arXiv_CL
-
Learning What to Ask for Tomorrow: Active Preference Acquisition for Long-Term Personalized Web Agents
2026-10-22 11:16:24Yiqi Wang, Zizhi Zeng, Jiaqi Zhang, Yiqun Duan, Mingkai Zheng, Taotao CaiAbstract
Web agents increasingly serve the same user across repeated online tasks such as booking travel, shopping, and scheduling, so their long-term usefulness depends on how well they know the user's preferences. Passive memory and personalization methods cannot obtain preferences the user never reveals, and clarification methods that value a question by its benefit to the current task overlook the benefit of reusing the answer in later tasks. We study Active Preference Acquisition (APA): which missing preferences are worth the cost of asking? We define the Prospective Personalization Value (PPV) of a missing preference as its expected cumulative utility gain over future tasks minus the effort and disclosure-risk cost of asking, and select questions under a per-interaction budget. We characterize when cumulative reuse justifies acquisition despite insufficient single-task benefit, and evaluate this regime with an initial-profile diagnostic. APA computes the counterfactual gain of each candidate preference on public task prototypes with the agent's own task solver and aggregates it by the expected reuse of each task family. We evaluate it on ActiveWebPref, a controlled longitudinal benchmark with hidden but queryable user profiles and exact task utilities. At a horizon of 20 tasks, 34.4% of budget-feasible valuable missing preferences are amortization-dependent; APA acquires 80.3% of them, whereas current-task value of information acquires 5.1%. Under per-user asking-cost caps matched to APA's realized expenditure, APA achieves higher net utility than four heuristic selection baselines. In our cost setting, asking about every unknown preference yields lower net utility than not asking, while APA recovers 84.2% of its utility gain with 27.6% of its questions. The value of acquiring a user preference should account for its reuse across future tasks, not only its benefit to the task at hand.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31107
PDF
-
RIFT: KV Cache Compression for Draft Models with Complementary Compressibility Training
2026-10-22 11:12:23Dor Tsur, Maor Ashkenazi, Maximilian Golub, Peter Hicks, Yaniv Galron, Talor Abramovich, Yoav Miron, Tiyasa Mitra, Bita Darvish Rouhani, Ran ZilbersteinAbstract
Speculative decoding is a popular decode acceleration method. As such, it is particularly relevant to long-output generation, which requires an increasingly large KV cache. KV cache compression offers a way to constrain costs in long-context drafting. However, compression methods that focus on prefill information can lose relevance as the output length grows. We term this phenomenon decode staleness and show that it degrades draft model accuracy under selection-based KV cache compression. We introduce RIFT, a family of selection-based KV cache compression strategies that balance prefill information and recent generated tokens. We show that refreshing the KV cache selection during token generation significantly improves the drafter's quality at the cost of additional decode latency. To alleviate the refresh costs, we propose an adaptive KV refresh strategy that reduces the number of score calculations while retaining the drafter's quality. We further demonstrate how a compressibility-driven drafter training phase, based on sparse attention masks, can improve the drafter's quality under a compressed KV cache. Combined with compressibility training, RIFT enables Qwen3.5-4B, drafting for Qwen3.5-122B, to retain over 98% of its full-cache acceptance length on 32K-token inputs, 4K-token outputs, and an 8K-token KV cache budget.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31098
PDF
-
LAMB: A Modern Contextual Language Model for Latin
Abstract
This article introduces Latin Modern BERT (LAMB), an encoder for the Latin language based on the ModernBERT architecture, pretrained in two phases on more than 74B Latin tokens using a novel morphologically informed masking regimen. We also present two corpora for the Latin language. Acervus contains 24B BERT WordPiece tokens, including over 760K samples from heavily damaged OCR transcriptions, repaired by a ByT5 model trained for Latin text repair. The second dataset, Corpus Trinum, contains a mixture of over 100M strong and weak Latin language semantic triplets. Corpus Trinum is the first Latin semantic triplet dataset, opening the door for Latin language models to be adapted to general sentence semantic similarity tasks. We construct an experimental grid using five tasks and seventeen subtasks, applying a variety of finetuning configurations and subword aggregation techniques for 15 pretrained encoders, 6 LAMB training checkpoints, and an ablation model designed to quantify the impact of different masking regimens. Using this benchmark, we rank Latin encoders and find that LAMB outperforms all other evaluated models on 16 of 17 subtasks, coming within 0.8 percentage points of state of the art on the seventeenth. We provide additional analyses relating Latin encoder performance to tokenizer behavior, syntactic structure, finetuning configuration, pretraining state, subword aggregation, and variation in source texts and domains. In these, we identify unique acquisition dynamics for Latin open class words during terminal decay phases, and establish that for some task categories, finetuning configuration selection can have a greater effect on downstream scores than additional pretraining tokens.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31723
PDF
-
What Pairwise Evaluation Misses: Reward Scores across Prompts
Abstract
Reward models score responses for preference optimization, filtering and ranking, and pooling scores across prompts needs them to be comparable. Yet pairwise evaluation compares two responses to the same prompt, so any score component shared by a prompt's responses cancels and goes untested. We study this component by comparing between-prompt variation in model scores with the same variation in human ratings of the same responses. We propose a human-referenced audit, built from a between-prompt variance share, a sibling-rating regression and prompt substitutions, and apply it to eighteen released scorers and six matched training conditions. We find that scores cluster by prompt more than the ratings do: on HelpSteer2, six reward models have between-prompt shares of 0.34 to 0.64 against a human share near zero. Replacing the original prompt with a fixed instruction keeps substantial clustering but weakens agreement with prompt-level human ratings, so clustering alone does not show agreement with humans. Part of that variation is useful: subtracting prompt means lowers the helpfulness of globally selected responses, and scores stay sensitive to a wrong prompt even when training shuffles prompt means. We propose that reward evaluation check between-prompt variation, sensitivity to a wrong prompt and usefulness for selection separately.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31090
PDF
-
SCOPE: Compiling Natural-Language Policies into Neurosymbolic Predicate Programs
2026-10-22 10:41:13Arnon Mazza, Elad LeviAbstract
Policy enforcement for LLM agents has shifted from single-turn content moderation to compound, application-specific policies evaluated over multi-turn transcripts and tool-call trajectories. Fine-tuned guardrail models are efficient but degrade on compound policies, and while LLM-as-a-judge approaches achieve higher accuracy, they incur high inference cost and tend to produce unstable predictions. Neurosymbolic methods offer a natural alternative, decomposing the policy into narrow neural decisions composed under explicit logic. However, existing symbolic layers lack the required expressiveness and are optimized against fixed data rather than the policy itself. To address this, we introduce SCOPE, which compiles a policy into an executable program of neural predicates executed by small models, deterministic extractors, and code that composes their outputs. The program is optimized jointly with its training data, using adversarial generation of counterexamples that target where the program deviates from the policy. We evaluate SCOPE on a benchmark of real-world policy tasks against prompt optimization, task decomposition, and workflow generation methods, achieving the highest average accuracy across tasks and models. Programs optimized with the weakest executor transfer to other models without further optimization. This allows them to be deployed in production on inexpensive models, where they reach accuracy comparable to frontier reasoning models. Unlike these monolithic models, SCOPE produces substantially more stable predictions, and each verdict is interpretable through the predicate outputs that produced it. Additionally, we validate the key components of our method through an ablation study. SCOPE thus enables reliable, interpretable, and efficient policy enforcement at production scale.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31062
PDF
-
Harm Is Behavioural, Moderation Is Lexical: How Language and Culture Decide Whom Automated Safety Systems Protect and Whom They Silence
2026-10-22 10:24:17Gabriel Lars Sabadin, Dr Nicola Harding, Sarah KingstonAbstract
Content warning: this paper quotes slurs and hateful statements from published test suites; the racial slur among them is masked. Automated systems decide at scale which messages are harmful, and they decide unevenly. Classifiers label African American English as offensive up to twice as often as other English, the accounts of prominent drag queens can score as more toxic than those of white nationalists, and safety behaviour that holds in English collapses in low-resource languages. We argue that over-flagging and under-protection share one cause: deciding harm from vocabulary and surface form rather than from what a message does. We organise the evidence into seven failure modes and argue, as testable hypotheses, that harmful behaviours (a request for secrecy, a demand for payment, an attempt to isolate someone) travel across languages better than the words that express them. We propose an evaluation protocol that makes the gap between the best- and worst-served language communities a headline result, and show that the naive gap is inflated under equal service and that its usual bootstrap interval is unreliable. Applied to five open systems in eleven languages, from a keyword lexicon to two current multilingual guard models, the protocol finds that none reaches native-speaker accuracy at separating harmful function from harmless form (51 of 55 system-language pairs fall short); that greater capability raises discrimination in nine languages while leaving Arabic and romanised Hindi near chance, so the gap between the best- and worst-served languages grows with capability; that one system both over-flags and under-protects the same language most; and that a language a system reads worst can look the most fairly treated when only false positives are reported.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31047
PDF
-
Chunk Recomputation under Real Document Revisions: Measuring Update Cost and Retrieval Quality of Fixed, Recursive, and Content-Defined Chunking
2026-10-22 09:54:17Yuning Bian, Lele Xu, Yifan Yuan, Xinyu Yin, Chengyan ZhanAbstract
Chunking strategies for retrieval-augmented generation are almost always evaluated on a static snapshot, yet the documents underlying a knowledge base continue to change. Because chunks are embedded independently, a boundary that moves after an edit forces unchanged text to be re-embedded. We measure this update cost on 120 pairs of adjacent Chinese and English Wikipedia revisions for three stateless chunkers (non-overlapping fixed-size, recursive paragraph-first, and content-defined chunking) under a strict byte-identical reuse rule and a frozen dense encoder. On the 116-article common-success pool, with frozen, length-calibrated parameters, fixed-size chunking recomputes 57-68% of its full new-version embedding-input workload, recursive chunking 38-56%, and content-defined chunking 43-54% (it failed on 4 of 62 Chinese articles), while at these parameters mean evidence recall under a fixed 2048-token retrieval budget differs by less than two percentage points within either language. A cost-only parameter grid shows that the ranking of the two content-aware chunkers changes with chunk-length settings, whereas non-overlapping fixed-size chunking has the highest recomputation cost in every length-matched condition. Controlled edits and a verbatim-presence diagnostic support boundary drift as one contributor to recomputation of text already present in the previous version. Update cost is therefore a substantial, parameter-sensitive dimension that static retrieval evaluations do not expose and that should be reported alongside retrieval quality.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31019
PDF
-
Mostly Reading: What Open-Book Scores Measure in Document-Level LoRA Adapters
2026-10-22 09:52:29Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Volker TrespAbstract
Single-document LoRA adapters have been proposed as units of knowledge to be merged and retrieved; yet open-book scoring, with the document in the prompt, can credit a document's own adapter (its specialist) for reading that document as readily as for holding it. Training one adapter for each of 150 Wikipedia articles on Llama-3.1-8B-Instruct, we find that open-book scores mostly measure a reading behaviour every adapter shares: a random adapter trained on another document recovers 82 % of the specialist's open-book advantage over the base model but only 2 % of its advantage on a cloze probe that asks for the same document's facts without providing the source text, a share we cannot distinguish from zero. Content does reach the weights, since the specialists beat the base model in cloze recall by 0.094, and the geometry of their updates follows the documents they read; yet, in a test of 30 queries, adapters chosen by weight-space proximity or text embedding are not distinguishable from random ones on either probe, although that test cannot exclude an advantage of about four fifths of what a query's own adapter gains over a random one. At 16 documents or more, correctly scaled merges recover 87 % to 89 % of the specialists' open-book advantage but only 12 % to 16 % of their cloze advantage, with intervals that run from below zero to at most 32 %, against 94 % for joint this http URL also report four silent errors, each of which produced a clean, wrong result.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31017
PDF
-
Padding to k: A Minimum Count in the Schema Makes Open-Weight Models Pad Their Extraction Lists
Abstract
Extraction pipelines ask language models for JSON that follows a schema, and the schema can set a minimum number of list entries. We ask what models do when the document holds fewer items than that minimum. Under a preregistered plan, we test eight open-weight models on 558 real documents from nine annotated datasets, with six more models on hosted services as a reference. Three findings stand out. First, models treat the count as a target: a minimum shown in the schema, but not enforced, raises the share of lists longer than the true count from 34.5% to 55.4%. Second, the padding is hard to catch, because most added entries are real text taken from the document; about half of them pass a check that each entry occurs in the source. Third, common fixes help only partly: a line saying fewer items are fine removes the effect on average but misses more true items, a count in prose ("up to k") also raises over-counting, an "N/A" option under an enforced minimum still leaves half the lists padded, and re-asking with the validation error makes padding far more common. We release the code, answer lists and run records at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.30997
PDF
-
ColaBench: Evaluating Memory-Guided Knowledge Acquisition in Personal Agents
2026-10-22 09:34:52Wenjie Zhang, Zhichao Shi, Xuhui Jiang, Xiaojun Wu, Cehao Yang, Chengjin Xu, Xiaoning SongAbstract
Personal agents need to draw on long-term interaction history to retrieve and apply external knowledge suited to a user's current request. Even an answer grounded in relevant documents can be unsuitable when it overlooks the user's circumstances. This creates two connected challenges: using historical conditions to identify the knowledge needed for a new request, and applying that knowledge to the user's situation. Existing benchmarks advance memory recall, personalized retrieval, and memory-guided action, while the joint evaluation of these two challenges remains underexplored. We introduce ColaBench, a synthetic benchmark in health and daily living, with 349 tasks across 10 Profiles and a knowledge base of 13,033 documents. Its Evidence Graph connects personal conditions to their disclosures across public interaction sessions, records prior document exposure, and links answer requirements to personal and newly acquired document evidence. We also introduce MemKB, which uses state summaries to inform queries and rule application, and memory-document links to retrieve knowledge associated with the user's history. MemKB achieves 86.25% Accuracy with an average of 4.07 interaction rounds per task. Task-level analysis finds larger gains on tasks integrating information across multiple historical Facts; error analysis further identifies unsupported personal assumptions and rule misapplication as obstacles to using retrieved evidence. ColaBench provides traceable tasks and an evaluation framework for studying how memory guides knowledge acquisition and rule application.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30989
PDF
-
Interpretive Variation and Discourse Structure in Detecting Enthymemes in Controversial Political Discourse
Abstract
Enthymemes are arguments that leave a premise or conclusion unstated, and they are a common persuasion device on social media. Because readers must fill in the missing part themselves, they often disagree on whether a text contains one. This paper asks how to detect and evaluate enthymemes when such disagreement is part of the task. On tweets about vaccination and immigration, each labeled by five annotators, F1 stalls around 0.67 for every fine-tuned model we train, regardless of data size, added features or label type. We therefore ask how these models should be evaluated instead, and propose metrics that compare predictions with the full distribution of annotator votes, together with a model that predicts this distribution directly and captures disagreement better at no cost in F1. We then ask what role discourse structure plays. Models learn to rely on explicit argumentative structure: they reliably recognize tweets without an enthymeme but rarely commit on tweets that contain one. Both the models and a discourse parser miss arguments whose links are left implicit and must be inferred by the reader. Finally, we discuss what these blind spots mean for content moderation systems trained on majority labels.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30987
PDF
-
A Reality Check of Inference-Time Context Augmentation Methods
2026-10-22 09:31:18Haeun Yu, Geert Heyman, Pepa Atanasova, Isabelle AugensteinAbstract
Large language models (LLMs) often fail to utilise the context given in the input, resulting in hallucinations or a fallback to parametric knowledge. Context augmentation methods aim to improve model performance by modifying the input, from inserting filler tokens that add computational time to adding relevant content such as elaboration of a context during inference. However, these methods have been evaluated in isolation and only by their performance gains, leaving it unclear how they compare and whether they reliably improve performance. In this paper, we present an evaluation framework spanning five methods that differ in how informative the inserted content is and apply them at the sentence level across seven models spanning multiple sizes on three datasets. We find that accuracy gains are not always genuine improvements but undirected perturbations: the methods correct some predictions while breaking others, particularly when the inserted text allows only for more computational time. Our analysis reveals that methods that bring no additional content primarily change model predictions near the decision boundary. Our findings highlight the need to evaluate the reliability of the accuracy gains before adopting them in practice.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30983
PDF
-
Joint or Per-Option? Compact Encoders for Closed-Set Decisions from 17 Million to 1 Billion Parameters
Abstract
Many language-model pipelines end in a closed-set decision, such as selecting a tool. Prompting a large language model for it is slow and order-sensitive. We ask how small an encoder for such decisions can be and whether it should read the options jointly or separately. In a three-seed study at 32 million parameters, per-option scoring beats joint scoring on a held-out sample of a 37-benchmark decision suite by 5.8 points of chance-corrected skill at matched optimizer steps and by 2.8 at approximately matched training compute; two external benchmarks agree, though not significantly on their unseen tasks. Crossing reader with initialisation shows that the advantage requires a reranker initialisation, whose query-candidate pair input matches per-option reading; from the raw encoder the two readers are indistinguishable. Per-option scoring needs 5 to 13 times more encoder passes: at most twice the median latency, but more than four times the mean. Of six released joint encoders (17 million to 1 billion parameters), the 68- and 150-million-parameter models exceed a 435-million-parameter DeBERTa-v3 zero-shot classifier by 4.9 and 5.9 accuracy points on 209 overlap-audited System1-Arena tasks and are at least on par with it on the 145 tasks unseen in training. The smallest answers a median request in 14 milliseconds on 16 processor threads, against 728 for the classifier. Reordering options changes 12 to 23% of the joint models' answers (per-option scoring: none). Up to 68 million parameters, they lead their size range on a public leaderboard's private, mainly in-distribution tests.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30970
PDF
-
Once a Tree, Always a Tree? Cross-lingual Transfer of Hierarchical Generalization in Language Models
2026-10-22 09:07:44Shin Yamaguchi, Masaya Taniguchi, Tatsuki Kuribayashi, Taro Watanabe, Yoshinobu KanoAbstract
Language models (LMs) often exhibit successful cross-lingual transfer through continued (pre)training, still their underlying process remains unclear. One possibility is that LMs can leverage the evidence toward language-universal principles, namely, hierarchical structure, from a source language (L1) to a new language (L2), while many other non-linguistic explanations, e.g., improved initialization of position embeddings, can be potentially possible. In this study, to quantify the cross-linguistic transferability of linguistic evidence in cross-lingual LM training, we introduce a new experimental design, namely, a cross-lingual poverty-of-stimulus (CPOS) setting that ablates how LMs' hierarchical generalization in L2 can be led by structural evidence provided in L1. We contrast counterfactual L1 and L2 learning scenarios where the L2 corpus is ambiguous between linear and hierarchical grammar rules, while the amount of positive evidence for the hierarchical rule is controlled on the L1 side. Our experimental results show that LMs can partially, yet not perfectly, transfer L1 structural evidence to L2, yielding only a small but detectable hierarchical bias, while the strength of transfer depends on language pairs. We further ablate the conditions for successful transfer, motivated by recent reports on the substantial impact of vocabulary overlap. We indeed find that vocabulary overlap is crucial even for the cross-lingual transfer of hierarchical syntactic generalization. This suggests that this also explains the language-pair-dependent tendencies in the main experiments and suggests that the transfer of linguistic evidence in continued new-language learning is only conditionally achievable.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30960
PDF
-
IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems
2026-10-22 09:01:29A. Jian Yang, An Huang, Bo Dan, Chen Mu, Chengran Yang, Chuan Shi, Feng Chang, Hongxin Li, Hongyuan Tao, Jinxing Zhang, Jujie He, Kun Wang, Lin Jing, Ming Yang, Qicheng Hu, Renyuan Li, Shark Liu, Shawn Guo, Shi Qing, Shi Zhong, Tianyu Zheng, Wei Liu, Wei Zhang, Wenrui Cai, Wenxiao Chen, Xuxing Chen, Yan Xing, Yi Xing, Yifan Luo, Yiran Xing, Yizhi Li, Yongsheng Kang, Yucheng Ding, Yue Song, Yun Qing, Yunqi Liu, Zelong Huang, Zhengmao Ye, Zhentao Fan, Zhiteng Li, Ziling Huang, Bihong Tang, Bingli Wang, Ce Jun, Changhao Chen, Chen Hu, Chen Xiao, Chi Liu, Chunjiang Ge, Danyang Li, Deyang Kong, Dian Qi, Ding Tan, Evan Yang, Fanglin Xu, Fei Bai, Gang Liu, Guochang Li, Hao Tang, Haoran Feng, Haowen Wang, Huahao Hu, Huanjin Yao, Huatong Song, Jia Deng, Jiajun Wu, Jialiang Cheng, Jiayi Liang, Jiayi Wu, Jie Deng, Jiejun Tan, Jing Yang, Jingda Xu, Junhang Cheng, Junxiao Liu, Kai Lv, Kaixiang Pan, Kecheng Lu, Kun Zeng, Kun Zheng, Kunbo Ding, Liang Qiu, Lingyi Sun, Luan Zhang, Mengxiao Lv, Ning Dong, Pei Wang, Peng Zhang, Qingsong Cai, Qixiang Yin, Qiyao Wang, Renbiao Liu, Renfei Dang, Ruihang Li, Ruirui Huang, Shanhui Yin, Shen Jian, Shengjie Fang, Shikai Li, Shinan Zhang, Shiqi Chen, Shuai Liu, Shuangqi Li, Shuo Tang, Shuqing Deng, Sijin Hao, Siwei Wu, Siyue Wang, Tao Hai, Tao Ren, TingYang Chen, Wei Shao, Weicheng Gu, Weihan Peng, Wenlong Miao, Xavier Yu, Xiangyuan Ji, Xiaocheng Li, Xiaodong Huang, Xiaohan Ge, Xiaoxiang Shi, Xiaoyang Zhang, Xiaoyu Yang, Xidong Wang, Xin Li, Xinyu Tang, Xiyuan Yang, Xudong Ma, Ya Tong, Yachao Wang, Yang Tian, Yaxin Du, Yeli Ruan, Yi Wang, Yi Yu, Yibin Feng, Yibo Tan, Yihang Wang, Yiming Ju, Yinda Chen, Yinda Chen, Yingjie Du, Yiwei Liu, Yizhi Zhou, Yuan Wei, Yuli Chen, Yuqi Zhu, Yuwen Li, Yuxi Dong, Yuxuan Zhang, Yuyang Song, Zhange Zhang, Zhaohan Zhang, Zhaoyi Li, Zhen Zhang, Zhengbo Zhang, Zhihang Hu, Zhiwei Yu, Zhixin Xie, Zhiyu Chen, Zhizhao Luo, Zhongyuan Wu, Zhou Fang, Zhu Lin, Zhuocheng Gong, Zixi Zhao, Ziyang Zeng, Ziyi Wang, Ziyu Zhao, Ran Tao, Mingjie Tang, Chuan Hao, Peihao Wu, Jianzhou Wang, Nan Jiang, Yuan Zhi, Bryan DaiAbstract
We present IQuest-Q1, a sparse Mixture-of-Experts (MoE) agentic foundation model with 320B total parameters (15B activated). IQuest-Q1 is developed through a three-stage process: (1) stable and efficient pre-training, combining Stable SplitNorm, Muon, and muP with a two-stage curriculum whose second stage increases the coding-data mixture, preserving a wide stability margin between safe and optimal learning rates; (2) fine-grained mid-training, expanding the data mixture across domains and formats and progressively extending the context; and (3) multi-scaffold with multi-version post-training, combining SFT, domain-specialized RL, and MOPD across various scaffolds to integrate complementary training signals and strengthen agentic coding capabilities. Furthermore, model merging combines different experts and checkpoints from different stages while mitigating conflicts among their parameter updates. The resulting IQuest-Q1 demonstrates capabilities in both coding and general tasks. We release IQuest-Q1 as an open-source model to support community research and development.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30949
PDF
-
RetiJev: Fast and Source-Grounded Ophthalmic Decision Models
2026-10-22 08:59:47Bo Liu, Lianyu Wang, Meng Wang, Ching-Yu Cheng, Huazhu FuAbstract
Ophthalmic clinical decision support requires fast, accurate, and privacy-preserving inference, yet autoregressive large language models (LLMs) incur substantial latency and computational overhead for decisions among predefined actions. We therefore investigate a decision-oriented alternative to generative reasoning. Specifically, we first introduce RetiJev-Data, a source-grounded ophthalmology corpus comprising 78,158 clinical-decision and diagnosis questions derived from articles in PubMed Central and a MeSH vocabulary covering 531 eye diseases, and 50,074 guideline-grading questions spanning 11 grading systems, along with a 24,662-question evaluation benchmark. Based on this, we develop RetiJev, a family of 0.8B-9B decision models that directly score candidate answers without autoregressive decoding. Across all model sizes, RetiJev outperforms both a 27B LLM and the commercial Jev 1.13 decision model on guideline grading by 10.4-23.0 points. RetiJev-9B further achieves the highest accuracy on clinical decisions, improving over both baselines by 5.4-6.5 points while matching them on diagnosis. It answers a case in 158 ms on one GPU, 3.5x faster than LLM generation and 470x faster than reasoning. Models are released at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30947
PDF
-
Tracing Hallucinations in LLM Reasoning for UML Class Diagram Generation: An Empirical Taxonomy and Mitigation Study
2026-10-22 08:59:27Mojdeh Rahmanian, Foutse Khomh, Ashkan Sami, Yanchao YuAbstract
Large Language Models (LLMs) can generate UML class diagrams from natural language requirements, but hallucinations may compromise the correctness and completeness of the resulting diagrams. Existing evaluations mostly assess the final diagram, without examining the intermediate traces of reasoning-capable LLMs, which can reveal where these hallucinations arise. We develop an empirical taxonomy of eleven hallucination sub-types organised into three dimensions: Prompt Faithfulness (deviations between requirements and reasoning), Code Faithfulness (discrepancies between reasoning and generated PlantUML code), and Reasoning Quality (problems within the reasoning). The taxonomy was derived from 60 manually annotated DeepSeek-R1 traces, applied to 120 further DeepSeek-R1 traces through a dual-LLM annotation pipeline, and examined on 20 DeepSeek-V4-Pro traces. Code Faithfulness was the most frequent dimension for DeepSeek-R1 in both datasets, whereas Prompt Faithfulness dominated for DeepSeek-V4-Pro, suggesting model-dependent failure patterns. In the manually annotated traces, reasoning-trace length showed no significant monotonic association with hallucination frequency, whereas requirement complexity was negatively associated with it, particularly for Prompt Faithfulness errors. In a mitigation experiment on 13 Golden UML ModelSet cases with three Qwen3-based models, adding Prompt Faithfulness and Reasoning Quality constraints to the generation prompt raised the pooled mean F1 score by 4.0 percentage points, whereas a post-hoc Code Faithfulness review provided no consistent further benefit. Relationship modelling remained the lowest-scoring element throughout. These results suggest that localising hallucinations within the generation process can inform targeted, prompt-based mitigation. Whether these findings generalise beyond the DeepSeek and Qwen3 model families remains to be tested.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30946
PDF
-
Verbalizing Activation Differences with Natural Language Autoencoders
2026-10-22 08:54:58Andrea Blasi N\'u\~nez, Gianluca Barmina, Federico Torrielli, Thiusius Rajeeth Savarimuthu, Peter Schneider-Kamp, Lukas Galke PoechAbstract
Activation verbalizers translate the internal representations of large language models into human-readable text. The recently proposed natural language autoencoders (NLAs) learn this by reconstructing activations through a natural language bottleneck. However, it remains unclear how well these methods generalize beyond the setting they were trained for. Here, we study this question for NLAs along four closely related questions: (i) Do NLA verbalizations exhibit compositional structure with respect to their input? (ii) Do NLAs generalize to finetuned versions of the model they were trained for? (iii) Can NLAs recover finetuning objectives in model diffing? (iv) Can NLAs be specialized for model diffing? Across four models from three families, our experiments provide evidence supporting a largely positive answer to all four questions. Specifically, verbalizations of composed analogy activations mostly name the target while rarely mentioning the subtracted word, and an NLA can be reused on lightweight finetunes of its original model without retraining, with reconstruction quality degrading noticeably only for the largest model. Verbalizing activation differences between Taboo model organisms and their original models recovers the concealed secret word for most organisms, and reinforcement learning on activation differences further improves this recovery on held-out organisms. We also find that NLAs can recover the finetuning objective without the finetuned model generating any token, since the most informative token positions precede the user prompt. NLAs do not yet match state-of-the-art model diffing methods on the emergent misalignment detection task, but outperform a blackbox baseline. Our findings suggest that NLAs can meaningfully verbalize differences in activation space, rendering NLAs a promising direction for future work.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30934
PDF
-
ReflectPO: Reflection-Guided Policy Optimization for Long-Horizon LLM Agents
2026-10-22 08:54:24Rumei Li, Xunliang CaiAbstract
With external tools, large language models (LLMs) can tackle complex tasks through sustained environmental interaction. LLM-based agents are therefore increasingly used to tackle complex long-horizon tasks in real-world environments. However, these tasks involve sequential, interdependent decisions and changes in environmental states. Early execution errors affect subsequent interaction histories and environmental states, and their effects accumulate along the decision chain. To address this challenge, we propose Reflection-Guided Policy Optimization (ReflectPO), a reinforcement learning framework for autonomous error recovery in long-horizon tasks. ReflectPO comprises two core components. The first is Reflection-Guided Sampling. We use branch differences and repeated visits in interaction trajectories to construct hints for error recovery. The second is Tree-Structured Credit Assignment. We organize sampled trajectories into trees and assign action-level credit using parent--child reward differences and recovery gains. Experiments over two long-horizon tool-use benchmarks show that ReflectPO outperforms multiple existing baselines in most domains. It also achieves the best overall performance on average. Further analyses show that ReflectPO effectively improves agents' ability to recognize their own deviations, recover from erroneous states, and continue making progress toward task completion.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30933
PDF
-
Beyond Attack Success Rate: Cost-Aware Metric for Evaluating Iterative LLM Jailbreaks
Abstract
Attack success rate (ASR) evaluates jailbreaks against large language models without accounting for repeated calls to the victim model. Yet each additional call gives the deployed system another opportunity to detect the attack and intervene. We introduce $\mathrm{ASR}_{\mathrm{STEP}}$, which penalizes additional victim-model calls while retaining binary success judgments. Our evaluation considers GCG, PAIR with a step budget, and a single-step PAP baseline on JailBreakBench across three model families. GCG's scores decrease by 7.0-8.7 percentage points after discounting, while PAP's remain unchanged. While for PAIR attack, we see a wider discounting range between 3.4-12 percentage points, showing that the "step" quantity differs per iterative strategy. Reporting both metrics makes the discount due to repeated victim-model interaction explicit. Code is available here: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.26589
PDF
