Shiyu Ji, Yixuan Wang, Yijun Liu, Qingfu Zhu, Wanxiang Che
The increasing memory demand of the Key-Value (KV) cache poses a significant bottleneck for Large Language Models (LLMs) in long-context applications. Existing low-rank KV compression methods reduce this footprint by modifying model projections, limiting the flexibility to switch back to standard full-cache inference when sufficient memory is available. In this paper, we propose EchoKV, a flexible KV cache compression framework that supports on-demand transitions from full KV caching to compressed caching. Unlike traditional compression-decompression paradigms, EchoKV utilizes a lightweight network to reconstruct the discarded KV components from a partial subset, exploiting intrinsic inter-layer and intra-layer similarities among attention heads. We further introduce a lightweight two-stage fine-tuning strategy, requiring only a few minutes on a single A100 GPU for Llama-3.1-8B. Experimental results on LongBench and RULER demonstrate that EchoKV offers a competitive quality-cache trade-off across multiple compression ratios and backbone models while preserving the throughput of full-cache inference in short-context scenarios.
Chaewan Chun, Meruyert Aristombayeva, Jiyoung Choi, Mahjabin Nahar, Delvin Ce Zhang, Dongwon Lee
Modern misinformation is often heard before it is read, yet fact-checking systems are still evaluated mainly on clean written claims. Spoken dialogue remains different even when systems operate on transcripts: claims may be distributed across speakers and turns, depend on prior context, and become harder to verify when Automatic Speech Recognition (ASR) errors distort the available text. Prior spoken dialogue fact-checking resources are small, English-centric, or focused on annotation rather than end-to-end benchmarking, leaving no large multilingual benchmark with paired speech and turn-level labels. We introduce TRILOGUE (TRIlingual spoken diaLOGUE fact-checking), a large-scale trilingual benchmark of source-grounded spoken dialogues in English, Russian, and Kazakh. It contains nearly 12K dialogues, 187K turns, and 390 hours of paired audio with ASR transcripts and word-level timestamp alignments across all three languages, including nearly 5K human-recorded Russian and Kazakh dialogue files. TRILOGUE supports claim check-worthiness detection, source-article evidence retrieval, and claim verification with claim-only, gold-evidence, and retrieved-evidence inputs. Baselines show that ASR degradation and cross-lingual transfer remain challenging, especially for Kazakh, while retrieved source evidence substantially narrows the gap to gold-evidence verification.
Reinforcement learning with verifiable rewards (RLVR) has become an effective approach for improving mathematical reasoning in large language models. However, a fundamental question remains: when a model fails to solve a problem, does it truly lack the underlying capability? In this work, we reveal that direct failure does not necessarily imply inability. Some reasoning failures can be recovered by exploring the model's own trajectories, exposing a recoverable capability frontier: a policy-dependent region where solutions are not reliably generated through direct decoding but can be elicited through trajectory-level search. Based on this observation, we introduce ReSI-Math, a framework that explores recoverable capability frontiers and converts recovered capabilities into data-efficient RLVR supervision. ReSI-Math uses Self-Judge-guided evolutionary search to identify recoverable failures under a fixed policy and adaptively selects frontier problems for sequential RLVR as the policy evolves. On Qwen3-VL-8B, correct-incorrect trajectory crossover produces 272 verifier-confirmed correct offspring from 621 attempts (43.8%). After three sequential one-shot RLVR stages, the MATH500 and WE-MATH average improves from 78.71 to 80.16. Under matched single-problem RLVR, random selection reaches only 53.31 versus 54.16 for frontier-guided Q1.
Meruyert Aristombayeva, Jason S. Lucas, Chaewan Chun, Dongwon Lee
While text-based hallucination detection is well studied, reference-free detection of factual alterations in speech remains underexplored, especially for low-resource languages. Our spoken benchmark comprises 12,013 English, Russian, and Kazakh news samples with three synthetic alteration types and three severity levels, pairing source articles with rewrites as text, synthesized audio, and ASR transcripts. We add 290 fact-checked misinformation items collected in Russian (225) and Kazakh (65), translated into the other language and rendered through the same TTS-ASR pipeline. We evaluate fine-tuned multilingual encoders and zero-shot multimodal decoders on text, transcripts, and audio. Detectors receive only target inputs without source articles or external evidence; the task evaluates reference-free classification rather than evidence-grounded verification. Encoder degradation from source text to transcripts generally tracks per-language ASR error on the binary task. Among decoders, only Gemma-3n exceeds the binary majority-class baseline in macro-F1, on transcripts only; the other four fall below their respective baselines. Comparisons between audio and transcripts are confounded by differences in evaluation coverage and class balance. Synthetic-trained detectors achieve 0.82--0.88 macro-F1 on real-world misinformation source text; Russian provenance analysis reveals veracity-related and model-dependent machine-style signals, a key confound in synthetic hallucination benchmarks.
Qwen3-8B reviewing Llama-3.1-8B in math adds 33.2 points, yet Qwen3-8B alone is 6.4 points more accurate than the result. Reviewing another model's answer is usually judged by its gain over the generator. We show that this confuses verification, which adds accuracy beyond either model alone, with capability transfer, where a stronger reviewer supplies answers it could have given on its own. Measured against the stronger model alone, across 48 pairs of 12 open-weight models (3B to 72B) in math, knowledge, code and reasoning, review beats it in only 9 of 96 review directions, and does worse the larger the capability gap. The reason is anchoring: the reviewer mostly follows the generator, discarding its own correct answers. Review costs 2.6 to 2.8 times as much as the stronger model alone, while a selector using the same two calls is 2.7 to 4.2 points more accurate than review and cheaper. Studies of verification should report the stronger model alone, not only the gain over the generator.
Large language models (LLMs) face significant challenges in processing long contexts due to the linear growth of the key-value (KV) cache and quadratic complexity of self-attention. Existing approaches address these bottlenecks separately: Multi-head Latent Attention (MLA) reduces the KV cache by projecting tokens into a low-dimensional latent space, while sparse attention reduces computation. However, sparse methods cannot operate natively on MLA's compressed latent structure, missing opportunities for joint optimization. In this paper, we propose Latent-Condensed Attention (LCA), which directly condenses context within MLA's latent space, where the representation is disentangled into semantic latent vectors and positional keys. LCA separately aggregates semantic vectors via query-aware pooling and preserves positional keys via anchor selection. This approach jointly reduces both computational cost and KV cache without adding parameters. Beyond MLA, LCA's design is architecture-agnostic and readily extends to other attention mechanisms such as GQA. Theoretically, we prove a length-independent error bound. Experiments show LCA achieves up to 2.5$\times$ prefilling speedup and 90% KV cache reduction at 128K context while maintaining competitive performance.
Hallucinations in Large Language Models (LLMs) remain a major challenge, as they can produce misleading, factually incorrect, or unverifiable information, thereby undermining their trustworthiness and reliability. However, reliable ground-truth labels remain a key challenge in evaluating hallucination detectors. Recent studies often construct labels automatically by comparing model generations with reference answers using similarity metrics such as BLEURT. This implicitly treats BLEURT similarity as a proxy for hallucination, despite its English-centric supervision, unbounded score range, and design as a similarity metric rather than a factuality measure, making threshold-based ground-truth construction potentially unreliable. Through comprehensive experiments using 15 automatic supervision judges, seven hallucination detection baselines, two generator models, two QA datasets, and five languages, we demonstrate that detector performance varies substantially with the choice of supervision judge. Human validation reveals substantial differences in labeling reliability, with QGen-9B achieving the lowest overall misclassification rate (12.86%), while BLEURT and BLEURT-n exhibit rates of 47.23% and 47.32%, respectively. Our findings highlight the limitations of automatic similarity-based labeling and the need for more reliable judging methods that better capture semantic and factual correctness.
Large language models (LLMs) are increasingly expected to extract and synthesize structured information from complex real-world documents, yet they typically produce unstructured, hard-to-verify paragraph-style outputs. We introduce AOE (Arranged and Organized Extraction), a bilingual benchmark that evaluates LLMs' ability to distill long-form, multi-document inputs (64k to 1.8M tokens) into structured tables. AOE comprises 12 tasks across four domains (academic, financial, legal, and medical), each grounded in genuine analytical needs, with schema guidance provided to isolate extraction quality from formatting variance. We evaluate 12 frontier LLMs (both open-weight and proprietary) under four generation settings of increasing autonomy: direct full-document input, RAG-based retrieval, offline document-browsing agent, and online web-search agent. Even the best-performing model achieves only 42.7% Cell F1, and retrieval-augmented approaches often degrade rather than improve performance, revealing that structured extraction from long contexts remains a fundamental challenge - an "Area of Effect" (AOE) that impacts all tested systems.
To prepare for future outbreaks, researchers need to know what studies have been done and what data they use. Mapping a large literature by hand is slow, while large language models alone can misstate or conflate sources. We present EpiLens, a lightweight, open-source retrieval-augmented generation package that applies a fixed retrieval template and output schema to every paper in a corpus and links outputs to source passages. We build a corpus of 2,297 papers from the sources of a multi-pathogen parameter study and evaluate EpiLens's paper-classification workflow against 214 manually labelled papers. Run locally with an open-weight model, so that licensed full texts never leave our hardware, the classifier is comparable to supervised baselines trained on the manual labels, and ahead of them on geography, despite the lack of task-specific training data. The study's early extraction sheets cite published sources for 327 estimates. We link each estimate to a cited document or to corpus studies included by a review. Then, we use the classifier's labels to compare the papers associated with different parameters and pathogens. EpiLens labels more papers linked to reproduction numbers than fatality risks as model-based (53% versus 15%, compared with 62% versus 24% in the manual labels). The gap is consistent with the frequent use of transmission models to estimate reproduction numbers and largely persists within pathogens.
This monograph develops a unified account of training and inference in Power Law Decoder Representation language models (PLDR-LLMs). Exact finite work identities decompose changes in the absolute energy of the row-centered learned map into parameter contributions, signed interactions, and numerical observation defects. Positive affine blocking retains restarts at the row-constant face, while the augmented AdamW state supplies the complete dynamical description. Predictive renormalization acts on the complete conditional training law for a single pass over distinct corpus target blocks, retaining optimizer memory, remaining data, schedule, and numerical policy. Autonomous reductions require closure; approximate reductions carry successor and emission errors. Finite-population covariance, matched physical clocks, matrix fluxes, and signed temporal energy connect row dynamics to model-wide observations. Absolute row collapse, relative row concentration, operator stabilization, and predictive accuracy are distinguished. Experiments reveal observer and optimizer dependence, reject the tested autonomous row-state candidates, and support finite conditional prediction and state-specific operator reduction. Independent single-pass families exhibit moving finite fluctuation regions without establishing a thermodynamic critical class. Conditional symmetry, head limits, covariance flows, and readout error budgets specify assumptions needed to transfer scaling laws to inference. The theory separates exact identities, conditional dynamical claims, and finite empirical findings, with proofs, selected formal checks, and compact numerical evidence.
Identifying the primary diagnosis alone does not fully characterize a model's diagnosis-generation performance. We introduce CardioVascBench, a retrospective benchmark for prioritized diagnosis generation from real-world Chinese cardiovascular inpatient records. It contains 3,262 de-identified cases, with 2,174 for training and 1,088 for evaluation, linking clinical narratives and textual ECG reports to expert-verified, ordered discharge diagnoses standardized to ICD-10. An English translation of the evaluation set supports aligned cross-lingual comparison. Our evaluation protocol combines terminology normalization and rank-aware scoring to separately assess primary-diagnosis identification, reference-diagnosis coverage, and list precision. Evaluation across 36 systems shows that higher primary-diagnosis accuracy does not necessarily imply higher list precision against discharge references. We also provide CardioKG-Agent, a knowledge-enhanced baseline combining entity normalization, constrained graph retrieval, and guideline evidence. In a staged comparison of three baseline models, supervised fine-tuning plus the agent achieves a 73.5% higher mean composite score than fine-tuning alone, representing a system-level gain. These results measure retrospective agreement with recorded diagnoses; agreement between metric rankings and physician judgments remains to be validated. CardioVascBench supports research on diagnosis prioritization and knowledge-enhanced generation.
Kristian Schwethelm, Jonas Kuntzer, Conrad Ketzer, Eva Wende, Moritz Schlager, Daniel Rueckert, Georgios Kaissis, Friederike Jungmann 等 9 位作者
MIMIC is one of the most widely used public hospital datasets for machine learning on electronic health records, yet clinically important information, such as past medical history, allergies, and diagnostic findings, remains confined to free-text discharge summaries. Studies requiring this information must therefore build their own extraction pipelines, duplicating substantial effort across projects. We release MIMIC-Note-Extract (MNE), a large-scale dataset that uses a hybrid extraction pipeline combining deterministic parsing with schema-guided large language model generation to transform 16 sections from 300,720 discharge summaries into structured tables. MNE adds information that has no structured MIMIC counterpart and enriches related tables with broader coverage and greater clinical detail, including 10 times more patient-history entries and medication histories for more than twice as many admissions. Validation by medical doctors confirms extraction accuracy above 92% across sections. Using MNE, we construct a clinical decision-making benchmark for LLM agents, demonstrating how reusable structured representations of clinical narratives reduce data curation effort.
Recent works on language identification and generation have established tight statistical rates at which these tasks can be achieved. These works typically operate under a strong realizability assumption: that the input data is drawn from an unknown distribution necessarily supported on some language in a given collection. In this work, we relax this assumption of realizability entirely, and impose no restrictions on the distribution of the input data. We propose objectives to study both language identification and generation in this more general "agnostic" setup. Across both problems, we obtain novel interesting characterizations and nearly tight rates.
Supervised fine-tuning (SFT) is a key component of language model post-training, enabling effective task adaptation and serving as an important foundation for subsequent reinforcement learning. Recent methods seek to improve SFT by introducing token-level weights into the training objective, where these weights are typically derived from a model's output distribution at each token position. To better understand how such weighting schemes affect SFT, we analyze their optimization objectives and theoretically characterize the target distributions they induce. Our analysis shows that dynamic weighting methods, such as Dynamic Fine-Tuning (DFT), induce a heavily sharpened distribution, whereas static weighting methods fit a target distribution proportional to the product of the data distribution and a frozen prior. These findings suggest that static weighting provides a more stable training target, while its effectiveness critically depends on the quality of the prior distribution. Building on this observation, we propose Self-Privileged Fine-Tuning (SPFT), which uses a frozen copy of the initial model conditioned on privileged information (such as ground truth for mathematical questions) as a prior to provide guidance for token-level weighting during fine-tuning. Through a Bayesian decomposition, we further show that SPFT reweights the standard static prior according to how informative each target token is for predicting the ground-truth outcome, thereby assigning larger weights to more predictive tokens. Experiments on mathematical reasoning and code generation show that SPFT improves over standard SFT and other baselines across the main settings.
Johnny Alali, Nicolás Benjamín Ocampo, Davide Ceolin
Prolonged exposure to hate speech on social media can cause lasting psychological distress among users. Counterspeech has emerged as an effective, non-censorial mitigation strategy, but its deployment requires rigorous quality assessment. Manual assessment is unscalable due to the large volume of content on the Web and imposes severe emotional strain on human annotators. Large Language Models (LLMs) offer a scalable alternative, yet concerns remain regarding their reliability. In this paper, we investigate the extent to which LLMs can effectively perform this task. We test 12 open-weight LLMs of different sizes and families under zero-shot, few-shot, and chain-of-thought prompting, to approximate pointwise and pairwise human expert assessment of counterspeech quality across four dimensions: relevance, coherence, aggressiveness, and suitableness. We find that (i) LLMs show low agreement and moderate rank correlations with human experts: our best configuration, Qwen 2.5 72B on coherence, achieves up to a linear-weighted Cohen's $κ_l = 0.19$ and Spearman's $ρ= 0.42$ in pointwise evaluation and an unweighted Cohen's $κ= 0.25$ in pairwise evaluation. (ii) Converting pointwise scores into pairwise labels substantially improves human--LLM agreement for selected configurations. For same-instance coherence, Qwen 2.5 72B under zero-shot prompting achieves $κ= 0.35$ with converted labels, compared with $κ= 0.25$ when judging pairs directly, and Mixtral 8$\times$7B under few-shot prompting improves from $κ= 0.07$ to $κ= 0.34$. (iii) LLMs agree more with one another than with human experts, particularly in pairwise evaluation, where the strongest model pair, OLMo 2 32B and Llama 3.3 70B on coherence, reaches substantial agreement with $κ= 0.76$, whereas their agreement with the experts in the same configuration does not exceed $κ= 0.21$.
Maxime Toquebiau, Jérôme Botoko Ekila, Paul Van Eecke, Katrien Beuls
Recent studies of language evolution with neural agents often use models where agents can either speak or listen, where there are only two agents in the population, or where agents can share internal representations. These choices prevent the study of how language emerges through cultural evolution, i.e. through the cumulative adaptation of individual behaviours to fit the communicative needs of the community. We tackle this issue with a new model that trains populations of bi-directional neural agents in a fully distributed fashion to solve a referential game. We demonstrate the effectiveness of this approach on multiple image datasets and show that it is able to scale to large populations. We also show that more conventional languages are easier to learn, providing a direct advantage for cultural transmission. Overall, this work provides a novel approach to neural emergent communication that opens up new directions for evolutionary linguistics.
Conspiratorial discourse on social media is not always expressed through explicit claims or stable lexical markers. The same surface content may express endorsement, legitimate concerns, criticism, satire, or mockery. The main challenge is therefore not only recognizing conspiracy-related claims, but inferring the speaker's intent -- the utterance's illocutionary force. We argue that this can be achieved through the use of relevant social contexts and propose an agentic framework, equipped with a set of tools supporting social queries. We demonstrate the benefits of our approach on a unique dataset of Hebrew tweets, covering 80%--90% of the public Hebrew tweets published over a four-year span (late 2018-- early 2023), encompassing several election cycles as well as the COVID pandemic years and related vaccination campaigns. This extensive coverage can be used in recovering different social contexts. Evaluating our framework on a manually-annotated adversarial dataset, we find that context-aware workflows consistently outperform text-only classification and that the agentic framework performs better than other frameworks and settings, with significant gains over a non-agentic model exposed to the same contexts available to the agent. We further provide an analysis of the results, the errors and efficiency (token economy) tradeoffs. These findings support viewing the task of conspiracy detection as a socially embedded interpretation task, in which effective classification depends not only on access to contexts, but also on adaptive reasoning in which the agent uses tools on a per-case basis, asking only for evidence relevant to its current reasoning step.
Javier Garcia Gilabert, Rico Sennrich, André F. T. Martins
Quality-Aware Decoding (QAD) improves machine translation through test-time scaling, generating multiple candidate translations and selecting the final output with a quality-aware decoder. Usually the same candidate budget N is spent on every source sentence, assuming that all inputs require the same amount of search. We show that this assumption does not hold and that the quality gain obtained from additional candidates varies across sentences. We study search utility through three quantities: total scaling gain, how much quality a sentence can gain from a much larger candidate pool; marginal scaling gain, the expected gain from generating one additional batch of candidates; and the sampling saturation point, the candidate budget required to realize most of the total scaling gain. We investigate two sources of information for predicting these quantities: signals available before candidate-set expansion, and statistics computed from the candidate pool accumulated during search. We find that sentence-level signals correlate with total scaling gain but are less informative about the sampling saturation point. In contrast, properties of the current candidate pool provide useful signals of marginal scaling gain. Finally, we show that these signals can identify sentences that obtain larger gains when additional search is selectively applied.
Multi-teacher on-policy distillation (MOPD) distills several domain experts into one student and is used in the post-training of recent large language models. It trains on all domains at once, so every expert must be available at the same time. In practice, capabilities often arrive one after another, and the experts of earlier domains may be gone when a new domain is learned. Distilling the experts in sequence then suffers from catastrophic forgetting of the earlier domains. Some recent pipelines store a snapshot after every stage and end with a multi-teacher recovery stage taught by all snapshots. The number of stored models then grows with the number of stages. Different from these pipelines, this paper formulates continual MOPD by what is kept from earlier stages and studies replay toward the previous student. Inside every stage, prompts of earlier domains are replayed with the student of the previous stage as their teacher, so neither earlier experts nor stage snapshots are kept. Two design choices of replay are explored. The replay budget of a stage is either shared among the earlier domains or fixed per earlier domain. Replay steps either distill the token distribution alone or add alignment of hidden states, region-pooled self-attention scores and value relations. A five-domain benchmark of logic, math, code, safety and instruction following is built from public experts of one 1.5B student. Experiments on its 14 benchmarks show that, without storing any snapshot, replay toward the previous student closes most of the gap between sequential distillation and joint MOPD. With a fixed replay share per earlier domain, it performs on par with recovery from all stored snapshots.
Yeonsu Kwon, Jiho Kim, Junseong Choi, Paloma Rabaey, Minseo Kim, Sujeong Im, Jeewon Yang, Jun-Min Lee 等 13 位作者
Data consistency between unstructured clinical notes and structured tables in Electronic Health Records (EHRs) is essential for patient safety and clinical decision-making. However, existing work on note-table consistency verification mainly relies on surface-level matching of numeric values or simple events. Such approaches fail to capture the reasoning underlying real-world EHR documentation, including clinical interpretation, event relations, and temporal changes. To address this gap, we introduce EHR-ReasonCon, a reasoning-intensive benchmark for note-table consistency verification. Built on MIMIC-III with expert-guided annotations, it comprises 8,042 entities derived from clinical notes and provides high-quality ground-truth labels. The annotation protocol is supported by specialized table-exploration tools to ensure systematic evidence retrieval and reliable consistency assessment. We also propose EHR-INSPECTOR, an LLM-based framework that segments notes, extracts anchor entities and temporal references, and uses table-exploration tools to verify consistency against structured tables. Evaluated using expert-validated LLM-as-a-judge metrics under harsh and lenient criteria, EHR-INSPECTOR achieves state-of-the-art performance across multiple model backbones. Analyses further demonstrate the effectiveness of its components and highlight differences from human verification. The code and dataset are available at https://github.com/dustn1259/EHR-ReasonCon.
Time-series interpretation often requires both numerical observations and textual context. Integrating these inputs into a pretrained language model raises a representation question: how should a signal enter the model's token sequence? We present TEMPO, which extends an LLM's vocabulary with discrete codes from a context-aware Transformer tokenizer using finite scalar quantization. A frozen tokenizer maps four-step patches to a 625-code vocabulary; signal tokens, scale statistics, and text are then processed by the LLM's existing self-attention layers. Adaptation updates the new token embeddings and DoRA adapters. With a Qwen3-4B backbone and task-specific fine-tuning, TEMPO achieves macro-F1 scores of 89.0% on activity recognition, 85.0% on sleep staging, and 86.0% on bearing-fault diagnosis. Time-series question answering and qualitative bearing-diagnosis examples further illustrate how this interface combines signal representations with textual context. Together, these results demonstrate the utility of discrete vocabulary extension for time-series understanding.
Audio language models are meant to answer questions from the audio itself, but an hour-long recording rarely fits in their context, and the answer often depends on a few seconds of it. Transcribing the recording first makes it fit but discards sounds and voices, while systems that return to the audio must be trained for this behavior. We propose ReListen, a training-free method in which a frozen audio language model finds its own evidence before answering. In a Jev-style scan, the model scores each 20-second window by the probability that it helps answer the question, and it then answers from the highest-scoring minute. This minute overlaps the annotated evidence far more often than transcript retrieval, for 88.5% of VoiceGiraffe questions. Without any change, the same scan also grounds events on LAT-Bench as well as a model trained for temporal grounding. Answering from it matches or exceeds hearing the whole recording with a small fraction of the audio, and on VoiceGiraffe it exceeds a transcript of the whole recording by 9.2 points. The selected evidence is also worth hearing rather than reading. Given the same excerpts, reading their transcript matches listening on speech content, but listening is clearly more accurate on questions about sounds and voices. Moreover, ReListen answers faster than every transcript-based system, and its memory stays constant even for recordings too long to hear whole on one GPU. Therefore, a frozen audio language model can search hour-long recordings by itself, without a transcript or additional training.
While conversational recommendation systems provide a more natural user interface, they face the challenge of progressively discovering a user's needs before they can recommend an item that satisfies them. To evaluate the ability of AI agents at this elicitation process, we introduce $τ$-Rec-Books, a benchmark that holds each user request fixed while varying only when the user reveals their requirements: all at the start (volunteer), one at the start and the rest when asked (mixed), or some never (hidden). A hidden requirement can be found only by proposing books and inferring it from the user's accept or reject verdicts. We generate 450 tasks over a catalog of 20,000 books from Wikidata, develop a rule-based simulated user to answer the agent, and score recommendations against catalog metadata without an LLM judge. Across four agents and two search interfaces, success rates fall for every agent and interface, on average from 0.975 (volunteer) to 0.756 (mixed) and 0.283 (hidden). Mixed tasks fail mainly because agents do not ask, and when told to ask, every agent reaches at least 0.931. On hidden tasks, most failures come from agents recommending a book they never proposed to the user. When told to propose first, agents only match a simple baseline that proposes popular books, and stay well below one that uses rejection to rule out possible requirements. We release the catalog pipeline, task generator, simulated user, validators, reference agents, and all agent traces at https://github.com/nbharaths/tau-rec/tree/main/data/tracks/books-wikidata-v2.
He Wang, Yeye Liu, Ruijie Li, Xuemei Jiang, Junyu Wu, Lining Liu Zhuoqi Man, Tao Li, Yanjie Song
Agile Earth observation satellites must coordinate imaging tasks within limited observation windows, while accounting for attitude maneuvering and onboard energy constraints. Competition among observation opportunities makes scheduling performance sensitive to the heuristic rules used for task selection. This letter proposes a large language model (LLM)-assisted adaptive heuristic search method for agile Earth observation satellite scheduling. The LLM generates an initial heuristic rule according to the characteristics of the scheduling instance. During the subsequent ant colony search, search performance and scheduling information are fed back to the LLM to refine the heuristic rule. The proposed modifications are evaluated through paired solution constructions before being incorporated into the search process. A unified scheduling procedure determines task selection, satellite assignment, and observation timing under the operational constraints. Experiments on six scheduling scenarios involving 50--200 imaging tasks demonstrate that the proposed method achieves 4.3\%--9.6\% higher mean observation benefit than the strongest baseline.
Among 291 surveyed open-weight language model repositories (280 distinct embedding tables), 52.6% of untied gated and hybrid checkpoints (20/38) contain reachable input embeddings with norms below $10^{-6}$, compared with 0% of weight-tied models. Such near-zero rows can arise during pretraining when unseen rare or template tokens receive zero input gradients while AdamW weight decay shrinks untied embeddings. In tested Transformer configurations, non-initial positions or a BOS token prevent this failure mode. In gated sequence models (Mamba, Gated DeltaNet, xLSTM), a near-zero token can amplify input-embedding gradients by more than 30 orders of magnitude (up to 35). At zero input, current-token readout factors (SiLU gates or linear queries) vanish, suppressing historical-state readout. Residual RMS values can then fall below the square root of the RMSNorm floor across consecutive layers, causing Pre-RMSNorm to amplify backward gradients by up to $10^3$ per layer. When a gradient entry exceeds $1.84 \times 10^{19}$, FP32 global-norm clipping in PyTorch, Hugging Face Trainer, and FSDP2 can overflow to infinity and zero the current batch's gradients. In Telugu continual pretraining of Nemotron-3-Nano-4B, this discards updates on approximately 20% of steps and leaves triggered rows frozen. We propose two remedies. For released checkpoints, GroupClip decouples input-embedding clipping using a 64-bit or rescaled norm; its rescaled implementation eliminates update loss and heals triggered rows in one step with negligible measured end-to-end overhead. For pretraining, GateFloor ($\mathrm{SiLU}(z) + δ$, $δ\in [0.01, 0.1]$) keeps the gate positive at zero input, reducing near-zero gradients by 28 orders of magnitude in experiments up to 370M parameters, with perplexity comparable to the baseline.
Language models are accurate on most long-arithmetic problems, but their accuracy is estimated on samples and carries no guarantee for an unseen input; tool use is exact but moves the computation outside the network. We show that exactness at all input lengths becomes a finite, checkable property of a network when each of its steps reads only a bounded view of an external working memory. In Sketchpad, Qwen3.5-0.8B with a low-rank adapter executes arithmetic procedures on a rendered sheet. Each forward pass reads a three-cell view and emits one step: a digit, a carry or borrow, and the next operation. A fixed runtime moves the view and writes no digit of a result. We prove in Lean that the model is then exact on all inputs of any length if and only if it is exact on the finite set of views the runtime presents, that no finite test suffices when steps also read earlier steps, and that the guarantee tolerates bounded numerical perturbation and extends to compositions of procedures. All 2,446 views pass on the deployment device, a laptop GPU running the model in a web browser. The certified model performs signed addition and subtraction, long multiplication and division, square roots and greatest common divisors at arbitrary length. It solved 200 of 200 pre-registered problems in the browser, and 220 of 220 sums of up to 1,000 digits and products of up to 20 digits, of which GPT-6.1 Sol, the strongest of ten other models, solved 217. A symbolic runtime built on these procedures solves, expands and factors polynomials with every number computed by the model, and a multiplicative gate on the adapter recovers the base model exactly when the sheet is not in use.
Yuqing Li, Jiangnan Li, Zheng Lin, Ziyan Zhou, Junjie Wu, Jie Zhou, Mo Yu
Humans understand long and complex texts by relying on a holistic semantic representation of the content. This global view helps organize prior knowledge, interpret new information, and integrate evidence dispersed across a document, as revealed by the Mindscape-Aware Capability of humans in psychology. Current Retrieval-Augmented Generation (RAG) systems lack such guidance and therefore struggle with long-context tasks. In this paper, we propose Mindscape-Aware RAG (MiA-RAG), the first framework to formulate mindscape-aware retrieval and generation as a unified conditioning paradigm for LLM-based RAG. MiA-RAG builds a mindscape through hierarchical summarization and conditions both retrieval and generation on this global semantic representation. This enables the retriever to form enriched query embeddings and the generator to reason over retrieved evidence within a coherent global context. We evaluate MiA-RAG across diverse long-context and bilingual benchmarks for evidence-based understanding and global sense-making. It consistently surpasses baselines, and further analysis shows that it aligns local details with a coherent global representation, enabling more human-like long-context retrieval and reasoning.
Existing metalinguistic reasoning benchmarks require both inference of rules and their application, making failures hard to diagnose. We introduce typological feature induction (TFI) which targets the former step. Given sentences in an unnamed language paired with translations in a known pivot language, models must predict a typological feature of the unnamed language. From Grambank and multi-parallel corpora, we automatically build evaluation questions over 195 languages. We find that LLMs identify 59--80% of these languages from only 16 sentence pairs; and misaligning the translations barely affects their accuracy, suggesting reliance on prior knowledge of the language. Enciphering text to prevent recognition increases the corpus' utility, but mainly for features where target and pivot languages agree. On languages which models cannot identify, neither open model outperforms a majority-value baseline. Since induction from raw text aligns with field linguists' use-case, this work indicates the models tested are not yet ready for this role. TFI offers a scalable path to track progress in this direction.
Yining Zhang, Mingyang Yi, Chen Wang, Xuwen Xiang, Tianhe Jia, Zedong Dan, Chengqing Zong, Yue Wang
High-performance GPU kernels are essential for efficient LLM deployment, yet optimizing them remains expertise-intensive. Recent LLM-based code generation makes automatic GPU operator generation promising, but operator optimization remains a hardware-aware search problem. Existing LLM-based methods face a granularity mismatch: coarse hints are reusable but hard to execute, whereas detailed memories are actionable but enlarge the search space and obscure optimization bottlenecks. The key challenge is therefore to organize optimization experience at an appropriate granularity. To address this issue, this paper proposes HTAM (Hierarchical Transition-Attended Memory), a coarse-to-fine framework for LLM-based operator optimization. HTAM builds a two-level Hierarchical Transition Graph (HTG) to organize coarse global directions, detailed local strategies, and transition experience between optimization steps. During each evolution step, HTAM selects a global direction from the current state and recent optimization history, retrieves the corresponding local strategy memory, and uses it to guide concrete CUDA code generation. Experiments on KernelBench demonstrate that HTAM improves correctness, fast-solution rate, and speedup over LLM-based baselines, while backend and Robust-KBench studies indicate transferable benefits from structured memory.
The "Min-Max" theory of language behavior proposes that humans engage simultaneously in minimization of processing effort and maximization of information content, and that this is a primary driver in language evolution (Ring, 2025). This claim seeks to unify observations from language processing (Mahowald et al., 2013; Futrell et al., 2015; Yadav et al., 2020; Hawkins, 2014), interaction (Hörberg and Sjons, 2023; Tucker et al., 2025; Jaeger and Tily, 2011; Smith and Culbertson, 2025), and corpora (Levshina and Moran, 2021; Levshina, 2023; Newmeyer, 2017), which suggest that considerations of processing efficiency and information content are both implicated in human communication. The interaction between these two processes provides an explanation for human linguistic behavior at various levels (Linders and Louwerse, 2023) and predicts historical developments in language, i.e. related to word order (Ring, 2025). Given that the original formulation of the Min-Max theory is based on observational data, the current paper seeks to validate the theory mathematically. Through a series of experiments that use computational models to investigate the theory's predictions, pretrained and iterated learning models are queried with synthetic data. The experiments show that manipulating the variables of token length and the weight of information content in a tabula rasa model results in word order change, as predicted by the theory. The experiments also identify important differences between computational models and observations of human language data, suggesting that current transformer-based language models may not accurately represent psycholinguistic processes. The overall findings indicate that evolutionary trajectories in language are highly susceptible to minimization of processing and maximization of information content.