跳到正文
研踪PaperTrail

论文库

检索计算机领域论文，按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+0新增论文
+0其中 AI 相关
+0已有论文修订
+93新增 AI 导读

arXiv 最近完成扫描：2026-11-16 01:34 · 全站最近入库：2026-11-14 15:09各信源进度 →

统计口径与数据新鲜度

统计 2026-11-15 04:02 — 2026-11-16 04:02（北京时间） · 点击数字查看对应简报；AI 相关包含在新增论文内。

最近完成检查：2026-11-16 01:34最近发现新增版本：2026-11-14 15:09库内论文最新更新时间：2026-11-13 02:59

每分钟刷新已保存的数据。空检查不代表采集异常；新收录可能包含历史论文。AI 导读按新增版本计数。

23,348 已收录论文19,580 AI 相关6,729 有修订112 HF 社区发现

浏览论文库

研究日报 / 周报 →按信源浏览 →

时间范围跟随排序：版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常，查看状态 →

找到 4,989 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-16 01:33（北京时间）
新论文 · v1arXiv:2611.16803

检测不等于预警：评估语言模型失败中的Token级警报

Detection Is Not Early Warning: Evaluating Token-Level Alarms on Language Model Failures

Marco De Negri, Luca Sartori, Lorenzo Baggi, Moritz Reihs, Eduard Frank Ďurech, Mennatallah El-Assady, Valentina Pyatkin, Anna Hedström

AI 速读 基于摘要

提出PREWARN评估协议，在重复退化案例中发现响应级指标饱和，而线性探针能在多数退化响应开始前发出预警。

cs.LG · cs.CL

arXiv ↗PDF ↗
查看原文摘要

Many failures of language models begin at a specific token, the onset. A monitor that reads the model during generation can classify the finished response, detect the failure once it is under way, or warn before it begins, and only a warning leaves time to stop or steer the generation. Monitors are usually evaluated with response-level metrics such as AUROC, which ignore when the monitor fires. We introduce PREWARN, an evaluation protocol for any failure with an annotated onset. It calibrates each monitor's threshold on validation data to an error rate chosen for the application, runs the monitor, and scores its alarms relative to the onset on all three tasks. As a case study we apply it to repetitive degeneration in three open language models, with early warning as the main task, using linear probes on activations and six untrained baselines. Response-level metrics saturate on this failure, and a rule that reads only the response length scores almost perfectly, while probes that AUROC ranks as equal warn very differently under PREWARN. Probes warn before the onset on most degenerate responses of two of the three models, low-rank adapters below the probe more than double the share of tokens just before the onset that it flags, largely by flagging degenerate responses long before the onset, and no untrained baseline warns much more often than random alarms. We release the probes with calibrated thresholds at https://huggingface.co/collections/marcodene/prewarn-degeneration-probes

版本更新 · v4arXiv:2611.05355

Reading Why: Born-Measured Attention with Exact Word Evidence on Light and Qubits

Teerawat Tongloy, Jiraporn Pongsopa, Rangsan Jomtarak

原文摘要

When a language model answers, we can usually only estimate why. We build a transformer in which every nonlinearity is a quadratic detection - an interference, a balanced reading or a camera - and the output is a Born probability: attention |<q|k>|^2, the form of the Born rule, normalised per row, a measured product for the feed-forward nonlinearity, and a Born readout; between detections, classical control rescales and re-encodes. With its denominators held at the input, each class probability is a polynomial in the word amplitudes, so its integrated-gradient evidence is computed by a finite quadrature with no quadrature error: the words' contributions sum to the class probability above its no-information share to 10^-13. On e-SNLI natural language inference it reaches 77.45 +- 0.91 % accuracy, level with a same-size softmax transformer (76.62 +- 0.36 %) and an order-free bag of embeddings (77.79 %). Exactness buys the sum, not the best ranking. Deleting the word its evidence ranks first changes the answer in 44 % of pairs, of the 53 % some deletion changes, against 31-32 % for the softmax model's integrated gradients; the model's own gradient x input does as well (43 %), as does a bag of embeddings' (43 %, seed 0), and leave-one-out does better. Holding the denominators helps the low end: the five words ranked last change the answer in 1 %, against 6 % free. The same weights compile exactly to simulated optoelectronics, and their attention to qubits. On ibm_kingston, reading only the attention, the device gave the model's answers in all three Hadamard-test runs (the two shortest pairs and one without its deciding word) and both compute-uncompute runs (one circuit per overlap); uniform attention answers otherwise on two of the three inputs. With the device's row sums the evidence correlated with the exact one at 0.997 and 0.962 by Hadamard tests; with compute-uncompute, 0.971 and 0.48.

quant-ph · cs.CL · cs.LG · physics.optics

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16714

Towards Green Text Generation: Memory-Based Speculative Decoding

Antal van den Bosch

原文摘要

Speculative decoding accelerates text generation with large language models by letting a cheap draft model propose tokens that a large verifier checks in a single forward pass. We introduce Olifant, a draft model built on a memory-based classifier. It predicts the next token from a four-token window in under 0.3 ms on CPU, requiring no GPU, no backpropagation, and no accelerator memory. We measure throughput and energy under two verifier regimes on consumer hardware. Without a KV cache, drafting is nearly free: a domain-matched Olifant model reaches 38.1% acceptance on EU legal text, delivering 2.43x throughput and 2.51x lower energy per token than the neural verifier model alone. With a KV-cached verifier, itself a ~10x energy saver, drafting costs matter: a TinyLlama draft model halves throughput and triples energy despite high acceptance, whereas Olifant still yields 1.25x throughput and the lowest energy in the study (93 J per 100 tokens, 14x below the uncached baseline). We further show that KV caching decouples time from energy: verification of longer drafts raises power draw, shifting the energy-optimal draft length below the speed-optimal one. Speculative decoding is not dead in the age of fast verifiers, provided drafting is nearly free.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.16837

A Scaled-Up Empirical Study of Syntactic Language Models

Pengcheng Wang, Zhexuan Zhang, Chao Lou, Kewei Tu

原文摘要

Syntactic language models (SLMs) that generate sentences along with their syntactic tree structures have shown promise in both syntactic generalization and downstream tasks. Prior studies, however, have mostly evaluated smaller models trained on less pretraining data and tested on narrower task suites. In addition, the design space of tree linearization and attention masks remains underexplored. We conduct a large-scale empirical study of 12 controlled configurations---one terminal baseline and 11 SLM variants spanning different linearizations and masks---at scales up to 1B parameters and 100B training tokens, evaluated across 11 downstream tasks and two syntactic benchmarks. The experimental results show that SLMs with standard causal attention match or exceed structural-mask variants across downstream and syntactic benchmarks. They require no architectural modifications and remain compatible with FlashAttention. We also observe that when scaling up to the 1B-parameter scale, the top-performing SLM variant no longer shows a clear advantage over a compute-matched non-structural baseline, suggesting that SLMs may become less competitive with scale. Our code is available at https://github.com/qlwpc/TG-Interpolation.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16704

Forecasting Sales Delay through Time-aware Analysis of Sales Events in Textual Reports

Masaya Tsunokake, Yuta Koreeda, Yuhei Niwa, Terufumi Morishita, Teppei Inoue, Hiroaki Ozaki, Yasuhiro Sogawa

原文摘要

Business-to-business (B2B) companies develop and deliver a one-of-a-kind product in response to a client's orders. To prepare resources for development beforehand, companies estimate dates when they will receive orders from clients in their sales process, called scheduled order dates. However, dates when the companies actually receive orders can differ from the scheduled order dates. This difference increases costs in the development. If the difference could be forecasted before receiving orders, companies could optimize their resource preparation. Towards reducing the costs, we introduce sales delay forecasting, a novel task of forecasting delay days from the scheduled order dates during sales process where sales representatives (sales reps) interact with the clients to receive orders. Sales delay forecasting is challenging because the sales process progress inherently involves uncertainty and depends on the client's willingness to order. To address this task, we propose confidence features which represent sales reps' confidence as a proxy measure of the client's willingness to order. To represent it, our method extracts dates of future sales events anticipated by sales reps from textual reports. To consider whether the confidence objectively looks reasonable, we also introduce status features which describe sales process status based on current date and past events. In experiments using real sales data, our method achieved a 16% lower mean absolute error compared with a feature-based baseline similar to a prior work. Moreover, our method outperformed non-feature-based baselines using large language models, indicating that it is crucial to explicitly create our features.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16694

When Expert Disagreement Hurts: Prestige and Selectivity in LLM Decision Revision

Yupeng Tang, Mingfeng Lin

原文摘要

When language models revise a decision after another agent disagrees, neither reversal counts nor final accuracy shows whether they are fixing mistakes or abandoning correct answers. We introduce a controlled two-pass audit that holds the question, the evidence, and the model's initial answer fixed and varies only the revision context: neutral reflection, anonymous disagreement, or disagreement attributed to a domain expert. For binary decisions the accounting is exact: revision raises accuracy only if it is selective enough, reversing wrong answers more than $a/(1-a)$ times as often as correct ones, where $a$ is initial accuracy (nine times at 90%). Across five models and 1,989 evidence-grounded decisions in medicine, science, and law, no model clears this bar under expert disagreement, and accuracy falls by up to 16.1 points. The expert label, a prestige cue, adds reversals for four of the five models, yet its effect on accuracy ranges from a 5.2-point loss to a 2.4-point gain, so reversal counts are a poor proxy for harm. With identical arguments from both sources, the label still raises reversal for three proprietary models and in a pooled analysis of four newer models. Models also respond to stated source reliability. Revision pipelines should therefore be evaluated with paired, conditional metrics, per model and protocol.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16662

MuSeToM: Multi-Sensory Evaluation of Belief Revision with Textual Cues Across Six Senses

Chunkit Chan, Kawai Chung, Sze Heng Douglas Kwok, Yuxuan Liu, Yauwai Yim, Tianshi Zheng, Haochen Shi, Qing Zong 等 16 位作者

原文摘要

Theory of Mind (ToM), the ability to infer mental states in others, is pivotal for human social cognition. Existing evaluations of ToM in large language models (LLMs) mainly focus on static belief attribution and false-belief reasoning, leaving a critical aspect of social reasoning underexplored: can LLMs decide whether perceptual evidence is sufficient to justify another person's belief revision? To address this gap, we present MuSeToM, a human-validated, text-based benchmark for evaluating epistemic belief revision in ToM reasoning. MuSeToM contains 2,415 instances adapted from BigToM and OpenToM, with textual cues spanning six sensory categories, and tests whether models can determine when a protagonist should revise a belief and infer the resulting updated belief state. Using MuSeToM, we systematically evaluate state-of-the-art LLMs under direct and chain-of-thought prompting, revealing a pronounced dissonance: while models often detect that a belief may change, they struggle to identify the updated belief and frequently over-update on perceptually salient but epistemically insufficient cues.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2611.13607

One Byte, One Token: Byteifying a 27B Hybrid-Attention LLM with Self-Generated Data

Liwei Zhang, ZhiMing Tong, Mochen Guan, Linzhe Li

原文摘要

Byte-level BPE fragments Chinese text (for Qwen3.8 on Chinese Wikipedia, 3.84 bytes per token on average; 64% of tokens are single characters) and hides spelling from the model. We ask whether an already-trained 27B subword model can be converted directly into a pure byte-level model (one byte, one token), without a patching structure or a local encoder, while keeping most of its ability. Starting from Qwen3.8-27B (48 Gated DeltaNet layers + 16 full-attention layers), we shrink the vocabulary from 248,320 to 384 entries (the 256 byte symbols already in the BPE vocabulary plus special ids), leave every other weight untouched, and continue training in a two-stage curriculum (encoding first, then reading) on 4.77 GB of supervised bytes: all dialogue/behavioral supervision is generated on-policy by the original model itself (character dictionary, word dictionary, general dialogue, coding sessions), and knowledge reading uses encyclopedia text (54% of supervised bytes). Training takes 178.6 h on 8xH200. On the same GPU and serving stack, the byte model's gap to the original falls within the paired bootstrap interval on MATH-500, HumanEval+, MMLU-Pro and non-thinking GSM8K, and it loses 4-6 points on C-Eval, thinking-mode GSM8K and non-thinking IFEval. Without thinking it leads on every character-level task (CUTE 75.2 -> 84.0, EXECUTE-zho 81.6 -> 93.9, our Chinese character suite 54.4 -> 64.0), but lookup knowledge about Chinese characters such as pronunciation and stroke count drops, and thinking does not recover it. The cost is throughput: per-step speed is identical, but each step yields 3-5x less output; Chinese characters per second is about 1/4 of the original, and prefill at 64K bytes is 6.5x slower. We also document four byte-model failure modes (greedy runaway, sampling breaking UTF-8, ineffective penalty terms on a small vocabulary, repetition) and their fixes.

cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16612

Unibuc - NLP at "Voight-Kampff" Generative AI Detection PAN 2025

Teodor Marchitan, Claudiu Creanga, Liviu P. Dinu

原文摘要

This paper describes the approach of the Unibuc - NLP team in tackling the "Voight-Kampff" Generative AI Detection task from PAN at CLEF 2025. In the Subtask 1: Voight-Kampff AI Detection Sensitivity we have explored different approaches of leveraging embeddings extracted from LLMs, using either classical ML models or further fine-tuning LLMs with a classification head for the downstream binary classification task. Our experiments showed that using a majority vote ensemble of classical ML models (Light GBM, XGB, Logistic Regression and SVM) is comparable with fine-tuning for downstream task achieving around 99.4% F1 score. For the Subtask 2: Human-AI Collaborative Text Classification we explored a way of combining features at different layers extracted using transformer backbone with layer-wise projection and attentive pooling which proved modest results.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.29922

Unintended Side Effects of Value Training in LLMs

Siddhesh Pawar, Chen Cecilia Liu, Anna Korhonen, Isabelle Augenstein

原文摘要

As LLMs are increasingly deployed in value-sensitive contexts, aligning them with human values has emerged as a post-training objective in NLP, aimed at instilling additional values and guiding model behaviour. However, current evaluations remain narrow. Because values are interconnected, it is unclear how fine-tuning for specific values alters unrelated capabilities and behaviours. In this work, we adapt four instruction-tuned LLMs across individual Hofstede's value dimensions and evaluate them on both cross-value transfer and broader behaviours. We find that value-specific training produces unintended side effects on untrained values, risk preferences, social cognition, and safety traits. Consequently, we suggest that value alignment after post-training should incorporate broader behavioral evaluations, and we highlight the need to investigate where additional value training might best fit into the pipeline. We release all 48 value-tuned model checkpoints at https://huggingface.co/sidicity/cultural-value-models

cs.CL · cs.AI · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16581

Do the Details Help the Whole? One-Way Transfer from Person Actions to Group Intent in VLMs on JRDB-Social

Chengguang Gan, Zhixi Cai, Yunhao Liang, Hanjun Wei, QingHao Zhang, Shiwen Ni

原文摘要

Social scene datasets label people at two levels: what each person is doing and what the group is there for. Training on both levels at once is often assumed to help each of them. We test this assumption for vision-language models (VLMs) on JRDB-Social, treating person actions as the details and group intent as the whole. We fine-tune Qwen3.5 4B under eight training arms with three seeds, repeat the core arms on Qwen3.5 9B and Qwen3-VL 4B, and judge every arm with a verdict rule fixed before testing. Writing each member's actions before the intent raises intent macro-F1 from 30.1 to 31.0, and writing the intent first reaches 31.2, while action F1 stays within 0.2 points of an action-only model. Under the selected recipe, no arm trained with intent improves actions. The gain requires correct, image-grounded actions: empty action lists or actions borrowed from another group with the same intent lower intent F1. Forced-prefix tests, occlusion, member masking, linear probes and component grafting show that the intent follows the written actions, relies more on people and partly resides in the vision tower. Supplying actions as input instead of as targets hurts both levels.

cs.CV · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2608.14198

Multimodal Zero-Shot Prediction for Transaction Data at Scale

Parameswaran Kamalaruban, Viktor Drobnyi, Maeve Madigan, Julia Rozanova, David Sutton, Stuart Burrell

原文摘要

Sequential financial transaction data are increasingly being used by financial institutions to train payment foundation models at scale, producing rich contextual embeddings that support applications such as fraud prevention, credit risk assessment, and offer personalization. However, such models are not designed for flexible instruction-following reasoning and require bespoke downstream models for each application. In contrast, large language models (LLMs) provide a natural-language interface for flexible transaction reasoning, yet naively serializing transaction histories into text representations is inefficient and costly. Multimodal fusion architectures bridge these paradigms, injecting transaction embeddings directly into flexible language reasoning pipelines. In this work, we present a systematic study of how transaction embeddings, pretrained at the scale of a global payments network, interact with LLM-based reasoning through embedding injection, transaction-language alignment, and instruction tuning. To do so, we train the Multimodal Instruction Network for Transactions (MINT), which combines a transaction encoder pretrained on billions of real-world payments with multimodal instruction tuning on millions of transaction question-answering examples. We evaluate predictive and extractive question answering, zero-shot generalization to unseen question types, and inference efficiency. Our results reveal a consistent trade-off: compact transaction embeddings improve predictive QA accuracy while substantially reducing input tokens, latency, and memory relative to a text-serialized baseline, whereas serialized histories remain highly competitive for extractive QA. Overall, our empirical findings provide a set of actionable insights and practical guidance for building multimodal transaction-aware language systems at network scale.

cs.LG · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2611.13636

An Interpretable Stylometric Framework for Early Chinese Buddhist Translator Styles: A Case Study on the Attribution of Disputed Works of An Shigao

Yating Pan, Xinyue Zhang, Jun Wang

原文摘要

Attributing early Chinese Buddhist translations requires evidence that scholars can trace back to the texts. We develop an interpretable stylometric framework that uses features defined or reviewed by Buddhist philologists, including function characters, sentence formulas, sentence-final characters, sentence length, and part-of-speech n-grams. We construct stylistic profiles from 31 authenticated translations by seven translators from the Eastern Han through the Western Jin, then compare eight disputed texts with An Shigao's profile. Two texts, T1557 and T0101, show the strongest stylistic similarity to the authenticated An Shigao corpus, with scores of 0.783 and 0.687, respectively. The attributions agree with those reported by a published BERT-based approach on the same set of disputed texts. The framework requires no task-specific training and allows scholars to inspect which linguistic features support each proposed attribution.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16500

EvenVLM: A Text and Multimodal Dataset for the Even Language with Vision-Language Fine-Tuning

Daniil Makovetskii, Ivan Bondarenko

原文摘要

We present EvenVLM, the first text and multimodal (VQA) dataset for Even (Lamut), an endangered Northern Tungusic language spoken by approximately 6,800 people in Siberia. Our dataset comprises 50,521 structured Even-Russian entries extracted from dictionary and phrasebook sources via GPT-4o Mini-based parsing, from which we synthetically generate 107,543 training examples spanning translation, grammar, reading comprehension, and semantic classification. To address tokenization inefficiency on Even's agglutinative morphology, we propose a three-step tokenizer expansion algorithm that adds Even-specific graphemes (eng, oe, o-umlaut), whitelisted vocabulary, and BPE merges, reducing average tokens per word by 27% (4.99 to 3.64). Using LoRA fine-tuning of Qwen3-VL-4B, we adapt the model for text tasks (0.53% trainable parameters, rank 32) and separately for visual question answering (rank 64, frozen vision encoder). Our best configuration achieves 11.4% exact match and a 13.6 weighted score on text benchmarks, 0.1283 chrF for sentence translation, and 65% accuracy on VQA (up from 0% for the base model and text-only fine-tuning). We release our datasets, evaluation benchmarks, and training pipeline to support further research on Even and other low-resource languages.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16477

Where in the Input Does a One-Pass Decision Encoder Look?

Mohammed Nayeem Khan, Allan Kevin T

原文摘要

Laya is a small model that answers a multiple-choice question about a text in one pass: the question, the options (each with a name and a description) and the text form one sequence of at most 512 tokens, the window, in the English release. A longer text is silently cut at the end, and we measure what that costs. Test documents place a held-out row's text, the evidence, at a chosen token offset among distractor text; 177 items, five tasks, three checkpoints (released, short-row fine-tune, long-document fine-tune), four window sizes and 93,275 scored requests. When the evidence lies beyond the window, accuracy falls to 26.4% for the released checkpoint (27.2% for both fine-tunes), near the 27.0% of guessing and below the 35.6% of always giving the commonest label. Confidence gives no warning (AUROC 0.50 to 0.55 for telling cut from read requests; 0.5 = chance), and the runtime's truncation flag also fires on requests whose evidence was visible. The released checkpoint reads visible evidence badly once text precedes it: 38.5% pooled against 60% when the evidence comes first, so a larger window alone gains +6.2 points. Fine-tuning on long documents adds +27.6 points on visible evidence over the short-row fine-tune (95% interval +22.3 to +33.1) and nothing on cut evidence. That is an in-family result: training rows and test documents share generator and question types, and the main filler is of the training kind. With encyclopedic filler never seen in training the gain is +13.2 (+9.9 to +16.6), with rows of other tasks as filler -2.0. Among training-free rearrangements a 312-token window scan helps most (+11.9 points over the default for the released checkpoint) at six to eleven forward passes. Documents are synthetic, labels are LLM-made and one backbone is studied, so these diagnose this model, not long documents in general.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16466

How Should Planning Agents ReSHAPE Their Search Space?

Yilin Zhang, Xinran Zhao, Tong Chen, Ziming Yuan, Tongshuang Wu

原文摘要

Planning agents and benchmarks treat a user's initial specification as fixed. Yet the requests users write exclude plans they would gladly take, since they rarely discover the possible trade-offs they would accept until seeing the options. Agents already revise requests through interaction, but whether this helps planning is unknown. We formalize this step as the ReSHAPE task, which judges a revision by whether its search space matches the acceptable plans and a planner finds one there. Using ReSHAPE Planter, which builds instances from planning benchmarks with a human-validated simulated user, we analyze 5,800 agent-revised specifications in travel and shopping. We find that reshaping matters: one correct trade-off makes the same planner reach an acceptable plan up to 4.6x as often, and search space quality predicts what the planner finds. Current agents, however, admit at most 32% of the acceptable plans, where that trade-off alone admits over 52%, because they ask about preferences instead of showing plans; in travel, showing plans raises F1 by 9.7 points over asking, with two model families. Planning agents should thus treat a request as a hypothesis to test against the user's reactions to concrete plans to recognize user trade-offs.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16433

daVinci-Prover: Agent-Native Mid-Training for Formal Theorem Proving

Wenhong Zhu, Xuefeng Li, Jiashuo Zhang, Yu Chen, Tiantian Mi, Qipeng Guo, Pengfei Liu

原文摘要

Advances in large language models (LLMs) are rapidly transforming formal theorem proving. Yet further progress remains constrained by the lack of a transparent and effective data recipe for systematically improving theorem-proving models. In this work, we argue that effective agentic provers require strong mathematical skill priors to ground theorem proving and agentic trajectories to enable long-horizon, interactive proof construction. To this end, we synthesize an agent-native 33.7B-token mid-training corpus covering decomposition, search, self-verification, and revision, together with verifier-grounded agentic trajectories. We further introduce the daVinci-Prover harness, which enables models to construct proofs and interact with a Lean verifier in an isolated workspace, and use it to collect approximately 36K high-quality agent trajectories for supervised fine-tuning. Our mid-training recipe yields an average improvement of 13.7 points over the matched Qwen3-8B SFT-only baseline across MiniF2F, FATE-M, and ProofNet. When applied to the Qwen3.5 series, our approach also yields strong performance across theorem-proving benchmarks and achieves near-saturating MiniF2F performance with memory compaction. The results highlight agent-native mid-training as an effective foundation for developing stronger interactive theorem-proving agents.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v4arXiv:2610.32955

Measure Before You Train: Typed Decision Models vs. a Classical Baseline vs. a Generative LLM for Routing a Personal Multilingual Inbox

Federico Lucca

原文摘要

Personal email agents decide, for every message, its type and urgency and whether to draft a reply, create a task or ignore it. In production these decisions are delegated to a generative LLM or to rules, and their error rate is rarely measured: the stored labels were written by the system itself. We describe Foyer, a self-hosted personal inbox agent, and a measurement protocol (stratified sampling, frozen snapshots, thread-grouped splits, secret redaction, human labels in the review loop) that produces ground truth before any training. On 231 emails in Italian and English we compare eleven systems: the stored production type, a TF-IDF + logistic-regression baseline, local and hosted typed decision models (Tev1, Nimble, Jeva, Jev) and generative LLMs. Four results. The stored production type scores 28% against a 57% majority class. With about two hundred labels a logistic regression has the highest point accuracy on every field (type 0.77 +/- 0.02 over five seeds) at a thousandth of the latency, but on every field at least one hosted model is statistically indistinguishable from it (exact McNemar, Holm-corrected). A generative model recovers part of the sender's relationship to the owner from the text (0.60); inheritance from the contact is right 98.6% of the time where the sender was seen before: it belongs in the contact graph, not in the prompt. Per-email accuracy is not the product metric: composed into actions, the classifier ignores 18% of the emails that needed action, against 4% for the best generative model; a cost-chosen threshold cuts that to 10%, indistinguishable from the hosted decision model's; on the adversarial corpus the thresholded classifier is the cheapest system above a missed-email price of 4. Two blind annotators agree with the reference at kappa 0.81 and 0.86 on type. v4 corrects threshold results computed on a truncated corpus in v1-v3.

cs.CL · cs.HC · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16409

Experience-Handoff Reinforcement Learning for Recursive Self-Improvement

Doyoung Kim, Byeongguk Jeon, Kailin Zhang, Sherry Yang

原文摘要

It is common to train Large Language Models (LLMs) for agentic tasks with Reinforcement Learning (RL) from scalar rewards. However, this throws away the fine-grained information about the environment that the agent experienced during the trial (e.g., error logs). To utilize this rich information, recent works wrap the agent in a harness that refines its own inputs through reflection or compaction and hands past experience to future rollouts, but the construction of the handed-off artifacts (e.g., a playbook) is itself not learned. We want the best of both worlds: gradient updates from a scalar reward, while also utilizing environment information to hand off experience. As a concrete testbed, we fix the Agentic Context Engineering (\ace{}) harness and train one model to act, reflect, and curate task-local playbooks. Our method, \acerl{}, evaluates candidate experience-handoff artifacts by their downstream effect on future actor rollouts, and co-optimizes the role-specific behaviors (action generator, insight reflector, and playbook curator) under a fixed scaffold. Across agentic benchmarks, \acerl{} achieves higher held-out performance than baseline methods. The results suggest that agent training should optimize not only the actions in the current rollout, but also the model-written experience that future rollouts inherit.

cs.LG · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2606.27025

Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role-Aware Policy Optimization

Zhenhua Xu, Dongsheng Chen, Jian Li, Yitong Lin, Zhebo Wang, Jiafu Wu, Yizhang Jin, Chengjie Wang 等 10 位作者

原文摘要

Building general-purpose role-playing agents that faithfully portray any character from a natural-language profile remains challenging. Explicit reasoning can help models deliberate before responding, but generic reasoning does not necessarily preserve character fidelity. We therefore propose Psy-CoT, a psychology-grounded chain-of-thought framework that decomposes pre-response reasoning into three role-specific steps: Interaction Perception, Psychological Empathy, and Logical Construction, so that the model thinks dynamically from the profile rather than merely mimicking surface patterns. While structured reasoning provides a foundation, it alone is insufficient; reinforcement learning is essential to further align the model with character fidelity. However, we observe that under LLM-based reward models, both generic phrases that hack the reward model and genuinely role-specific phrases receive identical gradient signals. This hacking accumulates over training, misleading the model into treating both as equally optimal choices. To address this, we propose Role-Aware Policy Optimization (RAPO), which uses profile-induced entropy differences to weight gradients asymmetrically, amplifying role-specific tokens under positive advantage while attenuating them under negative advantage. Experiments on CoSER, CharacterBench, and CharacterEval demonstrate that Psy-CoT outperforms existing role-playing CoT methods, and RAPO consistently surpasses GRPO across multiple model scales.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2611.14009

Coding Agent Memory Post-Training II: Unlocking the Memory Potential of Pre-trained File Operations for Long-Horizon Tasks via On-Policy Distillation

Lirui Luo, Kelong Mao, Xinwei Yang, Rongqing Li, Luyu Chen, Kieran Wong, Yudong Guo, Xinrui Wang 等 13 位作者

原文摘要

Language-model agents increasingly tackle long-horizon tasks whose interaction histories exceed the model's active context. They must therefore decide what to write down and what to read back, and when to do either. Recent work has begun to use on-policy distillation (OPD) to make memory control part of the policy, relying on memory tools and compaction protocols that lie outside either model's pre-training. However, this weakens the distillation on both sides: the teacher cannot give useful guidance on an interface it has never used, and the student samples too few meaningful memory actions for that guidance to take effect. To study this problem, we introduce Coding Agent Memory Gym II (CAMG-II), a suite of long-horizon agentic-RL environments spanning Shop, Coding, AutoResearch, General, and Visual. Alongside each environment's native task interface, CAMG-II provides ordinary Bash access and an episode-persistent workspace, enabling the policy to create, revise, search, and reuse files as memory across context replacements. We also introduce Coding Agent Memory Post-Training II (CAMP-II), which trains a single policy jointly across all five environments with fully asynchronous OPD to learn this file-based memory behavior directly from a frozen teacher with a prior for file operations. On CAMG-II, CAMP-II obtains 2.97 times the gain of the strongest OPD baseline. The memory behavior CAMP-II learns generalizes to two external benchmarks, Terminal-Bench 2.1 and SWE-Bench Pro V2, with improved success rates.

cs.LG · cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2608.23670

Automata from Agent Traces: Failure and Next-Step Prediction

Seonglae Cho, Franklin Cardenoso Fernandez, Umar Mohammed, Zekun Wu, Kleyton Da Costa, Ilham Wicaksono, Adriano Koshiyama

原文摘要

LLM-based agents execute multi-step tasks, but their behavioral structure remains opaque: long unstructured traces resist the safety auditing and runtime monitoring that deployment requires. Existing approaches operate per-trace or success-only, so they miss the cross-run topology that links next-step and failure prediction. To recover that shared structure, we collapse an entire trace corpus into a single, compact finite-state machine (FSM) that serves as a structural substrate for the otherwise unpredictable behavior of LLM agents. Across twelve public datasets, the FSMs are compact (7-43 states), replay held-out data at >=0.997 fitness with near-identical topology across splits, and build in milliseconds. This substrate addresses both prediction goals. For next-step prediction, FSM-state context outperforms Agent Workflow Memory on every ground-truth-matched dataset. For failure prediction, per-state behavioral features reach held-out AUROC up to 0.94, and an online monitor ranks failing runs above passing ones from a partial trace, triggering early stopping well before completion. Behavioral topology thus appears shaped more by the deployment harness than by the LLM, providing a model-agnostic structural primitive for safety auditing and runtime monitoring.

cs.AI · cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16375

Recursive State-Augmented Language Models: Persistent Latent-State Adaptation of Pretrained Language Models

Omer Haq

原文摘要

Language models rely on finite token contexts, limiting access to earlier interactions. We introduce Recursive State-Augmented Language Models (Recursive SALM), a framework in which a persistent latent state evolves with each observation and generates adaptations of a pretrained language model. A modular read-transition-write architecture separates observation encoding, latent dynamics, and predictor adaptation. We develop deterministic and stochastic formulations. The experiments evaluate only deterministic SALM with frozen GPT-2 small and exclude earlier evidence from its decoder context. On a four-turn hypothesis-inference task, latent-gated LoRA reaches 89.1% accuracy, comparable to 88.0% for a trained full-context LoRA baseline, with earlier evidence available solely through a 64-dimensional state. Zeroing or shuffling the state substantially reduces accuracy. In a two-turn task with codewords unseen in training, several interfaces improve candidate ranking over chance, but rarely generate the exact codeword under the evaluated training schedule. These experiments show that persistent latent-conditioned adaptation can support sequential inference when earlier evidence is absent from the decoder context.

cs.LG · cs.AI · cs.CL · stat.ML

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2611.08439

Label-Wording Augmentation in a Typed-Decision Encoder: Gains on Unfamiliar Words, Not on Abbreviations or Codes

Vladimir Beskorovainyi

原文摘要

Typed-decision models score options that the caller supplies at request time, so their accuracy depends on label wording that the model's developer does not control. We compare two releases of a 307M-parameter multilingual cross-encoder, Ru-Jev, trained on thirteen Russian-language tasks; the later release adds an augmentation that presents labels in several surface forms. Holding questions fixed and varying only how eighteen routing labels are worded, we evaluate ten label vocabularies at five and eighteen candidates, checked for overlap with the training data. On held-out vocabularies made of ordinary or borrowed words (jargon, office slang, department names), the augmented release has higher point accuracy in all fourteen cells, by up to 0.24. Several of these gains rest partly on one label that the earlier release almost never chose; fitting one bias per option removes part of them, while the jargon gain largely survives. On abbreviations and opaque codes the augmented release gives no usable gain: its loss on 17 abbreviations unrelated to training (-0.060) is one label and disappears after the same correction. The comparison is between two trained releases, not a controlled ablation, and the earlier recipe varies across seeds by up to 0.098 on the strictest slice. On two datasets neither release saw, accuracy moves by -8.4 and +5.6 points, narrowed to -3.4 and +1.4 by the per-option biases; on a third, scientific rubrics, the augmented release is ahead of both runs of the earlier recipe after the correction. Rankings of ten training checkpoints disagree across the three unseen datasets, so none of them should choose the checkpoint. We also report four evaluation hazards from comparing the two releases.

cs.CL · cs.AI · cs.LG

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2606.21666

A Contradiction Judge Counted Reported Teammate Errors as Hallucinations: An Assertion-Aware Re-Score of One Multi-Agent LLM Harness

Carson Rodrigues

原文摘要

LLM judges that score hallucination flag any statement that contradicts a reference fact. Agents in multi-agent LLM systems quote and dispute each other's claims, so such a judge can count a correct report of a teammate's error as a hallucination. An earlier version of this study, a preprint (arXiv:2606.21666v1), reported that broadcasting raised a hallucination rate from 0.492 to 0.658 (+34%) and called it contamination; this version withdraws that claim, and the preprint's runs cannot be re-scored. An unpublished revision of that harness reported 0.289 to 0.554 (+0.265, 95% CI [0.225, 0.305]), and its saved outputs are re-scored here. Under an analysis plan fixed before any judge call, three judges from three vendors labeled each of six facts in 1,832 outputs as asserted wrong, attributed to another source, asserted correct or absent. On values that outputs assert as their own, the broadcast effect was -0.0208 (95% CI [-0.0458, +0.0041]) on 10 Full-Broadcast trials: not distinguishable from zero, and equivalent to zero only within a post hoc margin of +/-0.0663 and under the plan's prompt. On these LLM-judge labels, reports of a teammate's wrong value accounted for nearly all of the Haiku travel effect. Scored for mentions, the judges reproduced the effect (+0.1996). With Llama agents, 2 of 40 teammate outputs asserted the wrong destination. One of twelve Claude judge-prompt combinations gave an asserted increase (Sonnet 4.6, one paraphrase: +0.0246, 95% CI [+0.0055, +0.0436]). In 40 new trials per arm, judged by Claude models like the agents, the asserted difference was -0.0505 when both judges agreed and +0.0161 (95% CI [+0.0024, +0.0299]) when either did. Main limit: the labels have no human validation; human labels for 120 outputs are pending. The outputs come from one harness and 13 hand-written scenarios, so the shares do not transfer.

cs.AI · cs.CL · cs.MA

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16359

IsharaBART: An efficient BART-Based Framework for Low-Resource Bangla Sign Language Translation

Aonmoy Das, Ananna Dev Aishi, Md. Fazlul Kader

原文摘要

Millions of Bangla Sign Language (BdSL) users lack reliable automatic translation tools. Existing systems are limited by data scarcity, unidirectional design, and computationally expensive architectures that are not suitable for low-resource deployment. We introduce IsharaBART, a bidirectional generative framework for gloss-sentence translation in BdSL. IsharaBART uses Bidirectional and Auto-Regressive Transformers (BART)-based conditional language modeling with denoising learning objectives. We developed a parallel dataset of 1,000 professionally annotated gloss-sentence pairs, balanced across tenses and structures. We expanded the dataset to 4,831 pairs through four augmentation techniques: rule-based tense transformations, masked token substitutions, back-translation, and one-shot prompting. A compact denoising sequence-to-sequence model was pre-trained from scratch on monolingual Bangla text from six different corpora, then fine-tuned on the annotated and augmented datasets. We compare IsharaBART against four larger baseline models under identical conditions, where it achieves comparable or superior bidirectional performance. On the gloss-to-sentence (G2S) task, it attains Bilingual Evaluation Understudy (BLEU-4) score up to 51.7, and on the sentence-to-gloss (S2G) task, it achieves 61.40, outperforming multilingual models despite having fewer than 10% of the parameters. Ablation studies demonstrate that pre-training provides 62.5-90.4% faster convergence and 11.5-118% higher BLEU-4 than scratch. Low-rank adaptation (LoRA) reduces trainable parameters by 99.3%, but with 28.0-52.2% loss of BLEU-4 compared to full-parameter fine-tuning.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2606.31272

The Decomposition Is the Fingerprint: Per-Component Identity for Agent Skills

Hongliang Liu, Yuhao Wu, Tung-Ling Li

原文摘要

AI agents increasingly acquire and execute skills at runtime: bundles of prompt instructions, executable code, and tool declarations fetched from marketplaces and other agents. Governing them needs a stable notion of skill identity, yet cryptographic hashing is engineered to destroy the very similarity we need, as a one-character edit scrambles the digest. We present a compact, locality-sensitive fingerprint that embeds each component of a skill and projects it to bits with a multi-bank SimHash, giving a fixed 120-byte signature compared in constant time by Hamming distance. Our central claim is that keeping the fingerprint as a per-component triple (prompt, code, tools), rather than a single score, is what makes it useful: the triple recovers skill-family identity through paraphrase, renaming, refactoring, and controlled code translation when another component remains shared, while independent multilingual reimplementation is not recovered; it also localizes which component carries the reuse. We claim lineage, not behavioral equivalence: identity supplies the structural axis of a registry and leaves safety to behavioral verification. The fingerprint reaches an area under the ROC curve (AUC) of 0.974 (95% CI [0.956, 0.994]) over 4,950 pairwise comparisons while using 77x fewer bits than the embedding it approximates, with ranking preserved in expectation and finite-bit concentration; the per-component split turns one number into relationship classification, families, novelty, and a portable "SkillBOM" for a skill registry. On a 906-skill injection benchmark the fingerprint recognizes injected skills as tampered copies of a known base and localizes the change, but recognition is not trust: it remains, by design, an identity signal complementary to behavioral verification rather than a safety verdict.

cs.CR · cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16323

Does Deep Knowledge Tracing Learn Semantic Relations Between Knowledge Components?

Kia Karbasi, Mahdi Saiedi, Kevin Hong, Gregory Pottie

原文摘要

Deep Knowledge Tracing (DKT) is a knowledge tracing (KT) model that updates a shared latent state from sequences of interactions, each pairing a knowledge component (KC) with a response. An interaction with one KC can therefore affect predictions for other KCs. We ask: Are the KC relations learned by DKT semantic? We construct text-derived semantic KC representations using nine pipelines to test whether the conclusions hold across alternative ways of constructing those representations. The first test compares DKT-derived KC relation graphs with semantic relation structures induced by pairwise cosine similarities among the KC representations. Across the evaluated datasets, embedding models, and pipelines, the two structures align only weakly and inconsistently. The second test introduces semantic information into DKT; the pretrained text-derived representations do not consistently outperform the standard DKT baseline. In the reduced-data experiments, semantic initialization yields a higher test area under the ROC curve (AUC) than the baseline at eight of nine nested sample sizes for ASSISTments 2017. KC-supervised fine-tuned BERT produces small positive AUC differences in some matched comparisons, but the study does not establish their stability across runs. These effects remain limited and inconsistent. The findings suggest that behavioral, temporal, structural, or psychometric signals may also shape DKT's learned relations. DKT-derived relations are not primarily semantic: the tested text-derived semantic structures do not strongly recover the DKT-derived relation structure.

cs.LG · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16321

From Large Language Models to Efficient System One Models via Adaptive Computation

Minh-Anh Nguyen

原文摘要

Transferring LLMs into System One decision models typically replaces the language modeling head with a decision head while retaining the full backbone, making every decision pay for all layers. Yet it remains unclear how much depth such decisions require. We find that decision information is concentrated in intermediate layers, with early layers building essential features and the final six layers removable without loss, while independently trained smaller models largely mimic the full model rather than developing specialized decision capabilities. These findings suggest that efficiency requires not simply smaller backbones, but shared computation across different depths. Building on this, we propose \method{}, which jointly trains a full model with prefix sub-models that share its final layers and decision head through self-distillation. A lightweight gate reads the hidden state at each exit to determine whether additional computation is needed, allowing one model to operate at multiple layers. Across seven decision benchmarks, our approach matches or outperforms independently trained models at every depth, while achieving better accuracy-efficiency trade-offs through adaptive computation.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.16289

Eliciting and Mitigating Specification Satisficing in Coding Agents

Ruixin Wang, Tianyi Zhang

原文摘要

Coding agents often fail to meet all requirements in the task description. In some of these failures, the agent considers an implementation choice that meets a requirement, but later decides to commit to another one that does not meet the requirement for a reason it finds acceptable. We call this specification satisficing. It is challenging to identify and mitigate specification satisficing because the agent accepts such choices for its own reasons, and existing self-correction methods are biased toward them. We propose SatisFix, which guides a coding agent to report these choices and their reasons, and then check each choice against the task description. We evaluate SatisFix on DeepSWE and SWE-bench Pro with five LLMs as backbones and two different agent frameworks, and compare it with three self-correction methods. SatisFix improves the pass@1 of the base agent by an average of 5.09% on DeepSWE and 7.06% on the hard subset of SWE-bench Pro. On DeepSWE, it outperforms the best baseline by 3.98%. Our ablation study shows that both steps of SatisFix contribute to the improvement.

cs.SE · cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗