Papers

Spin in clinical trials includes reporting practices that distort the presentation of results. This is particularly critical in medicine, where spin is present in more than 50% of randomized controlled trials that fail to reach statistical significance. The comparison of primary and reported outcomes is crucial for detecting several types of spin, including outcome switching. We used 300 pairs of outcomes labeled with semantic similarity to develop a system for automatic detection of outcome switching. We evaluated baseline text similarity models and open-source LLMs using generated similarity scores and the Youden index to determine the classification threshold. The proposed approach involves prompt engineering, classification based on token probabilities, and majority voting for the final decision. The results on the test set of 2,496 examples with an F1 score of 0.78 and an accuracy of 0.90 outperform baseline text similarity models but trail behind fine-tuned versions of BERT. We used LLMs to generate natural language explanations for the classified instances and manually assessed their quality.

Learning to act in unfamiliar environments requires agents to infer how the world works and revise that understanding as new evidence arrives. Yet limited observations can support multiple world models that explain past interactions but predict different outcomes in unseen states. We introduce Memento 3, building on the Memento series to enable frozen LLM agents to continually learn explicit world models through external memory. The agent maintains a natural-language rulebook as persistent semantic memory, recording revisable hypotheses about environment dynamics while leaving unknown aspects underspecified. It compiles this rulebook into executable code for prediction and planning. Through a continual loop of observation, reflection, rule revision, compilation, and verification, the agent uses prediction errors to refine both the rulebook and its code. Updated code is accepted only when the LLM judges it faithful to the rulebook and cell-exact replay reproduces the observed transitions. We investigate this process as a model-based route to recursive self-improvement (RSI): the agent autonomously explores the environment, revises its world model, and uses verified updates to guide subsequent interaction and learning, while the underlying LLM remains fixed. A population extension maintains multiple world models in parallel, sharing interaction evidence and using their predictions to guide exploration. On ARC-AGI-3, the single-model agent clears every level of all 25 public games, achieves a mean Relative Human Action Efficiency (RHAE) of 100.0, and uses 44% of the human action count. In an Atari Pong case study, a learned feedback controller wins 21:0 in each of three evaluated episodes with different openings, without further LLM calls.

Byte-level language models such as the Byte Latent Transformer (BLT) group bytes into patches and run their large global model once per patch. BLT starts a patch where a small model's next-byte entropy is high, so global compute goes where the next byte is hard to predict. We show that this rule has a systematic blind spot: positions whose type is predictable but whose value must be computed, such as the number after "=" in a worked math solution. Under tight patch budgets, entropy-triggered layouts skip these positions, and accuracy on them collapses. In Meta's BLT-1B with patch starts on 10% of bytes, the entropy rule puts a patch start at 16% of the computed results in GSM8K solutions and gets 19.0% of them exactly right; a boundary after each "=" at the same patch count gets 51.8%, and entropy combined with a label-free boundary-dependence signal gets 67.1% (default layout at 26% of bytes: 76.8%). The gap survives adapting BLT-1B to the budget with low-rank fine-tuning (32.9% vs 72.7%, three runs per rule, paired p < 1e-200) and grows with model size in byte models trained from scratch at a 10% budget: at 1M, 12M and 50M parameters, boundary dependence beats entropy on final answers by -1.6, +10.1 and +19.8 points, and at 50M it gets 35.9% of computed results against 13.9% (3 seeds each). BLT's entropy-jump rule helps neither target at 50M. The entropy trigger of Scratchpad Patching is likewise indistinguishable from random scratchpads on final answers (5.6% vs 5.9%, 5 seeds), while answer-start scratchpads give 38.1%. The effect is specific to computed values: copies and lookups gain little, and values the model cannot compute gain nothing. Boundary dependence, the rise in the model's own loss when a patch start is removed, measured per two-byte context, finds these positions without labels: combined with entropy it beats the hand-written rule on computed results.

Multimodal depression assessment offers a promising approach to analyzing behavioral patterns associated with depression. However, existing methods often rely on dense and opaque multimodal representations, making it difficult to interpret the behavioral patterns underlying their predictions. In this work, we introduce BehavDep, a sparse factor-based framework that decomposes multimodal behavioral representations into sparse latent factors and associates them with behaviorally meaningful concepts through a semantic bridge. To address the mismatch between user-level annotations and heterogeneous video-level behaviors, BehavDep further learns video-level depression tendency scores under weak supervision and aggregates information across multiple observations for user-level assessment. Extensive experiments demonstrate that BehavDep achieves the best overall assessment performance while revealing complementary modality contributions, heterogeneous behavioral patterns across observations, and prediction responses to concept-level editing. These results show that BehavDep provides a structured and interpretable approach to analyzing multimodal behavioral representations for depression assessment.

Long-term personalization requires language models to use interaction history to track users' preferences across sessions. Parametric memory encodes this interaction history into model parameters or adapters, reducing the need to include it in the inference context. However, independent context compilation leaves cross-session integration unspecified, while recurrent updates can attenuate earlier evidence. To address these challenges, we propose Dual-Path Parametric Memory (DPPM). Its Evidence path directly pools representations of the interaction history to preserve earlier evidence, while its Delta path sequentially updates an associative state to capture changes. Fusing both outputs produces history-conditioned LoRA adapters that combine evidence accumulation with ordered revision. Across multiple backbones, DPPM outperforms the evaluated baselines, achieving 54.22% on PersonaMem-v2 and 86.79% on PrefEval. These results suggest that DPPM provides a simple and effective design choice for cross-session personalized parametric memory.

Sparse Mixture of Experts (MoE) models scale more efficiently than dense models by routing tokens to modular expert networks that are only active for processing a fraction of tokens. A leading hypothesis for the performance of MoE models is that each expert specialises in a single, coherent domain. However, interpretability efforts that assume this hypothesis have generally been unsuccessful. We propose and present evidence for an alternative account that we call the Superposed Specialisation Hypothesis (SSH): experts specialise in a disjoint union of fine-grained features rather than one broad domain. Leveraging the SSH, we introduce RouterInterp, a method for interpreting expert routing that identifies Sparse Autoencoder features most predictive of routing decisions and produces unified natural language explanations. On gpt-oss-20b, RouterInterp explains expert routing with ${\sim}65\%$ higher detection accuracy than prior token statistics based methods. This work provides a scalable method for generating more accurate explanations of expert routing and increases our understanding of a previously uninterpretable component of foundation models.

Safety evaluations of large language models commonly summarize harmful-output behavior with attack success rate (ASR). Yet the same non-harmful outcome can arise for very different reasons. A model may recover a harmful task and refuse it, fail to recover the task, or respond to something else entirely. Distinguishing these cases becomes especially important under intent-obscuring prompts, where a low ASR does not reveal whether the evaluated task was actually engaged. To make this distinction explicit, we pair ASR with operative understanding rate (UR), which measures whether a response both identifies the evaluated task and treats it as the task to be answered. Across interfaces, this paired view reveals substantial variation hidden by ASR: similar ASR values can correspond to sharply different recovery rates. Controlled English reconstructions show that recovery consistently improves as compressed prompts become more explicit, whereas ASR does not follow the same pattern. A complementary contrast comes from FormalLogic, where high recovery can still coincide with frequent harmful assistance. Together, these results show that non-harmful outcomes are not equally informative about model safety, motivating the joint reporting of intent recovery and ASR in LLM safety evaluation. Code and experiment inputs are available at https://github.com/kevinjiang0121-cyber/IRIS.

Large Language Models (LLMs) increasingly ship with explicit "thinking modes", yet their counterpart, "no-thinking", has received far less attention. We study LLMs' no-thinking behavior along two axes. a. How to measure no-thinking? Prior work typically defines no-thinking through proxies such as a disabled thinking mode or the absence of long traces. These proxies are unreliable: disabled thinking modes may still emit reasoning, while long traces may contain filler rather than genuine inference. We instead normalize each response into a pre-answer trace and final answer, and evaluate it at three levels: (i) Empty-Thinking Rate for strict answer-only compliance; (ii) instruction-aware Question-Pre-answer Relevance for similarity between the question and pre-answer trace; and (iii) LLM-as-judge Explicit Inference Rate for visible explicit inference. Together, these metrics distinguish answer-only output, relevant but non-inferential text, and explicit inference. b. How does no-thinking vary across tasks and models? We evaluate six prompting interventions on six LLMs across Boolean, multiple-choice, and open-ended questions. We find that explicit no-think controls cannot reliably eliminate visible inference. Models instead exhibit "Thinking Inertia": explicit inference persists even under strict controls and becomes more prevalent as the answer space opens. Accuracy remains stable on Boolean and multiple-choice tasks, whereas open-ended tasks reveal a trade-off between answer-only compliance and task accuracy. Rewriting the same questions across answer spaces shows that supplying candidate answers makes answer-only responses easier to produce. These findings establish no-thinking as a non-trivial capability: stopping explicit reasoning cannot be assumed from model settings or instructions alone and deserves systematic evaluation alongside reasoning ability.

We describe a 4-tensor attention model that predicts the next semantic state of a scene, for video generation and robot planning. A window of states has positions (x, t) and two fibers, a semantic fiber and a temporal-context fiber, and one softmax normalizes attention jointly over the window. Frames and an agent's situation are written as those states; the encoder, the renderer, and the planner remain outside the update. To test the update on its own, we train on ROCStories, where each window poses the same next-sentence task at the semantic layer. On the validation split, with one seed per setting, the last-sentence cross-entropy on the three matched settings is lower for the 4-tensor model than for a free-running one-dimensional transformer by 5.3% at H=2, L=2, by 2.6% at H=4, L=2, and by 2.4% at H=4, L=3. At H=4, L=2 the parameter counts are nearly the same, 172.5M and 175.9M. On the same two GPUs that 4-tensor run finished in 2.4 hours and the baseline run in 45.2 hours; the baseline is trained by free-running decoding, one sequential forward pass per target token.

Hypernetworks that map a context directly to a LoRA adapter let a large language model carry that context in its weights, but prior work has demonstrated them only on base models of up to 14 billion parameters. We present the Internalizer, a state-of-the-art, portable Context-to-Parameter Mapping hypernetwork that generates document-specific LoRA adapters for the frozen 284B-parameter DeepSeek v4 Flash, a target two orders of magnitude larger than in any previous work. Most of its parameters live in a model-agnostic trunk with only thin entry and exit layers per base model, so it trains cheaply against small models before being ported to the large one. On unseen documents of up to 4096 tokens, the generated adapters reach 84.9% top-1 and 97.8% top-5 teacher-forced accuracy against 63.4% and 83.5% for the base model, with nothing in the context window but a three-word instruction. Once the hypernetwork is trained, a single forward pass turns any document into an adapter for such a model, which could be served alone for speed or alongside the document in the window to raise accuracy further.

This study addresses the problem of structured sentiment analysis, whose goal is to obtain a fine-grained sentiment graph where the nodes represent spans of sentiment holders, targets, and expressions, while the arcs define the relationships among them. Our proposed approach casts the task as dependency graph parsing, but departs from traditional parsing methods by solving it through sequence labeling. To do so, we leverage recent advances in linearized graph encodings that allow each word in the input to be assigned a label, effectively capturing the structure of the dependency graph. We conducted experiments on seven datasets spanning five languages (English, Spanish, Norwegian, Basque, and Catalan), showing performance competitive with leading, more complex single-model approaches.

Verifier scores now serve as both benchmark metrics and training rewards for large language model (LLM) agents, and a change in score is routinely read as a change in capability. It may instead reflect a change in the evaluation. We introduce TRACE, a protocol that turns a score change from a verdict into a testable diagnosis: it applies a targeted change to one part of an evaluation, compares paired runs, checks whether the agent's behavior changed, and rescores unchanged trajectories to test whether the scoring rule is responsible. In a controlled suite of 25 synthetic tasks, renaming tools lowers a scripted agent's score by 0.250 even though it performs exactly the same operations; restoring the original names at scoring time closes the entire gap, while the same mutation exposes a genuine behavioral failure in a second agent. On public $\tau^2$-bench tasks with four LLM agents, an initial 30-task study finds mixed reward changes whose one clear effect does not replicate. In a larger follow-up on 88 new tasks with repeated runs per condition, renaming tools or reformatting tool outputs leaves reward unchanged to within $\pm$0.10 for seven of eight agent-change pairs, whereas tool names that deliberately mislead lower every agent's reward by 0.20-0.44, showing that the setup can detect real effects. Identical reruns flip 15-36% of task outcomes, so single-run comparisons cannot separate presentation effects from run-to-run variation. Two frontier LLM judges give consistent verdicts when a fixed trajectory is presented differently, yet disagree with each other on 57% of the same records, largely because one grades procedure rather than outcome. TRACE thus separates what a score change says about the agent from what it says about the measurement.

On-policy distillation (OPD) supervises student-generated trajectories with token-level teacher signals. Its sampled-token variant avoids the cost of full-vocabulary probabilities. Yet we find that training on fewer tokens can outperform full-token OPD, challenging the intuition that more supervision improves learning. This motivates selecting tokens by learning value. Existing disagreement-based criteria ignore probability scale: tokens assigned negligible probability by both models, termed low-low tokens, can receive large log-ratio rewards and hinder learning. We propose DIAL-OPD, a token-selection method that bridges log-probability and probability spaces by weighting reward magnitude with the logarithmic mean of teacher and student probabilities. A parameter beta controls this weighting, and the highest-scoring tokens are retained. Across 4 teacher-student pairs and 7 mathematical reasoning benchmarks, we compare DIAL-OPD with 9 baselines. Retaining only 40% of tokens, it outperforms Vanilla OPD and its full-token variants, with mean accuracy gains reaching 5.25 percentage points over Vanilla OPD, and doubles AIME25 Pass@16 from 13.33% to 26.67%. It also achieves up to an 18% relative improvement in mean accuracy over the strongest token-selection baseline at matched retention ratios. With a 4B teacher, DIAL-OPD surpasses the strongest full-token baseline using an 8B teacher at both student scales, showing that effective supervision allocation can outweigh teacher scaling. Further analysis shows that moderate beta balances suppressing low-low tokens against preserving useful disagreements. Token-level evidence reveals that DIAL-OPD filters high-reward tokens with limited reasoning value while preserving supervision critical to reasoning correctness.

Improving a long-horizon LLM agent means evolving the harness around a frozen model or training its weights. We let a self-evolving harness make the system stronger first, then cross seed and evolved harnesses with base and trained weights to learn which gains the trained model keeps and which still need the runtime. We show that the right lever can be read off the agent's failure composition: labelling failed trajectories by the first signal that fires separates process failures (blocked calls, loops, exhausted step budgets) from content failures (a delivered plan that is poor). Harness evolution repairs the former, the behaviour it instils can be trained into the weights, and content failures are what weight training is for. On DeepPlanning, a self-evolving harness loop lifts the held-out score of Qwen3.5-4B from 0.16 to 0.30 and of Qwen3.5-9B from 0.32 to 0.44; for 4B, held-out delivery rises from 55% to 90% while content failures are left for the weights. LoRA adapters trained on evolved-harness trajectories internalise the gain: under the original harness they add +0.13 on held-out tasks for both sizes; on 4B they stack with the harness to more than double the held-out score, and on 9B the adapter alone matches the full evolution line, cutting content failures from a quarter of trajectories to one in twenty. A placebo adapter trained on answer-shuffled trajectories falls below the base model. The loop transfers to WebArena-Lite (+0.09 on 117 unseen tasks), where the gain lives in what the model sees and adapters do not add to it. The result is a diagnose-then-intervene rule applied twice: read the failure composition to choose between harness and weights, then read what the accepted edits changed to decide which gains to train in. Scores are four-rollout means against fresh anchors, same-night except where marked, across eight models from six families and two benchmarks.

German is a morphologically rich language whose syllable structure is exceptionally well-predicted by the Knuth–Liang hyphenation algorithm. We ask whether phonologically informed tokenization can serve as a competitive target for end-to-end speech recognition. We compare three tokenizer families on the Omnilingual ASR wav2vec 2.0 backbone fine-tuned with CTC: the pretrained multilingual character inventory, a data-driven Byte-Pair Encoding (BPE) over orthography, and phonologically informed units from Pyphen syllabification and grapheme-to-phoneme conversion. Across 40 fine-tunes, we evaluate on three German test sets spanning orthogonal shifts: in-domain read speech, dialectal spontaneous speech, and standard-German spontaneous speech. In-domain, all phonologically informed tokenizers match BPE and the multilingual character baseline on both WER and CER. Under domain shift the picture splits along vocabulary size rather than the linguistic axis of variation: at small vocabularies, syllable-aware tokenization improves on dialectal speech, where phonetic surface forms vary but syllable structure is preserved, and stays ahead on spontaneous speech, where new word-forms violate vocabulary closure. A phoneme-level confusion analysis further shows that all tokenizers commit the same canonical function-word errors, indicating that the acoustic encoder, not the tokenizer, dominates the error topology. Our findings suggest that tokenizer choice may depend on the vocabulary budget as much as on the distribution shift expected at deployment rather than reducing to a single universal optimum.

Evaluating user experience (UX) with automated computational methods has gained increasing attention, supported by empirical evidence from UXBench. However, binary preference prediction provides limited insight, while relying on a single user-agnostic reward model overlooks the inherent heterogeneity of users, whose expectations can differ substantially. In this paper, we present UXBench Pro, comprising 1{,}000 test instances derived from real user interactions across 12 task scenarios and 82 domains. Each instance is paired with a FACTORS user profile that characterizes the user through seven interpretable behavioral facets, differentiating user groups. To provide richer evaluation insights, we introduce a dual-perspective paradigm that combines a personalized User Reward Model (URM) for third-person judgment with Sim4Eval, a user simulator that enables multi-turn interactions and provides first-person evaluation across four cognitive state dimensions. To assess the reliability of these based evaluators, we further introduce two meta-benchmarks, URMBench and USimBench, that evaluate how faithfully they reproduce real human preferences and behaviors. Extensive experiments reveal seven key findings that highlight the importance of user modeling and multi-perspective evaluation, offering a fresh perspective on user-centric benchmarking and motivating personalized model optimization.

Journals and conferences have begun to screen submitted manuscripts for text written using large language models (LLMs). The reliability of this screening rests on benchmark evaluations against a fixed set of LLM versions, while the versions in actual use keep changing. Here we quantify how this LLM turnover affects the screening of scientific manuscripts. We paired 4,000 pre-ChatGPT abstracts from the Proceedings of the National Academy of Sciences with their rewrites by 23 LLM versions from three vendors, released between June 2023 and August 2026. We then trained detectors under maintenance scenarios ranging from a detector retrained on every new version to one trained once and never updated. Detectors trained only on a vendor's past versions can collapse at the boundaries between model generations: calibrated to falsely flag 1% of human-written abstracts, they catch above 99% of rewrites just before the sharpest boundary and 3.8% just after it. Detectors trained on later versions can also miss rewrites of earlier ones. Vocabulary differences between versions largely track where detection transfers and where it fails. In the two screening scenarios we simulated, screens covering all 23 versions either flagged one in eight human-written abstracts or missed one in three rewrites of the newest version. Indeed, a commercial detector missed most rewrites of the version just after the sharpest boundary while flagging almost no human-written abstracts. Research-integrity policy should therefore treat the benchmark accuracy of a detector as provisional, to be re-verified with every LLM release, including earlier versions.

Current frontier LLMs can theoretically process long contexts with 1M tokens or more. But to what extent can they go beyond simple retrieval and perform deeper reasoning over such long contexts? We empirically investigate long-horizon reasoning capabilities of LLMs, focusing on deductive logic expressed in Prolog. We construct ProloNg, a synthetic testbed to probe Prolog Long Reasoning, which systematically varies the complexity (reasoning depth) of problems, where the hardest case has a reasoning depth of 22 and 62k context length. We study 8 reasoning models across 5 families of frontier LLMs, and find that performance degrades substantially as reasoning depth grows, with the majority of models approaching chance beyond depth 10.

Existing evaluation methods for healthcare LLMs primarily assess factual correctness,safety, and fluency, while providing limited insight into whether generated interactions reflect culturally situated healthcare reasoning. This limitation is particularly important in maternal health, where care decisions are shaped by social and relational norms. We introduce MH-INDIC, a culturally grounded evaluation framework for maternal-health interactions in urban and semi-urban North Indian contexts that operationalises cultural behaviour through ten dimensions of maternal-health reasoning. Using a 26-item survey administered to 102 pregnant and postpartum women from urban and semi-urban North India, we evaluate ten LLMs. We distinguish population level cultural alignment from profile-level behavioural variation. Although several models approximate the human population-level distribution, all evaluated systems exhibit substantially lower variation across demographic and household profiles than the human cohort, revealing a gap between aggregate alignment and profile-conditioned sensitivity. As a downstream application of MH-INDIC, we use the strongest-aligned proprietary and open-source models to generate culturally conditioned maternal-health dialogues under zero-shot, self-conditioned, and human-grounded prompting. Human-grounded conditioning produces stronger profile alignment and dialogue quality ratings, suggesting that measured cultural profiles can improve the cultural grounding of generated interactions

Recent advances in large language model (LLM) pretraining highlight the role of high-quality training data in improving performance. While model-based filtering has proven effective in selecting high-quality subsets from web-scale corpora, especially for high-resource languages, low-resource languages face challenges due to limited availability of annotated data. This work explores extending quality filtering to over 100 languages by proposing a multilingual adaptation approach that converts an existing English quality classifier into a multilingual variant. Our approach proposes training a small multi-layer perceptron on top of Transformer encoder-only model embeddings, using multilingual text as input and scores obtained from English classifiers applied to machine-translated text as labels. Our 1B, 3B and 8B scale experiments show that our approach maintains the downstream LLM benchmark performance of existing multilingual model-based filtering baselines, without harming regional and cultural knowledge benchmarks. To further evaluate cross-lingual generalization, we compare classifier scores of high-quality synthetic data and web samples, and the correlation of classifier scores with LLM-based ones, revealing that the classifier can learn the scoring criteria of its original English variant, even for languages not included in its training data.

Historical pull requests record the design decisions, compatibility constraints, and implementation patterns behind a codebase's current state. Experience relevant to a new task can span related changes whose descriptions emphasize different concerns. We introduce Chronos, a test-time framework that makes this connected history available to large language model (LLM)-based code agents. Chronos distills merged pull requests into structured experience cards and connects them through a typed graph of code-level, developer-intent, and organizational relations. Semantic search identifies entry cards, and weighted multi-hop expansion retrieves connected changes for selective reading. The same memory guides candidate generation and patch selection: a patch-focused change agent and a validation-strategy agent each develop a patch, and an evolution steward consults history to select between them. On SWE-Bench Verified, the full workflow improves SWE-Agent across all six evaluated LLM backbones, raising the mean resolution rate from 69.2% to 72.9% and reaching 79.8% with MiniMax M2.5. With the same backbone, it raises resolution rates from 48.3% to 51.7% on SWE-Bench Pro and from 41.0% to 43.5% on FEA-Bench Lite. Both experience-guided single-agent variants also outperform the base agent. In a human evaluation on 100 tasks with ten cards retrieved per task, graph-grounded retrieval increases the mean number of useful cards from 1.24 to 2.87 over flat semantic retrieval. These results demonstrate the value of PR relations for retrieving useful repository experience and of the evaluated workflows for applying that experience during patch generation and selection.

Sparse Mixture-of-Experts models scale parameter capacity efficiently while maintaining a fixed compute budget per token. However, traditional training paradigms enforce a static choice of top-$k$ experts, which converts a continuous routing distribution into a rigid step function. This constraint introduces a brittle boundary where highly competitive experts are arbitrarily separated into full-supervision and zero-feedback zones based on minor score fluctuations. To address this issue, we propose Elastic Expert Routing, which stochastically samples the active expert budget from a localized discrete distribution centered at $k$. Over multiple training iterations, this mechanism softens the sharp threshold into a gradual probability distribution. Because the sampling neighborhood remains symmetric, this approach matches the expected computational cost of deterministic training, while preserving the inference budget. During supervised fine-tuning, elastic routing improves downstream macro-averages on OLMoE-1B-7B and Qwen3-30B-A3B by $+0.84$ and $+2.02$ points, respectively. In addition, in from-scratch pretraining, it reduces validation LM loss by up to $0.04$ and outperforms the static top-$k$ baseline by $1.6$ points on average across downstream tasks. Comparisons on Qwen3 fine-tuning and from-scratch pretraining further show higher average scores than Noisy top-$k$, expert FFN dropout, and DefaultMoE.

Existing Open-Ended Deep Research (OEDR) systems primarily generate reports from scratch, making them inefficient for scenarios where research reports need to be continuously maintained as new information emerges. We introduce Incremental Open-Ended Deep Research (Incremental-OEDR), a research setting that treats a report as an evolving research state and incrementally updates it by preserving valid knowledge, revising outdated or incomplete content, and incorporating newly available information. To support this setting, we propose Structured Harness, which represents reports as structured collections of outlines, sections, and supporting evidence, and provides structured retrieval, a persistent structured evidence pool, and structured generation for selective report updating and evidence reuse. We further establish a temporal evaluation framework spanning ten years, with Single-Step Task and Long-Chain Task to evaluate incremental updates over both individual transitions and long-term update chains. Extensive Experiments on DeepResearch Bench and DeepConsult under both the Open-source Configuration (OC) and Proprietary Configuration (PC) show that Incremental-OEDR maintains competitive report quality while substantially improving report continuity and reducing research costs. As shown in Figure~\ref{fig:profile}, it achieves up to 0.51 higher content-level ROUGE-L F1, 0.63 higher outline-level EM F1, 33% lower token consumption, and 61% fewer search calls than OEDR on DeepResearch Bench. For more details, please refer to our project page: https://ioedr-project.github.io/.

Coding assistants such as Claude Code and Codex have become a major application of LLM agents, yet existing benchmarks remain far from real-world use, particularly in task horizon and interaction length. Code assistants require completing long chains of development work in continuously evolving repositories, while repeatedly clarifying requirements and adapting implementations through multi-turn interaction. To address these gaps, we introduce SWE-Journey, a benchmark for more realistic evaluation of coding assistants. To address the task-horizon gap, we propose a weak-to-strong synthesis pipeline that automatically constructs long-horizon coding tasks. To address the interaction gap, we mine four representative user personas from real interaction data and build a user-simulation agent to reproduce realistic code-assistance interactions. On average, models pass over 75% of tests for requested functionality with software architects, but fewer than 25% with non-coders. These results show that current coding assistants still fall short of enabling reliable coding for non-coders. We further analyze the reasons for this gap and identify asking right, finding right, and fixing right as key capabilities during interaction.

While predicting prosody from text is an established task in the field, the opposite direction, predicting text that fits a given prosodic pattern, remains largely overlooked. We find this unfortunate, because this opposite direction could lead to some very interesting use cases. Therefore, in this paper, we make the first steps in the prosody-to-text direction by inves- tigating how much of the original sentence can be recovered from its prosodic pattern. To this end, we fine-tune the Whis- per model using only the 12 lowest Mel bins (low-pass filter with approximately 450Hz cutoff), and obtain surprisingly accurate results (WER 36%), with 10% of utterances be- ing recovered perfectly, and 40% of utterances having Word Error Rate at or below 25%. We also find that, given the correct prefix, the next token was predicted correctly in 79% of cases. Our results suggest that the relationship between low-frequency speech features and lexical content is much stronger than previously thought, and we believe that direct- ing more attention to this topic might open the door to new applications, such as using prosody to guide text generation of modern LLMs

Functional Web generation is increasingly optimized with executable rewards, yet existing methods largely focus on the quality of the final page and leave the process of diagnosing and repairing imperfect implementations underexplored. We identify a central challenge in this setting: the Generator and Refiner produce executable artifacts with direct environment rewards, whereas the intermediate Critic influences downstream behavior without a directly executable outcome. We introduce WebLoop, an execution-grounded framework that jointly learns generation, critique, and refinement within a shared policy. WebLoop trains an execution-free Critic with complementary signals for requirement-level discriminability and downstream helpfulness, first establishing reliable diagnosis and then introducing consequence-aware credit, while all three roles are jointly optimized with group-relative policy learning. With Qwen3.5-9B, WebLoop reaches 41.5 Overall on WebRise and 38.9% accuracy on WebGen-Bench, improving the base model by 11.3 and 15.4 points, respectively. The gains transfer to first-pass generation, persist at 27B scale, and generalize from text-only training to multimodal inputs. Controlled analyses further show that the improvement cannot be explained by an additional refinement pass alone, highlighting the importance of learning the Critic and the loop itself.

Multi-agent LLM systems negotiating with a stateful counterpart waste model calls in three ways: polite loops that never meet the counterpart's hidden acceptance condition, malformed outputs that trigger retries, and compliance deadlocks in which the counterpart demands something the agent must refuse. We study a three-part control stack - a 5-Pillar runtime constitution, a 4-tier swarm (Director, three-agent majority vote, Monitor, schema hard gate) and Cognitive Annealing (deterministic deadlock detection, atomic purge of the agent-side context, a canonical recovery message) - against a released adversarial Gatekeeper whose acceptance rules are fixed regular expressions and whose LLM only renders reply text. The testbed has a known solution: it measures whether the stack executes a constitution-aligned strategy against swarm drift and recovers from deadlock, not whether it discovers anything. In five runs per configuration (30 runs; Gemini 2.5 Pro agents, Claude Haiku 4.5 Gatekeeper) we find: (i) the constitution and Director make an acceptable framing possible but not reliable - 0/5 baseline unlocks versus 1/5 and 2/5 with the constitution; when the swarm unlocks it does so in one turn with 7-8 calls and about 15k tokens (67-73% below baseline); when it does not, it costs 17-38% more; (ii) the Monitor and hard gate do not reduce unlocks and leave an audit trail; (iii) under a honeytrap-to-compliance deadlock, LLM-only steering escapes 0 of 5 times while atomic purge plus a canonical strike escapes 5 of 5 (Fisher $p = 0.008$) at the same call budget, with zero calls for the strike. LLM-written strikes failed the deterministic pre-flight 5 of 5 times although an LLM Monitor had approved four. Pre-registered hypotheses on average call and token reduction were not supported. Cost is bounded in every arm by deterministic stop rules; the stack adds recovery at no extra model cost.

End-to-end speech parsing, a task recently proposed, consists in predicting both the transcription and the syntactic tree for a spoken utterance. Existing architectures for speech parsing often utilise intermediate neural networks. In this work, we examine the effectiveness of intermediate neural networks (NN) for parsing, and, specifically, what role do they play. We introduce a simpler end-to-end architecture for speech parsing, where we remove these intermediate NN units, reducing the parameters by 12%, while achieving comparable or better performance than prior method on both automatic speech recognition (ASR) and parsing. We demonstrate that intermediate NN units help reduce the representational gap when the pre-trained encoder is frozen. We do a comprehensive evaluation of speech parsing on French, and medium-low resource languages Slovenian and Naija. We further investigate the impact of the training data size and intermediate layers of the pretrained speech encoder on speech parsing.

Reinforcement learning with verifiable rewards (RLVR) and on-policy distillation (OPD) have become two main paradigms for post-training reasoning models. RLVR gives each response a single outcome label, leaving the steps inside it without separate credit. OPD provides token-level guidance at student-visited prefixes, but its pointwise signal does not directly reflect the pattern of teacher–student disagreement across the vocabulary. Dense, unbounded log-ratio supervision can amplify the teacher's influence, yet a strong solver is not necessarily a suitable guide when the student's solution paths depart from the teacher's. We propose Residual Advantage (\RA{}), which treats the teacher–student probability residual as a bounded one-step reward, subtracts the corresponding state value under the student policy to form a standard advantage, and centers the result within each response before adding it to the verifier advantage. The guidance term has zero mean within each response, so the verifier advantage remains the response's mean label and the teacher only redistributes credit among the steps within it. \CoRA{} further updates a teacher LoRA with verifier advantages on the same scored student batch and uses the updated teacher in the next iteration's residual, adapting guidance to the student's attempts. With Qwen3-1.7B-Base and Qwen3-4B-Base students and a Qwen3-8B teacher, \RA{} combined with GRPO or REINFORCE++ improves the underlying sequence-advantage algorithm in all 24 comparisons on three mathematical benchmarks, raising macro Avg@8 by 1.7–3.6 points and Pass@8 by 3.9–6.3 points. Both combinations surpass teacher-only OPD, and \CoRA{} adds a further 1.0–1.5 Avg@8 points.

Large language models (LLMs) often default to Modern Standard Arabic (MSA) when generating Arabic, even when prompted with dialectal Arabic. A natural explanation is that their internal representations are dominated by MSA. We test this hypothesis by adapting the language-dominance framework of Shani and Basirat (2025) (https://doi.org/10.18653/v1/2025.blackboxnlp-1.7) to 26 Arabic varieties. Across layers and model families, we find no evidence that MSA acts as a dominant internal representation for Arabic dialects. Instead, dialect representations form a dense and highly overlapping space: normalized mutual information drops sharply compared to patterns reported for more distinct languages. Moreover, the strongest separability effects are not confined to intermediate layers, but can shift toward later layers depending on the architecture. These findings challenge a common interpretation of MSA-biased generation: output preference does not necessarily reveal internal representational dominance. Analyses of multilingual and dialectal LLMs should therefore distinguish generation bias from the geometry of internal representations.