) layers carrying a small fixed-size recurrent state instead of a growing key-value (KV) cache. This makes ordinary decoding memory-efficient, but hurts speculative decoding. To verify a batch of draft tokens and then roll back the rejected ones, today's systems snapshot the full recurrent state at every draft position for GDN layers, and those snapshots cannot be shared across branches of a draft tree, so a wide, high-acceptance tree becomes memory-infeasible. We remove the snapshots. Using a tree-structured WY transform of the gated delta rule, we compute every draft node's output with a single triangular solve and reconstruct only the one accepted state on commit, storing a small pseudo-value matrix instead of per-node states; the derivation depends only on the gated delta rule, not on any other architectural detail. In serving benchmarks on two scal
arXiv:2608.20868v1 Announce Type: cross Abstract: Document processing pipelines traditionally cascade optical character recognition (OCR) engines with downstream models for structured information extraction, leading to multi-stage error propagation. We fine-tune SmolDocling, a compact 256M-parameter vision-language model (VLM), to perform end-to-end key-value extraction directly from document images, jointly solving identification, localization, and association in a single pass without OCR preprocessing. We extend DocTags with specialized key, value, region, and link tags, enabling many-to-many relationships in a unified output sequence. To address data limitations, we design an augmentation pipeline combining synthetic form filling and graph-based crops that preserve complete key-value subgraphs. We further introduce a layout-aware evaluation framework extending text matching with spatial bounding box verification. On FUNSD, XFUND, and a large-scale private dataset, our model outperfo
arXiv:2608.20801v1 Announce Type: cross Abstract: While Large Language Models (LLMs) have significantly advanced reranking in recommendation, effectively leveraging item-side information remains challenging. Real-world items are described by vast, heterogeneous, and unstructured metadata, where decision-relevant signals are often implicit, noisy, or buried in long descriptions. Moreover, feature salience is highly context-dependent, varying not only across items but also across users. Existing methods often rely on item titles, fixed attributes, or static item summaries, which limit personalized and fine-grained item understanding. To bridge this gap, we propose CAIRO, a user context-aware item profiling framework for LLM-based reranking. CAIRO first structures raw metadata and reviews into objective features and subjective traits, and employs a lightweight profiler to select the most relevant information for each user-item pair with limited serving-time overhead. The resulting profile
arXiv:2608.20729v1 Announce Type: cross Abstract: Language-model agents can improve after failure or carry text across episodes without revising what counts as success. We study the narrower attribution problem of criterion revision: when criterion K0 accepts an outcome violating a broader commitment B, what observations justify saying that the system formed and persistently used K1? We require five non-compensatory conditions: criterion-failure detection, a model-emitted proposal, new-episode transfer, intervention sensitivity on the claimed carrier, and preservation. We evaluate CMB-0.1 on twelve cross-domain cases and four arms: stateless inference, append-only history, model-generated but harness-committed state, and evaluator-written oracle state. Seven mechanism fixtures yield 84 deterministic scorer trials; four local quantized artifacts yield 96 calls and 192 model-case-arm trials. No model trial satisfies all five conditions, but this zero does not establish general capability
arXiv:2608.20685v1 Announce Type: cross Abstract: Retrieval-augmented generation (RAG) has no model of time: when a fact changes across a coding session - a function is renamed, an endpoint moves, a dependency is bumped - RAG retrieves both the old and new value with near-identical similarity and cannot tell which is current, so it serves the superseded value. Paper 1 showed, on synthetic single-value benchmarks, that a deterministic (subject, relation, object) supersession memory eliminates this failure. Here we validate it end-to-end on real software history. From 707 real GitHub issues (SWE-bench Lite + Verified) we extract 130 clean atomic state transitions, a fix that changes one identifiable value from a pre-fix to a post-fix form, and render each marker-free (the stale and current statements differ only in the value). On this set, MemStrata reaches 0.91 answer accuracy versus RAG's 0.57-0.59; and, the structural result, when forced to answer RAG serves the superseded value 36-38
arXiv:2608.20670v1 Announce Type: cross Abstract: Many agentic systems must repeatedly choose between acting and abstaining, making faithful reasoning important for oversight: an explanation is useful only if it reflects the computation that produced the action. We study this problem through intervention timing in multi-party conversation, where an assistant must decide whether to speak or remain silent. This setting exposes class imbalance, asymmetric action costs, and the possibility that exposing reasoning changes the policy being audited. Using Qwen3-8B, decoded with or without chain-of-thought reasoning, we compare direct decision policies, reasoning policies, supervised fine-tuning, and reinforcement learning. We find a capability-auditability tradeoff: the strongest direct policy achieves higher quality but exposes no reasoning to inspect, while the reasoning policy provides a trace at the cost of lower performance, particularly recall of true intervention opportunities. Supervi
arXiv:2608.20661v1 Announce Type: cross Abstract: Enterprise adoption of large language models in finance is constrained less by fluency than by trust: in Financial Planning and Analysis (FP&A) and other regulated workflows, an answer is usable only if it is traceable to authoritative sources and auditable after the fact. This paper argues that retrieval-augmented generation for enterprise finance should be evaluated on auditability alongside accuracy, and presents the Knowledge-Driven Analytics Framework (KDAF), which builds ontology-driven knowledge systems through six iterative stages and retrieves evidence via Context-Aware Relevance Propagation (CARP), so that every retrieved fact carries its relationship type, confidence, and source lineage. An evaluation on FinanceBench (145 questions) compares KDAF against zero-context inference, BM25, concept-weighted lexical retrieval, and ungrounded graph traversal. First, retrieval is necessary: zero-context inference reaches 4.1% correctne
arXiv:2608.20569v1 Announce Type: cross Abstract: Are frontier models able to introspect about their internal states? Recent work suggests that under certain conditions a complex enough model can audit its own internals, call out what changed, and report back confidently about it. We tested that claim on eight open-weight models from seven families and found no such ability: asked whether their own computation had been altered, none answered better than chance. To test it we built Open-Weight Masked Introspection (OWMI), a framework that intervenes on residual-stream sites, attention heads and sparse-autoencoder features, then interrogates the model about the change against the null conditions an answer has to beat: sham runs where nothing was altered, impact-matched random perturbations, and a text-only observer that sees only the visible output. Over 78,000 measurements, no model's report discriminates a real intervention from a sham beyond chance (AUROC ~0.5007), and an equivalence
arXiv:2608.20432v1 Announce Type: cross Abstract: Formal proofs in Lean 4 that pass the kernel's type checker can nonetheless vary widely in quality. We introduce ProofJudge, an agentic LLM-as-judge system that scores formal proof quality along five dimensions beyond correctness: library leverage, automation fit, structural clarity, statement quality, and Mathlib conventions. We evaluate ProofJudge on a novel dataset of 218 declarations drawn from distinct Mathlib PRs. The judge agent is grounded by tool access to the commit the PR is applied to, enabling it to query the library state when scoring. A judge is considered aligned with human preferences when it rates the version of the PR Mathlib accepted above the initial version that was sent back for revision. All six judge models evaluated recover the reviewers' preference well above chance, from 80.8% to 63.5%, and two open-weight judges reach roughly 70% at a tenth of the best judge's cost. We release the judge harness, evaluation d
arXiv:2608.20400v1 Announce Type: cross Abstract: Agentic memory under a fixed budget involves two stages: retention and retrieval. Existing retrieval-centered paradigms implicitly assume necessary evidence survives eviction, but we challenge this by isolating a pre-retrieval failure mode: structurally indirect prerequisite eviction, in which upstream blocks weakly aligned with the query are discarded under budget pressure. We provide an operational definition of this failure, a reproducible deterministic benchmark, and per-seed trace diagnostics. Finally, we evaluate Dependency-aware Semantic Garbage Collection (DSGC), a one-hop graph-aware rule. In our main suite, DSGC improves full-chain retention from 0.03 to 0.90 under a lexical encoder and from 0.23 to 1.00 under a sentence encoder. Robustness checks then identify the budget and scaling regimes where the one-hop rule holds or degrades. Our released pipeline and failure postmortem support mechanistic analysis of retention before r
arXiv:2608.20394v1 Announce Type: cross Abstract: Industry pipelines that turn speech into supervised fine-tuning (SFT) data via multi-stage refinement are increasingly adopted but, to our knowledge, have not been publicly ablated stage-by-stage, leaving each stage's marginal value unknown. We design a production-ready speech-to-SFT pipeline in which transcript refinement (Phase 0) and SFT data quality refinement (Phase 2) are independently toggleable, yielding a 2x2 factorial design. For each condition, we generate QA-form SFT data from Korean medical and finance conference recordings and fine-tune 9 models (5 LLM families, 2.4B-70B); we evaluate with four cross-provider LLM judges, a blind six-expert human evaluation, and 3 downstream MCQA benchmarks. Our central finding: under a fixed, standard SFT recipe, improvements in QA data quality do not transfer uniformly into downstream MCQA gains. 4-judge quality rises consistently, yet the cross-model mean MCQA gain is not significant; po
arXiv:2608.21325v1 Announce Type: new Abstract: Users increasingly turn to large language models for emotional support, yet little is known about how these models actually conduct a psychotherapy interaction. We introduce an ontology of ten therapeutic moves: compact, function-based categories grounded in the MULTI-60 inventory, validated through an annotation campaign with five licensed psychologists, and scaled with a judge-based approach that matches expert agreement. Applying it to real counseling transcripts and model-led sessions, we compare the move distributions between human clinicians and a panel of frontier models. Models over-use inquiry at up to three times the human rate, neglect psychoeducation, and are strongly context-anchored: they carry forward strategies initiated by a human clinician but rarely initiate them themselves. Exposing the ontology as a set of tools roughly halves the mean deviation from the human move distribution and improves turn-level alignment with h
arXiv:2608.21315v1 Announce Type: new Abstract: That a prompt's effect is not a property of the prompt is established: prompts optimised for one model degrade on another, and rankings reorder under neutral reformatting. That evidence is about task accuracy, which cannot say whether the interaction is a fact about task machinery or about the conditional distribution itself. We ask on a readout with no task in it: the fixed-point structure of the short-window argmax map x_{t+1} = argmax_x p(x | x_{t-1}, x_t), censused from 96 starts. It is deterministic, so nothing can be helped or hurt, and it exists only at short windows -- four of six models lose it entirely by window 16 -- so everything here concerns how a model reads a fragment. Two results. First, the interaction reaches this readout at full magnitude: nine tokens of conditioning move the fixed-point fraction across most of its range, change a four-way structural class, and reorder models, while instruction tuning worth 60.5 IFEval
arXiv:2608.21265v1 Announce Type: new Abstract: Large language models often rely on Chain-of-Thought (CoT) reasoning to solve complex tasks, but verbose reasoning traces introduce substantial inference overhead. CoT compression shortens generation, yet aggressive compression may disrupt logical coherence and degrade performance. We formalize this trade-off as the \textit{Context-Generation Substitution Law}, where explicit reasoning context substitutes for part of decode-time generation. Based on this principle, we propose \textit{Memory-Augmented Compression}, a training-free framework that constructs reusable reasoning memories from historical traces and retrieves them as prefill-side scaffolds. Rather than using raw demonstrations, these memories summarize reusable reasoning patterns, key constraints, and critical operations to compensate for information lost during compression. Experiments show that Memory consistently improves prompt-based Chain-of-Draft (CoD) compression across m
arXiv:2608.21252v1 Announce Type: new Abstract: Question answering (QA) over long, connected documents remains challenging because relevant evidence may span multiple entities and their relationships. Existing retrieval-augmented generation (RAG) methods typically index documents as raw chunks and retrieve them through embedding similarity. Their performance degrades when chunk boundaries separate entities from supporting evidence or when a question requires multi-hop reasoning across the corpus. We propose EnSI-RAG (Entity-Structure-Indexed Retrieval-Augmented Generation), a framework that constructs a query-independent, entity-centered index. Each record (e, t, k, v) represents an entity e, its type t, a semantic category k in {property, relation, aspect}, and a value v, while retaining links to the original source passages. At query time, these records serve as retrieval handles, and an LLM synthesizes the retrieved passages into the final answer. This design separates evidence loca
arXiv:2608.21249v1 Announce Type: new Abstract: While recent large language models (LLMs) have achieved promising results on individual patent drafting tasks, they fundamentally fail to investigate the core challenge of real-world patent drafting: generating a complete and legally coherent patent application directly from early-stage invention materials. Prior work predominantly assumes later-stage, highly structured, or already legalistic inputs. However, real patenting workflows begin with informal, de-legalized disclosures authored by inventors. To bridge the gap, we introduce Dis2Pat, a disclosure-to-patent dataset that reflects realistic patenting workflows by requiring the generation of complete patent applications directly from inventor-style, de-legalized disclosures. Given the inherent difficulty of long-form, legally constrained patent drafting and the strong privacy requirements, we further propose a strong baseline named Patent-MAF. It is a multi-agent framework for locally