Natural Language Processing Papers

ALT We replicate the riichi-mahjong track of "Can Large Language Models Master Complex Card Games?" (arXiv:2509.01328) under a drastically smaller budget: models of at most 2B parameters, one 16 GB consumer GPU, about five GPU-days. (1) With matched data and labels, text is the bottleneck: on the same 100k-decision training set, a 2.6M-parameter convolutional board encoder agrees with the strong AI Mortal on 80.6% of test decisions, above every text-prompt LLM we train (best 77.8%), and places better at the table (2.85 over 1,000 games vs. 3.0 for the LLMs; paired difference at n=100 not significant). Scaled to 1.71M decisions, a 4.9M-parameter Transformer reaches 94.7% agreement, beats the 0.8B LLMs on identical seeds (-0.45 ranks, sign test p=0.005), and is statistically equivalent to parity against three Mortal champions (TOST within 0.10 ranks, p=6e-4); parity against a teacher-opponent is roughly the ceiling of pure imitation. (2) Training toward the teacher raises teacher agreement m

ALT User memory helps personalized systems build on their understanding of users over time. In e-commerce, users convey their preferences and needs through both explicit language and implicit interactions with products. We study memory built from logs of these interactions, examining how it can remember concrete evidence and established interests, and use that evidence to infer underlying needs and preferences that inform judgments beyond well-observed product domains. Existing memory methods primarily target language-based interactions, and evaluations offer limited real behavioral data and limited separation of these capabilities. This leaves open how to jointly support and separately evaluate remembering and inference from behavioral logs. We formulate the problem of continual behavioral user memory around these joint requirements and develop a dedicated evaluation framework and memory method to address them. BehaviorMemBench is a benchmark built from real e-commerce logs covering 1,194

ALT Knowledge editing aims to update outdated or incorrect knowledge in large language models without full retraining. Existing parameter-editing methods remain limited for long-sequence knowledge: single-point editors provide insufficient coverage for extended targets, while fixed chunk-based editors distribute editing effort uniformly across rigid predefined windows, failing to prioritize difficult, information-dense tokens. In this work, we show that token entropy is closely associated with both editing difficulty and semantic importance: high-entropy regions are harder to edit and often contain key factual content. Motivated by this observation, we propose DynEdit, a dynamic framework for long-sequence knowledge editing. DynEdit uses token-level entropy to guide non-uniform sliding, increasing the coverage of difficult high-entropy regions, and further applies entropy-guided token-level optimization to strengthen updates on information-dense tokens within each editing window. Experimen

ALT Effective collaboration requires language models to understand what others know and believe, and to act accordingly. Models that assume others share their information may omit necessary communication and undermine coordination. Many theory-of-mind (ToM) benchmarks assess mental-state inference through story questions, but correct answers do not establish that models can use these judgments in subsequent tasks. We introduce a multi-level evaluation built on existing, verifiable tasks. We extend household collaboration and calendar scheduling with asymmetric information while retaining their original constraints and validators. Independent interfaces assess world understanding, information access, beliefs, decisions, and execution. Controls supply personal evidence or correct beliefs to distinguish state recovery from state use. Across six model configurations, these abilities diverge. On matched scheduling cases, Claude recovers all agent beliefs correctly but produces subjectively opti

ALT In high-stakes domains such as finance, critical information evolves through multi-party interaction. Recovering these changes and their sources is fundamental to reliable speech understanding. Existing speech benchmarks focus on transcription, local understanding, and turn-level state prediction, leaving joint decision evolution across speakers insufficiently evaluated. We introduce DecisionTraceBench, a speech state-tracking benchmark for Decision-Critical Information State Tracking (DCIST). It jointly evaluates decision event sequences, speaker attribution, scope, and temporal evidence, using trajectory-level F1 to measure complete event-chain recovery. Our dataset comprises Chinese financial dialogues involving two to four speakers, with 100.39 hours of synthetic audio and 8.61 hours of human-performed recordings. On a fixed cohort of human-performed recordings, five cascaded systems and two direct-audio models achieve DCIST scores of 5.73% to 21.35%. Ablations with speaker-labeled

ALT Knowing when they are reliable is essential for language models to support human-AI collaboration. Yet interaction feedback is not always complete or correct, and may mislead models' understanding of their own capabilities. We introduce LADDER to study how language models assess themselves under unreliable feedback. Our findings show that models can accurately report supplied performance information while still struggling to form trustworthy self-assessments from interaction experience. Apparently accurate assessments can obscure capability differences and do not necessarily predict future task performance. These findings suggest that trustworthy self-assessment should be evaluated against observed task performance and its predictive value. Our study provides empirical evidence of the limits of model self-knowledge and a basis for evaluating capability awareness in reliable human-AI collaboration.

ALT An agent that truly knows and acts on behalf of its user has been a long-standing promise of AI assistants. Computer-use agents bring this promise within reach by capably executing fully specified tasks across apps. However, real-world user requests are often underspecified. To act correctly for the particular user, an agent must infer some missing information from the user's activity history and ask for other information that only the user can provide, without making unsupported assumptions or burdening the user with unnecessary questions. We present iPrefBench, a benchmark for evaluating the infer and ask capabilities of personal mobile-use agents in an executable iOS environment with an interactive simulated user. Grounded in a coherent persona with 3,051 past activity events, iPrefBench comprises 130 underspecified user tasks across 26 apps. Each omitted user-dependent requirement is, by construction, either history-inferable or ask-only. iPrefBench evaluates personalized mobile ag

ALT Latency-sensitive spoken agents must know not only which tool to call, but also when they have heard enough to act. Waiting for the turn to complete delays execution even when the request is already actionable. However, neither existing spoken agents nor spoken dialogue corpora define the earliest point at which a request becomes actionable. We introduce a synthetic spoken-dialogue pipeline that provides earliest-actionable-point supervision, marking when the observed speech grounds a tool and its required arguments. Through controlled task generation, speech synthesis and alignment, and gated annotation, we construct 26,694 dialogues totaling 786 hours across 48 domains, with disjoint training and test argument values. We additionally create an audio rendering of τ²-bench for external evaluation. We train an omni-modal language model on this corpus, enabling causal, chunk-level decisions and partial-transcript generation for downstream call completion. The model achieves tool-state ma

ALT Entity matching is the task of deciding whether two records describe the same real-world entity. Current entity matching methods rely on pre-trained language models (PLMs) which require task specific fine-tuning or generative large language models (LLMs) which do not require task-specific fine-tuning but are slow and expensive at matching large record sets. Decision models are a new class of LLM-based models that return a probability for each predefined answer option rather than generate text. This paper investigates decision models as an alternative to generative LLMs and fine-tuned PLMs for entity matching. We evaluate six decision models in a zero-shot setting across eight benchmarks and compare their performance with that of the generative LLM GPT-6 Luna and the PLM-based matcher Ditto. Our experiments show that, without task-specific training data, the four strongest decision models achieve mean F1 scores close to that of GPT-6 Luna, with the best model slightly surpassing it. Thi

ALT A chunky gold-tone chain necklace bought online for 2.43 USD was photographed four ways: worn with formal attire against a wood-panelled wall, worn at a party, worn with casual clothing in a recycling yard, and laid flat with no person present. We asked six frontier language models to estimate its price in 1,164 independent and sequential API sessions (4,604 logged rows) and compared estimates across settings. Among the four models that answered image prompts, the same necklace was priced 1.9 to 3.3 times higher in the formal photograph than in the flat-lay (all p < 0.01). Against the yard photograph, the ratio ranged from 1.23 (Grok 4.5, not significant) to 3.57 (Kimi K3), with Claude Fable 5 at 3.32. A text-only version of the task, in which the setting is described in words, produced formal-to-yard ratios of 2.2 to 3.9 in all six models, so the effect does not require pixels. Flat-lay estimates themselves were 7.8 to 18.7 times the receipt price, so part of any headline "outfit mult

ALT Large language models (LLMs) are neural networks trained on huge text collections to generate text by predicting what comes next. Engineers using ChatGPT and other generative artificial intelligence (AI) tools need a clear picture of what training changes and prompts add. The article starts with the Transformer, which connects words in context, and pre-training, where models learn to predict text. It then explains how fine-tuning adapts models to requests, with reinforcement learning from human feedback using human answer rankings. Next, retrieval-augmented generation (RAG) supplies documents, while chain-of-thought prompting encourages intermediate steps and agents use tools. FlashAttention lowers training and running costs by reducing data movement during attention; quantized low-rank adaptation (QLoRA) saves fine-tuning memory. Readers get a life cycle figure and tables comparing methods and their limits. A short checklist for judging LLM evaluation claims flags test material in tra

ALT The largest open German text datasets are built from broad web crawls, which provide scale but limited coverage of specific domains. We introduce RefinedGAW, a processed and annotated resource derived from 26 semi-annual crawls of the German Academic Web, totaling 122 million documents and 314 billion tokens, with full crawl and capture history of every text. Documents labeled as academic research account for 84% of German RefinedGAW tokens across HTML and PDF. At 87.6 billion tokens, this exceeds the 64.9 billion tokens in the German subsets of FineWeb2 and FinePDFs, despite their almost five-times-larger combined size. RefinedGAW also contains substantial amounts of educational, math-heavy, and expert-oriented content with minimal commercial bias. Across HTML and PDF, 88% of German RefinedGAW documents have no near-duplicate candidate match in the three evaluated reference datasets, supporting complementary coverage. We release a queryable metadata index for fine-grained data selecti

ALT Estimating the correctness of large language model (LLM) predictions without access to ground truth is critical for reliable deployment in high-stakes settings. Existing approaches, such as entropy and perplexity, operate solely on final output statistics, discarding the rich information encoded in intermediate reasoning steps. To address this limitation, we propose STEP (Soft Trajectory Estimation of Prediction correctness), a training-free framework that scores each reasoning step via lightweight forward passes to obtain a soft probability distribution qi over candidate answers, from which we derive a set of soft trajectory features capturing per-step confidence and its progression toward the final answer. Extensive experiments on various benchmarks demonstrate consistent and substantial improvements, particularly in AUROC, over output-based baselines and hard-indicator trajectory methods. Notably, Temporally-Weighted Consistency, which assigns higher importance to later reasoning st

ALT A long-lived assistant that brings up memories without being asked must decide at every turn whether anything in its store is worth an interruption. We show that this decision is a multiple-testing problem whose difficulty grows with the store size N. On streaming replays of LongMemEval and LoCoMo, a similarity threshold tuned to 5% false alarms at N=8 reaches 87-100% at N=2048, across embedding models and for explicit and implicit cues; LLM judges show the same growth from a lower base. We prove that a gate with false-alarm rate α operates at a per-item false-positive rate of about -ln(1-α)/N, which pins its hit rate to the far left of the single-comparison ROC curve and, under regular variation, makes it decay as a power of N with an explicit constant. The resulting prediction matches the hit rate of a per-N oracle threshold to within one percentage point. The analysis explains why lexical and semantic channels need thresholds growing like ln N and √(ln N), and why per-cue normalisat

ALT The Voynich manuscript is a fifteenth-century book in a script that no one can read. One account holds that its text means nothing and that a scribe made it by copying his own words and changing them a little. The other holds that a real language lies hidden behind a cipher. We write each account as a program that fills the pages of the manuscript, fit every program with the same budget, and check it on statistics the fit never saw. The copying process of Timm and Schinner predicts these statistics far better than any cipher with one key. But ciphers that switch keys between the manuscript's two dialects, or that encode only the word classes of a real text, predict them equally well. Word statistics alone cannot decide. We look for a trace of language that needs no decipherment. In 28 texts in 16 languages the same word pairs keep sharing a line. The manuscript shows this pattern at half the strength of the weakest of these texts. A word code keeps the pattern above the manuscript unle

ALT In formal systems such as Lean 4, efficiently finding answers to mathematical problems and generating verifiable proofs is an important task in automated mathematical reasoning. In the traditional way where an upstream solver provides multiple candidate answers, the common workflow requires searching for proofs for each candidate separately, incurring redundant computational cost. To this end, we propose a joint proof framework that integrates candidate answer selection, answer supplementation, and formal proof generation into a unified reasoning process. At its core is a trained prover that selects given or proposes new answers and generates a proof for at least one answer in a single call. Joint proof can also reuse generated proofs to check for faithfulness errors. For problems with a unique answer, if two distinct answers both obtain verifiable proofs under the same formal statement, this confirms the statement does not faithfully preserve the original problem semantics. On MATH-50

ALT Understanding how individual token choices affect reasoning length is important for characterizing reasoning models. Building on prior findings that changing the first reasoning token (hereafter, the first token) can substantially alter reasoning length, we focus on this initial choice to investigate how its influence operates during subsequent generation. One possibility is that the first token changes the early text that follows, which in turn affects reasoning length. However, replacing only the first token of an existing reasoning prefix and resuming generation substantially changes the remaining reasoning length, even though all other prefix tokens are unchanged. To trace this persistent influence, we compare next-token distributions while varying only the first token and find that sensitivity is concentrated at a small set of generation positions. Blocking attention from these positions to the first token reduces the length gap between first-token conditions. Using next-token dis

ALT Structured semantic parsing is typically judged by whether the correct category is assigned, which leaves value recovery and instance-level associations unexamined: a system may recognize an event yet bind an argument to the wrong instance, or identify an intent yet supply an incorrect slot value. We present USP, a cross-family semantic contract that represents categories, fields, values, and instance associations through a common interface for named-entity recognition, relation extraction, event extraction, and intent-slot parsing. The contract is paired with five diagnostic matching views—category agreement, field recognition, field-value pairing, instance-bound assignment, and complete-instance recovery—that ask distinct questions about the same output. Using this contract, we normalize 61 datasets into a unified corpus with shared schemas, prompts, and matching rules, and compare six models on a fixed test subset. Our fine-tuned Xing4.0-29B-Intent attains 57.70 task-balanced Field-

ALT Large reasoning models are costly to serve because they are large and because they generate long chains of thought (CoT). Weight pruning and CoT compression address these two costs separately, but they interfere in that pruning adds redundant reasoning on problems the model still solves and incomplete reasoning on those it no longer solves, and on harder problems makes the model less compressible by subsequent length-preference optimization. We trace this to a blind spot on each side: pruning criteria do not distinguish necessary from redundant reasoning, and preference-based CoT compression receives no training signal on problems a pruned model always or never solves. We propose Contrastive Set-Difference (CSD), which identifies neurons enriched in redundant reasoning by contrasting their activations on long and short correct solutions of the same problem and prunes them preferentially, and we extend preference pairs to the problems a pruned model always or never solves. Pruning 20% o

ALT Extracting salient information from long-horizon dialogue histories remains a significant challenge for Large Language Models (LLMs) due to the restricted context windows and efficient retrieval challenges. Current methods typically rely on history organization or local token-level optimization, suffering from information loss, retrieval redundancy, and myopic attention. Inspired by the view in cognitive science that dialogue constitutes sense-making and memory association, we propose CoLink, a framework that enhances LLMs' global context awareness in domain-specific multi-turn dialogues through information linking and compression. CoLink first clusters historical contexts into topic graphs to establish cross-semantic associations with external knowledge. Guided by the adaptive control of thought, it then retrieves and activates knowledge via synergizing semantic relevance and graph-structural topology. Crucially, to achieve effective compression of history and knowledge, we ground mul