Papers

Large language models (LLMs) excel at complex tasks like question answering and summarization, thanks to their ability to handle long-context inputs. However, deploying LLMs is costly, not only due to the high computational demands of quadratic complexity of self-attention and auto-regressive generation, but also because of the significant memory overhead required for storing the key-value (KV) cache during inference. To reduce the memory cost, existing KV-cache eviction strategies leverage the sparsity in attention to selectively store a subset of tokens. While reducing the memory footprint, such approaches show a considerable drop in performance, especially in tasks that require long-context reasoning. We identify that the drop in performance is linked to a reduction in the coverage of unique tokens. Additionally, we theoretically show that reduced coverage limits the mutual information between inputs and outputs, thereby impairing predictive accuracy. To this end, we introduce K-VEC, a novel coverage-aware KV-cache eviction strategy that prioritizes token coverage while evicting tokens in the cache. K-VEC introduces a cross-head and a cross-layer coverage module to enhance token retention across attention heads and model layers, mitigating performance degradation caused by low coverage. Evaluated on 16 LongBench subsets, K-VEC exhibit up to 10.35 points improvement over the existing methods under the same eviction rate and memory constraint. Comprehensive evaluations validate the effectiveness of our approach and demonstrate its potential for efficient LLM deployment in resource-constrained settings.

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing tasks. However, their tendency to generate hallucinations, namely factually incorrect or unfaithful outputs, poses a critical obstacle to their deployment in high-stakes applications. Although recent hallucination detection methods have made encouraging progress, they typically rely on costly output-level consistency checks or static hidden-state probes that capture shallow dataset-specific patterns, leading to substantial degradation under cross-dataset evaluation. In this work, we propose AURORA, a novel hallucination detection framework that shifts the focus from static representations to the weight-gradient dynamics of LLMs. Our key insight is that hallucinated and faithful answers induce qualitatively different gradient update patterns on the model's parameters. Specifically, hallucinated samples trigger asymmetric and structurally misaligned gradients, which can be captured through two complementary features: (1) the skewness of the cosine similarity distribution between weight matrices and their gradient update directions, and (2) the rotation ratio, which quantifies how much the gradient update reorients the singular-vector basis of weight matrices via SVD. AURORA achieves strong hallucination detection performance across four model families and four benchmark datasets. Further analyses demonstrate that our method scales effectively across model sizes and transfers to out-of-domain tasks, including mathematical reasoning and vision-language scenarios.

Large language models (LLMs) can leave subtle stylistic traces in assisted text; one of the most cited is the em-dash (Unicode U+2014). Yet no one has measured whether em-dash use has changed in the scientific literature. This study, pre-registered on the Open Science Framework (HFT8C), used the full set of medRxiv full-text XML preprints from the official Text-and-Data-Mining resource. The primary cohort was first, original versions deposited 2020-2025 with an extractable Discussion section of at least 500 characters (N = 69,632). The primary endpoint was the presence of at least one em-dash in the Discussion; the principal measure was the absolute change in its prevalence between the pre-ChatGPT era (before 30 November 2022) and the post-ChatGPT era, estimated with a logistic model with standard errors clustered by first author. The analysis plan (six supporting analyses, six sensitivity analyses, two falsification tests) was frozen before any confirmatory result was computed. Em-dash prevalence in Discussion sections rose from 4.23% before ChatGPT to 11.58% afterward, an absolute increase of 7.35 percentage points (95% CI 6.94-7.77; odds ratio 2.96, 95% CI 2.77-3.17). The rise was not a sharp jump but a gradual, delayed acceleration: near 4% through 2023, 8.0% in 2024, and 20.3% in 2025. The effect survived every feasible sensitivity analysis (7.35-7.60 pp) and both falsification tests; a placebo split within the pre-LLM era showed no meaningful change (+0.13 pp, 95% CI -0.33 to +0.58), and was essentially absent in boilerplate sections. Independent LLM-associated lexical markers and within-paper section comparisons pointed the same way. The em-dash is a population-level indicator, not a per-paper detector of LLM use, and the design cannot establish causality; it shows that something in how scientific literature is written changed markedly in the early 2020s, and roughly when.

Popular ASR test sets adopt inconsistent conventions for numbers, disfluencies, entities, and casing, while standard normalizers erase the format distinctions users care about. Current benchmarks therefore cannot measure whether a model follows user preferences for output style. We introduce PreferenceASR, a test set evaluating ASR systems on their ability to follow natural-language preference instructions across four categories: normalization, entities, disfluencies, and case. Built from seven open-source corpora via a two-stage LLM-assisted pipeline with human verification, it is evaluated with a preference-aware normalizer that selectively skips steps matching the active instruction. Benchmarking four models shows rankings shift across preference types, exposing quality differences traditional evaluation obscures. We publicly release the dataset.

Activation steering can reliably change an agent's output by modifying its internal activations. Yet arriving at the same answer need not involve the same computation: behavioral equivalence does not imply mechanistic equivalence. We test whether an activation edit changes the state used in subsequent computation or biases that computation toward the desired output. In a controlled state-tracking task, trace supervision creates an editable register: changing its internal value makes the model apply the next operation to the edited state. We next ask whether widely used mean activation steering admits a similar interpretation: does steering recreate the internal configuration the model naturally uses when performing a task? If so, transplanting that activation should be effective, and steering should remain effective at the scale of the natural source-to-target activation change. Neither prediction holds in the two selected model-task settings. Replacing the activation at one layer produces less than 2% of the target-answer margin gain from patching through all remaining layers, while natural-scale steering is similarly ineffective. The steering vectors have norms 77 and 26 times the median natural change in Qwen and Llama, producing 54% and 82% of the reference effect. Thus, a successful steering intervention need not reproduce the natural target activation at the intervention layer. More generally, an intervention should be interpreted as changing the computational state only when a later computation uses the edited value according to the semantics of that state.

The verbose context problem occurs when structured concepts have token-inefficient textual representations. This bottleneck is acute in population health: cohort-level analysis of longitudinal patient records requires reasoning over thousands of medically-coded events, often exceeding 400K tokens in total. We present PopMedQA, a benchmark isolating this problem through computational tasks on groups of longitudinal patient records. We construct the benchmark using neopatient, a new library for language-controlled generation of artificial patient records. Through extensive ablations – including prompting strategies, prompt compression, and agentic decomposition – we find that domain-independent methods fail to alleviate the verbose context problem. There remains significant opportunity to exploit domain-specific structure in language model inputs for population-scale reasoning.

Skill memories can improve agentic reinforcement learning by reusing past experience as textual guidance, but retrieved skills are not oracular: they may help in one state while misleading the same policy in another. This makes the common privileged-teacher assumption fragile, namely that a skill-conditioned prompt can be treated as a fixed teacher for the no-skill prompt. We introduce UCOB, a framework for learning to utilize and evolve agentic skills via credit-aware on-policy bidirectional self-distillation. UCOB treats skill-conditioned and no-skill prompts as two on-policy context views of the same model, compares their return-to-go within the same task and anchor state, and uses the higher-return view as the local teacher. This local credit signal internalizes useful skill-conditioned behavior, corrects misleading skill usage, and guides task/state skill memory updates, utility-aware retrieval, and reflection self-training. Experiments on agentic tasks, including ALFWorld, WebShop, and Search-QA, show that UCOB outperforms skill-free RL, skill-memory baselines, and self-distillation methods across model scales, with up to 23.5 and 18.0 point gains over SOTA baselines on ALFWorld and WebShop. Ablations and analyses further validate its core mechanisms, continual adaptation across environments, and modest training overhead. Code is available at https://github.com/TU2021/UCOB.

When humans translate, not every word depends equally on the surrounding context. Some tokens, particularly function words like pronouns and auxiliaries, rely heavily on preceding or following sentences, while others, such as proper nouns, do not. Understanding this inherent context sensitivity is essential for evaluating whether machine translation systems use context in human-like ways. However, existing approaches to analysing context usage rely on discourse-specific test sets or model internals, making them narrow or model-dependent. We propose a post-hoc, model-agnostic framework to quantify context sensitivity at lexical and syntactic levels using two measures derived from word alignments: fertility (number of target tokens generated per source token) and entropy (stability of fertility patterns across contexts). Using reference translations for three language pairs (German $\leftrightarrow$ English, English $\rightarrow$ Hindi) under four context conditions, we show that context selectively redistributes generative responsibility from source to context tokens without altering overall fertility. Function words show the largest fertility reductions, while content words remain stable, suggesting that context resolves ambiguity rather than adding new information. Our framework provides a ground-truth characterisation of selective context usage in human translation, establishing a diagnostic baseline for evaluating machine translation models.

While reinforcement learning (RL) significantly enhances LLM reasoning, its efficacy is severely undermined by Pre-RL data overlap, where RL datasets overlap with pretraining or SFT corpora, causing models to exploit shortcuts by memorizing correct answers and fabricating post-hoc reasoning. To address this, we introduce HIPPO, a novel RL framework that integrates hint-injected aggregation with a tailored pairwise reward model. By utilizing hint injection to deliberately trigger overlap-induced behaviors, the resulting traces naturally serve as explicit anchors for pairwise comparison. This provides highly discriminable preference signals, enabling a lightweight judge model to reliably distinguish genuine reasoning deduction from shortcut-driven rationalization, while the pairwise formulation ensures stable and robust optimization compared to standard PRMs. Extensive experiments demonstrate that HIPPO yields substantial improvements over standard baselines and generalizes effectively to out-of-distribution general tasks, showing it extracts authentic, transferable reasoning skills rather than superficial shortcut patterns.

Medical question-answering benchmarks rarely cover the maternal, neonatal, child, and reproductive-health questions a nurse-midwife asks, and, to our knowledge, no public chunk-level relevance benchmark exists for maternal-health guideline retrieval. We release two benchmarks that fill these gaps. mamabench is a scope-filtered QA set of 25,949 items assembled from seven existing expert-authored sources across multiple-choice, short-answer, and rubric-graded tracks; to help users calibrate the LLM judge that scores the rubric track, we re-scope HealthBench's physician-labelled meta-evaluation to the domain. mamaretrieval pairs 3,185 clinical queries with graded (0-6) relevance labels over a 63,650-chunk maternal-health guideline corpus, using a decomposed rubric that distinguishes a chunk that answers a query from one merely on its topic. Three decisions shape both: assemble and filter expert sources rather than author questions, grade relevance rather than binarise it, and measure and disclose the limits of the labels – scope-classifier agreement, a frontier-judge check, and a pooling-completeness audit – rather than treat them as an oracle. A companion paper uses the benchmarks to evaluate a deployed on-device assistant; both are released openly for research.

Inverse design of metal-organic frameworks (MOFs) requires navigating combinatorial spaces with costly property labels and opaque machine-learning models. We introduce LLM4MOF, a closed-loop multi-agent framework that converts a natural-language target into chemical hypotheses, constraints, diagnostic tests, and feedback. One agent proposes interpretable hypotheses over metal nodes, linkers, pore geometry, and functionality. Another converts them into constraints selecting MOFs defined by a node, linker, and topology. The Matchmaker forms four beams to attribute gains to geometry, chemistry, or metal choice: full hypothesis, chemistry, metal only, and random baseline. Blind to database landscapes, LLM4MOF enriches top performers across six adsorption, separation, and electronic-structure tasks within 400 evaluations. It also designs and live-simulates de novo MOFs spanning H2 storage, SF6 capture, and C2H6/C2H4 separation, deriving a distinct design rule for each objective. Under an identical nominal evaluation budget it consistently outperforms random search, Bayesian optimization, and genetic algorithms, and the outcome is insensitive to the language-model backend. All of this uses not a single property-labeled training structure, whereas generative alternatives train on thousands to hundreds of thousands. A new objective requires only a natural-language request: interpretable inverse design at a fraction of the data cost of existing approaches.

Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists. We evaluate five defense paradigms (no defense, static steering, CAST, AlphaSteer, probe-gated) across seven instruction-tuned models (7-31B) and five attack types (GCG, AutoDAN, DeepInception, prefilling, intent laundering). Our central finding: prompt-time activation defenses are structurally blind to prefilling attacks. AlphaSteer achieves 0% attack success on GCG, AutoDAN, and intent laundering but 50% on prefilling. We prove a corollary: any defense that gates intervention on a single layer's activation alignment with a benign reference (cone, subspace, or null-space) is blind to attacks that craft activations to lie inside that reference, whether checked at prompt time or per token. As its constructive contrapositive we introduce response-time probing: a linear probe on the model's hidden state at the first generated tokens, with AUROC 0.97-1.00 across all seven models. Combined with a halt, it cuts prefilling attack success to 0/40 on every model with 0% benign false positives, outperforming Llama Guard 3. Cross-template generalisation depends on probe depth, so we scope the claim to the canonical prefilling-template family. Composing the response-halt with AlphaSteer's null-space steering gives an orthogonal split (the halt catches prefilling, AlphaSteer catches semantic attacks), reaching defense success 0.983 on Mistral and 0.994 on Llama and dominating both components. We further show MMLU fails to capture steering's true utility cost, which appears as behavioral hedging rather than factual loss, and that diverse negative training sets cut probe false positives from 80-100% to near zero. Code, attacks, per-sample results, and the judge prompt are released.

Existing multi-agent debate frameworks suffer from two critical limitations: they rely on static architectures where agent roles and coordination patterns are fixed at design time, and they require instantiating multiple model copies, incurring substantial computational overhead. We propose Mixture of Debaters (MoD), a unified framework that enables dynamic self-debate within a single model by leveraging the Mixture-of-Experts paradigm. We address three key challenges in adapting MoE for dialectical reasoning: (1) dual-routing that decouples role allocation from process flow, dynamically determining when to debate versus when to synthesize; (2) momentum switching that smooths token-level routing with local context, reducing expert-switch jitter; and (3) unified self-debate that encapsulates diverse debating personas into lightweight expert modules, eliminating inter-agent communication while preserving behavioral diversity. Extensive experiments on multimodal benchmarks demonstrate that MoD outperforms both single-model baselines and conventional multi-agent systems, achieving superior accuracy with 3.7x lower latency and 87% reduction in token consumption.The source code can be accessed at https://github.com/YongLD/MoD.

Model routing balances solution accuracy and computational cost by selecting among models of varying capabilities. While recent multi-round frameworks interleave reasoning and planning, we identify a structural failure mode termed Trust Region Collapse. We demonstrate that the deep coupling of reasoning and routing, exacerbated by the dominance of strong pre-training priors under sparse supervision, leads to degenerate local optima where capable experts are systematically suppressed. To decouple these processes, we propose $\textbf{EntroRouter}$, a single-round routing framework that treats entropy regulation as a core objective. We first initialize the policy via Soft Supervision, fitting a distribution of suitable models to establish a high-entropy prior for exploration. Subsequently, we stabilize Reinforcement Learning using a Soft Anchor, which utilizes offline capability estimates to orchestrate controlled entropy contraction within a safe trust region. Extensive experiments demonstrate that EntroRouter retains 98.3% of the strongest expert's accuracy while reducing computational costs by 48.25%.

There has been increasing interest in exploring the capabilities of advanced large language models (LLMs) in the field of information extraction (IE), specifically focusing on tasks related to named entity recognition (NER) and relation extraction (RE).Although researchers are exploring the use of few-shot information extraction through in-context learning with LLMs, they tend to focus only on using correct or positive examples for demonstration, neglecting the potential value of incorporating incorrect or negative examples into the learning process.In this paper, we present LC-ICL a novel few-shot technique that leverages both correct and incorrect sample constructions to create in-context learning demonstrations. This approach enhances the ability of LLMs to extract entities and relations by combining positive samples with negative samples annotated by error-cause labels. These labels expose more detailed error features in erroneous examples, enabling the model to understand why similar predictions fail and avoid repeating such errors during inference.Specifically, our proposed method taps into the inherent contextual information and valuable information in hard negative samples and the nearest positive neighbors to the test and then applies the in-context learning demonstrations based on LLMs. Our experiments on various datasets indicate that LC-ICL outperforms previous few-shot in-context learning methods, delivering substantial enhancements in performance across a broad spectrum of related tasks. These improvements are noteworthy, showcasing the versatility of our approach in diverse scenarios.

Sinhala is a morphologically rich abugida spoken by roughly 16 million people in Sri Lanka, and to date, there are no publicly available real-world datasets for page-level Sinhala OCR. All previous studies for assessing Sinhala OCR models have used artificially generated data. To bridge the gap, we introduce sinhala-ocr-lk-acts-1010, an annotated dataset of 1,010 page-level images and their transcriptions collected from Sri Lankan Legislative Acts published between 1981-1989 and 2000-2019, split into 707 training examples, 101 validation examples, and 202 testing examples. Three models based on deep learning-based visual language processing, namely DeepSeek-OCR V1, DeepSeek-OCR V2, and LightOnOCR-2-1B, are fine-tuned using QLoRA in 8 experiments conducted on consumer and cloud GPUs. LightOnOCR-2-1B is the top performer, achieving a CER of 1.05% across all test examples, outperforming state-of-the-art open-source OCR models such as Surya-OCR (8.84%) and Tesseract v5 (10.69%), as well as commercially available OCR models such as Google Document AI (2.06%). Our results suggest that LightOnOCR-2-1B outperforms other baselines on real-world OCR tasks and maintains consistent performance across all print periods, even when documents are severely degraded.

Medical large language models are commonly adapted with a fixed low-rank budget, even though medical questions differ substantially in confidence, clinical coverage, and cross-domain difficulty. We study adaptive rank budgeting for parameter-efficient medical question answering: for each question, the adapter decides whether to activate a small, medium, or large subset of LoRA rank channels. The central challenge is that a naive adaptive budget router can collapse to unstable choices or spend capacity without improving shifted benchmarks. We propose TriageRA-CCF, a source-side teacher for adaptive rank-budgeted LoRA. It combines three signals computed only from source training data: base-model answer confidence, metadata-cell clinical coverage, and a counterfactual close-miss proxy. These signals supervise a straight-through budget router over active ranks {2,4,8}, together with budget-cost, entropy, and rank-balance regularization. Under a matched CMB-source training protocol, TriageRA-CCF achieves the best average accuracy among LoRA, DoRA, and MoELoRA baselines on both Qwen3-8B and Llama3.1-8B. The gains are modest and non-uniform across benchmarks: +0.21 average points over the strongest external baseline on Qwen3-8B and +0.16 on Llama3.1-8B. Component ablations show that confidence, coverage, and counterfactual signals all provide useful budget supervision, but their combination is not monotonically best on every backbone.

We identify intervention bias as a previously unquantified failure mode of zero-shot large-language-model (LLM) educational advisory agents: without task-specific training, they recommend action when a hindsight-optimal oracle policy mandates inaction. In a six-arm ablation on the Open University Learning Analytics Dataset (N=800 students, four temporal cutoffs), at day 56 – when the oracle designates 70.1% of students as needing no intervention – zero-shot GPT-4o recommends action for 73%, a 43 percentage-point false-positive rate. Commercial RAG and SQL-augmented retrieval are comparably miscalibrated; at 10,000 students this implies about 4,300 unnecessary advisor contacts per cycle. Supervised policy learning eliminates this bias: a trajectory-conditioned ONNX Decision Transformer (DT) and a snapshot XGBoost classifier, trained on the same oracle-labelled trajectories under strict prefix-only features, both achieve near-zero calibration error. The DT reaches macro-F1 0.79 (macro-recall 0.85) across all five action classes, predicting even the rare load-reduction action without collapsing, at a 0% action flip rate and sub-5 ms CPU decision latency. The two supervised arms are on par; the DT's edge over XGBoost at the final cutoff is indicative only (unpaired across cohorts). Scope: we validate Stage-2 decision-making (EAV state vector to supervised policy) under controlled oracle input from structured OULAD data; high fidelity reflects feature-oracle alignment, not general high-stakes-AI capability. The most robust finding is the intervention-bias contrast, not the absolute accuracies. We also show an Evaluation Gap: LLM-as-judge scoring (DeepEval G-Eval) is blind to intervention bias, rewarding fluent over-prescription rather than decision quality.

LLM agents carry conclusions across steps and sessions in compressed memory, and memory products (e.g., mem0, LangMem) rewrite conversation into stored "facts" that later steps trust. We show this rewriting manufactures confidence: across our constructed agent settings, a casual, hedged remark becomes a confident, dated assertion the agent then obeys like a verified fact, granting every above-clearance request it faces. No attacker is needed: a role that was true once and never corrected is stored as a flat fact and acted on like a deliberate injection. We then isolate what the agent responds to. It is not the source: attributed, unattributed, and even forged "system of record" claims all grant alike. It is the confidence of the phrasing. A hedge is discounted, a flat assertion is obeyed, and this holds with no special keyword. Not all hedges are equal, though: the evidential register is the least-discounted, with "reportedly" obeyed like a flat assertion on most models. The obvious fixes fail. A passive "unverified" tag is ignored, and an active "do not trust this" instruction escalates even correct memory, so it is safe only by refusing to decide. The real fix lives in the store: keep the tentative phrasing rather than upgrade it. But that is hygiene, not a defense against an attacker who can simply write a confident lie. The deployable lesson is narrower and constructive: a single load-bearing memory is the hazard, and one redundant source restores correct decisions. We release the harness and demonstrations.

We introduce the Complexity Ceiling Benchmark (CCB), a controlled evaluation of how language-model reasoning decays as the number of required sequential steps grows. CCB fixes the semantic content of a task and varies only its depth N in {5,...,50} across three structurally distinct regimes: grounded spatial state-tracking, abstract symbolic pointer manipulation, and transitive relational inference. Across 6,000 trials over five frontier and open-weight LLMs we find a consistent pattern of geometric per-step decay with widely separated domain ceilings: on the first two regimes the strongest models retain pd>0.92 across N=50; on the third every model collapses by N=5, with the best model's 50%-success horizon at H0.5~4.7 steps despite pd=0.863. A trace-level metric (TFBC) shows that 14.5% of correct answers across the benchmark are reached via incorrect intermediate reasoning. Forced verbose state-tracking does not move the ceiling (McNemar p=1.000), and the mean step at which reasoning first diverges, k*, predicts within-domain accuracy better than parameter count. CCB and the geometric decay model together reduce a model's long-horizon reasoning profile to one interpretable number per task family.

Emotion recognition of song lyrics is a challenging task since lyrics may not necessarily align with the overall emotion of a song. As a result, lyrics annotation remains largely underexplored. Drawing inspiration from research in large language model (LLM) assisted annotation, we examine the alignment between humans and LLMs for annotation of lyrics by creating a new sentence-level dataset of lyrics. Our observations highlight the subjectivity of the task and the inherent challenges. Following this, we present a hybrid annotation framework that optimizes human and LLM annotation by predicting potential misalignment in annotation.

Reasoning large language models (LLMs) have recently made much progress in complex problem-solving, leveraging internal reasoning (or thought) to guide their solution generation. However, existing LLM-based counseling agents, including those using Motivational Interviewing (MI), generate responses without explicitly aligning thoughts with counseling techniques, limiting their effectiveness. We propose MIThinker, a lightweight thinking model that generates therapeutic thoughts to guide MI counseling agents in strategy selection and response generation. To overcome the lack of annotated thought data, we introduce AugR1-MI, an automated pipeline that reverse-engineers counselor's thoughts from observed responses. Through two-stage training combining supervised fine-tuning and reinforcement learning, MIThinker demonstrates improved theory-of-mind assessment and strategy alignment. Comprehensive evaluations show that MindfulMI, our agent leveraging MIThinker, achieves MI competency comparable to state-of-the-art systems with an order of magnitude less computation.

Large language models (LLMs) demonstrate broad reasoning abilities but struggle with accuracy and reliability in specialized domains such as travel, where reasoning depends on precise definitions, rules, and expert-defined conceptual frameworks, and where confident but unfounded outputs arise from a reasoning failure in which the model has not internalized the underlying domain graph rather than from missing domain knowledge alone. We propose a modular pipeline for building a travel-domain reasoning LLM grounded in an expert-designed knowledge graph (KG). Our pipeline integrates a travel KG that encodes domain entities and their relationships, a bottom-up construction procedure that walks the KG to produce multi-hop question answer (QA) pairs, a supervised fine-tuning stage that embeds the domain knowledge into a reasoning-capable LLM using the generated QA pairs as auditable reasoning traces, and a travel-domain benchmark dataset that measures the fine-tuned model's accuracy and calibration. We evaluate our approach using Qwen3-4B with LoRA adaptation. Our reasoning model achieves an $82.4\%$ exact match on the benchmark. This performance significantly outperforms the pretrained Qwen3-4B baseline at $22.4\%$. A calibration analysis decomposes the residual $17.57\%$ of errors into two distinct failure modes: an over-confident multi-label decoder that predicts both correct answers plus one spurious option on most dual-answer mistakes, and a smaller reasoning failure on single-answer questions where the supporting facts are present in the KG but the model fails to reconstruct the correct multi-hop path. This split confirms that explicit KG-grounded reasoning substantially improves the accuracy and uncertainty interpretation of LLMs in specialized domains, and isolates per-option calibration and trace-length-aware decoding as the next axes of improvement.

Despite the capability of parallel decoding, diffusion large language models (dLLMs) require many denoising steps to maintain generation quality, motivating recent research on efficient decoding strategies. However, existing studies have reported inconsistent evaluation results even under seemingly identical evaluation settings, risking biased conclusions about dLLM decoding methods. To understand this evaluation concern, we conduct a rigorous evaluation of current decoding methods for dLLMs across diverse evaluation settings. Surprisingly, our analysis reveals that the ranking of decoding methods is highly sensitive to the choice of prompt templates. Single-template evaluation can lead to an illusion that decoding methods improve inference efficiency without performance degradation. Through comprehensive experiments, we find that current parallel decoding methods consistently underperform the single-token decoding baseline, failing to overcome the speed-quality trade-off. We further identify this evaluation inconsistency as the high sensitivity of parallel decoding methods to minor variations in prompt templates. Our experiments show that an effective prompt template can achieve strong evaluation results even with fewer denoising steps, markedly outperforming the marginal gain from increasing denoising steps. Beyond prompt templates, our experiments indicate that overlooked evaluation settings can also notably affect the assessment of decoding methods. Based on these findings, we propose practical guidelines for the reliable evaluation of decoding methods in dLLMs.

LLM agents handle user requests on behalf of organizations through tool calls and must follow the company policies stated in their system prompts. Prior work approaches this as a safeguarding problem – external checks that block non-compliant agent actions. We argue that policy adherence is a broader problem: real workflows unfold across many turns, require explicit user confirmation and prerequisite reads, and hinge on the content of the dialogue rather than on any single argument value. Meeting this bar requires (i) full conversation context, (ii) self-reasoning over the policy and the current dialogue, and (iii) conversation-specific remediation that guides the agent's next turn – three capabilities that prior safeguard work has often underestimated. We introduce POLICYGUARD, a sub-agent verifier that shares the agent's view of the dialogue, reasons over the policy in context, and provides actionable feedback for the agent's next turn. On tau^2-BENCH airline across three vendors (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro) with four trials per setting, POLICYGUARD improves PASS4 by +12.0 / +6.0 / +12.0 pp. Per-call analyses show POLICYGUARD achieves higher policy-violation recall while blocking roughly half as often as argument-level guards.

Block Diffusion Language Models (BD-LMs) improve diffusion-based text generation with KV caching and flexible-length generation. A natural next step is to extend them from Single-Block Diffusion (SingleBD) to Multi-Block Diffusion (MultiBD), where a running-set of consecutive blocks is decoded concurrently for inter-block parallelism. However, existing BD-LMs are mostly trained under teacher forcing, where the model observes only one noisy block conditioned on a clean prefix. While the recent diffusion forcing strategy introduces visibility among multiple noisy blocks, its training states still differ from MultiBD inference, where decoding operates on a bounded running-set with heterogeneous slot-wise noise patterns. To bridge this gap, we propose Multi-Block Diffusion Language Models (MBD-LMs), obtained by post-training BD-LMs with Multi-block Teacher Forcing (MultiTF). MultiTF integrates teacher forcing and diffusion forcing by training on bounded noise-groups conditioned on clean prefixes, with randomized noise-schedulers that better match MultiBD inference states. To make MultiBD practically executable, we further introduce an optimized decoding algorithm based on the Block Buffer mechanism that preserves prefix-cache reuse, keeps input shapes static, and translates increased decoding parallelism into wall-clock acceleration. Empirically, MBD-LLaDA2-Mini increases average Tokens Per Forward pass (TPF) from 3.47 to 6.19 and improves average accuracy from 79.95% to 81.03%; when combined with DMax, MBD-LLaDA2-Mini-DMax reaches an average TPF of 9.34 with only a 1.02% accuracy drop on math and code benchmarks.

OCR systems, ranging from classical engines to specialised OCR vision-language models (OCR-VLMs) and frontier multimodal LLMs, report strong results on English and Chinese document benchmarks, yet their behaviour on Indic scripts is largely uncharacterised. We benchmark ten systems on Devanagari (Hindi): classical EasyOCR; open VLMs (Qwen2.5-VL-3B, Qwen3-VL-8B, olmOCR-7B); specialised OCR-VLMs (DeepSeek-OCR, Unlimited-OCR); and frontier closed models (Gemini 2.5 Flash, Claude Opus 4.7, GPT-5.5, Mistral OCR), across four synthetic degradation conditions and 300 real printed scans. We report four findings. First, on clean rendered text all ten cluster within chrF++ 91 to 98, so synthetic text does not separate them. Second, under degradation the specialised OCR-VLMs are the most fragile: DeepSeek-OCR suffers rare but catastrophic repetition failures (outputs up to 71 the reference length) that wreck its corpus mean even though its median is the best of any system, which is why we report median and catastrophic-rate instead of the mean. Third, on real scans nine of the ten systems collapse (EasyOCR falls from chrF++ 93.6 to 58.3) and the field spreads across a 76-point range, so synthetic renders badly overstate Devanagari quality. Fourth, strong English OCR does not predict Indic OCR: GPT-5.5 drops to chrF++ 58.5 (tying classical EasyOCR) and olmOCR-7B, the model behind olmOCR-Bench, falls to 40.5, while the open Qwen3-VL-8B (75.2, runnable on a single 24 GB GPU) beats GPT-5.5 and approaches Mistral; Gemini and Claude lead at 86.3 and 82.2. An error taxonomy separates surface errors (numerals, punctuation) from structural ones (conjuncts, matras, nukta), and a byte-level (ByT5) post-corrector improves a cheap engine on its own error distribution (chrF++ +1.2 to +1.5) but does not transfer across engines. We release the benchmark, code, and models.

Do language models know when they are being tested? This question matters for AI safety: a model that recognises an evaluation context could alter its behaviour strategically, making downstream benchmarks harder to interpret. Using 11 models spanning Qwen 2.5, Gemma 2, and Llama 3.2, we find a systematic size-dependent shift in representational depth: in both Qwen 2.5 and Gemma 2, the layer at which evaluation-awareness is most linearly recoverable moves from late layers in smaller models to early layers in larger ones. This suggests that scale changes not only the strength of evaluation-awareness but also where it is most linearly recoverable in the network. This depth shift helps explain why within-family scaling trajectories are non-monotonic or inverse rather than smooth and family-general, showing that a simple universal power-law account is not supported under denser within-family sampling. Finally, white-box probe signals are consistently stronger than black-box behavioural expression, and the relationship between the two varies by family in ways not predicted by probe AUROC alone.

Open-ended scientific discovery with large language models (LLMs) increasingly operates as a long-horizon loop of hypothesis search and verification, where a reward signal guides which hypotheses to test next. A notable recent example is AutoDiscovery, which uses "Bayesian surprise" - the belief shift an LLM undergoes after observing evidence for a hypothesis - as both a discovery metric and a reward for search. We first observe that AutoDiscovery treats surprisal as a static quantity, while surprisal in human reasoning is non-stationary - it is defined relative to beliefs that evolve with experience, a prerequisite for continual scientific discovery. We address this mismatch with evidence-informed LLM beliefs: priors updated with evidence from previous hypotheses to compute non-stationary surprisal for new hypotheses. We compare in-context belief-updating mechanisms and find that embedding-based retrieval-augmented generation over prior discoveries best anticipates eventual posteriors, identifying 37.5% of static surprisals as spurious. We then modify search to avoid these spurious rewards and prioritize hypotheses that remain surprising under non-stationary beliefs. Concretely, we introduce two complementary changes to the original search procedure: belief-update filtering and diversity maximization. Across five discovery domains, our method increases accumulated non-stationary surprisal by 30.62% on average compared to the original search procedure, demonstrating that continual scientific discovery with LLMs requires not only better belief measurement but also search procedures that avoid redundancy and encourage diversity.

When does retention matter for memory-augmented LLM agents? We study this with TraceRetain, a lightweight framework for bounded external memory in frozen LLM agents that scores entries by interpretable features (success, age, access frequency, redundancy, specificity, similarity, downstream utility) and evicts the lowest-scoring ones at capacity. On clean ALFWorld with gpt-5-mini, external memory robustly improves over no memory across two seeds, but differences among bounded retention policies fall within Wilson 95% CIs: clean ALFWorld at T=100 to T=200 does not naturally exhibit the memory pollution retention is designed to address. Under a controlled noisy-write stress (75% synthetic distractors), unbounded memory and FIFO-K50 degrade on Precision@5 (20.2% to 12.4% and 15.8% to 3.8%) while TraceRetain-CEM is essentially unchanged (16.9% to 16.6%) and preserves 97/100 task success. The mechanism: unbounded memory has the highest mean similarity (0.87) but lowest precision, indicating failed distractors close to the query in embedding space. Held-out in-distribution evaluation shows memory-augmented policies solving 47 to 49 of 50 tasks vs. 39/50 for no memory. Bounded retention buys memory and step efficiency on saturated clean benchmarks at no task-success cost, and only differentiates from cache heuristics when streams contain noise.