Computation and Language · Cross-listed from cs.LG
Understanding and Exploiting Anisotropy in Post-Training
Samyak Jha, Harshvardhan Saini, Yizhen Liao et al. (+2 more)
LLM post-training combines supervised fine-tuning (SFT), a mode-covering forward-KL objective, with reinforcement learning (RL), a mode-seeking reverse-KL objective. Frequency-weighted likelihood training leaves a well-known signature: anisotropy, in which a few residual channels carry disproportionately large activations. Anisotropy is widely documented and usually treated as a defect, yet its function and its interaction with post-training remain unclear. We first analyze it. A label-free outlier rule isolates about 5% of residual channels that are essential for
language modeling: removing them raises perplexity from 10 to over $10^6$, versus 35 for count-matched random channels. Yet they barely distinguish correct from incorrect
reasoning. SFT reshapes them, whereas RL leaves them largely intact and adapts the complementary channels. These channels therefore form the model's coherence substrate, and
reasoning adaptation happens elsewhere. We then exploit this. SphereGate learns one bounded gain per residual channel on a frozen backbone. Its activation-weighted gradients provably limit movement of high-energy coherence channels and leave the remaining channels free. With 0.1M trainable parameters, SphereGate outperforms parameter-efficient baselines by 2.0–7.3 points on MATH-500 across Qwen2.5 (0.5B–7B) and Llama-3-8B, is comparable or exceeds full-model GRPO. Anisotropy is not a defect but a division of labor that post-training can exploit.
Read the full abstract ▾
All authors Samyak Jha, Harshvardhan Saini, Yizhen Liao, Yiming Tang, Dianbo Liu
Computation and Language · Cross-listed from cs.AI
CLAIRE: A Schema-Grounded Hybrid Workflow for Healthcare Administrative Form Completion
Garapati Keerthana and Manik Gupta
Healthcare administrative staff transfer structured information from electronic health records, referrals, claims systems, provider rosters, and work queues into dynamic forms. We developed and evaluated CLAIRE (Clinical
Language and Agentic Intelligence for
Reasoning and Entry), a hybrid workflow that separates field-state discovery, source-to-field mapping, deterministic validation, bounded correction, escalation, and audit tracing. We tested five synthetic healthcare administrative schemas, 1,000 source records, four interface variants, two data-quality suites, and six comparators, yielding 24,000
benchmark episodes. A separate strict-output audit evaluated direct mappings from Qwen2.5-1.5B and Qwen2.5-7B, and a trace-derived operational simulation covered 6,000 episodes. Under the evaluated synthetic
benchmark conditions, full CLAIRE achieved 1.000 episode success, field accuracy, required-field completion, and dependency completion in both suites; removing validation reduced stress-suite success to 0.500. In the simulation, 100.0% of clean and validation-stress episodes reached a staff-reviewable draft, compared with 68.6% of escalation challenge episodes, unsupported cases were blocked. Scenario-based savings were 149.7-165.5 seconds per case, not observed staff times. The findings support schema-grounded, validation-first healthcare administrative automation in which
language-model components assist mapping but do not authorize unsupported or consequential actions.
Read the full abstract ▾
Computation and Language
On the Behavioral Traits of LLM Agents
Haokai Zhao, Jie Gao, Yunze Xiao et al. (+5 more)
Users increasingly describe different AI
agents as distinct colleagues to work with. AI personality research aims to quantify such impressions by attributing human-like "traits" to
agents. However, existing measures fall short: models' self-reports (S-data) diverge from their actual behavior, while informant ratings from LLM judges (I-data) are costly to scale and cover few everyday scenarios. In this paper, we propose A-B-D to infer traits bottom-up from behavioral data (B-data), namely how
agents act on their environment and communicate with users, as recorded in existing trajectories. From 345,667 real-world trajectories spanning 80 models, 12 tasks, and 50 harnesses, we extract 318 candidate features that capture both the actions an
agent takes at each step (functional) and the
language accompanying them (linguistic). We retain only features that show instance-level stability, cross-task consistency, and model discriminability. Factor analysis of the remaining 79 features uncovers six stable, model-attributable factors, two functional and four linguistic. For example, Kimi-K3 exhibits the most planfulness, whereas GPT-5.5 and GPT-5.6 are the least energetic. Moreover, we quantify the "knowledge-action gap" in the wild: these factors correlate only weakly with self-reported Big Five scores, even for conceptually matched pairs such as extroversion and energetic (r = 0.07, p = 0.58). Our work offers a new lens for understanding AI personality, with implications for users, developers, and researchers from both computer science and social science.
Read the full abstract ▾
All authors Haokai Zhao, Jie Gao, Yunze Xiao, Xintao Wang, Weihao Xuan, Aditya Joshi, Mark Dredze, Jen-tse Huang
Computation and Language
C-HAT-Bench: Benchmarking Chinese AI-Text Detection Beyond Fully Generated Text
Qing Yang, Zixiang Luo, Zhenyu Mao et al. (+6 more)
Large
Language Models (LLMs) increasingly participate in writing by modifying or extending human drafts, causing machine involvement to vary in both form and extent. Yet most Machine-Generated Text (MGT) detectors are evaluated only on fully human-written versus fully AI-generated text. Because human–AI collaboration can weaken or redistribute cues associated with machine generation, strong performance under this binary setting may overstate detector reliability. This mismatch remains underexplored in Chinese: detection cues are shaped by tokenization and
language-specific text distributions, yet controlled resources spanning production settings, domains, and generators remain limited. To fill this gap, we present a Chinese Human-AI Collaborative Text Detection
Benchmark (C-HAT-Bench), a unified
benchmark that links $5,000$ human-written source texts from five domains to more than $240,000$ variants produced using six generative models under Prefix-Conditioned Continuation as a reference setting and three collaborative production modes. We evaluate $21$ detectors through four protocols spanning zero-shot and pretrained supervised document-level detection, boundary localization, and cross-condition generalization. Relative to Prefix-Conditioned Continuation, mean AUROC across document-level detectors is $12.0\%$ lower on the collaborative production modes, with the largest detector-specific relative decrease reaching $44.4\%$. Transfer across collaborative production modes is also asymmetric, indicating that performance in a given production setting is not a reliable predictor of performance in other production settings.
Read the full abstract ▾
All authors Qing Yang, Zixiang Luo, Zhenyu Mao, Zezheng Wu, Xinghe Cheng, Haibo Chen, Qinggang Zhang, Jiapu Wang, Jingwei Zhang
Computation and Language
How Far Do Persona Effects Generalize in Language Models?
Yufan Zhou, Yuxuan Liu, Enze Ma et al. (+6 more)
Persona
prompts ask
language models to answer as particular kinds of people. We test whether relationships learned from these effects predict responses to new questions and remain useful across models and
prompts. Across 57 attributes, three behavioral domains, and seven pairs of open 7 to 9B checkpoints, persona effects can be predictable without being portable. Separate attribute and task gains improve prediction beyond shared scaling significantly in OLMo-3 and Qwen2.5, with the most robust evidence in OLMo-3. In that model, target refitting significantly outperforms gains borrowed from each of the other six pairs. Across model transfers, borrowed gains with one amplitude underperform shared scaling in most directions; allowing two target parameters removes the significant losses but yields no significant benefit over target shared scaling. After rewording, refitting significantly outperforms reuse with one amplitude in all six tested pairs, while changes of examples or country context often preserve reuse value. In the tested
prompt transfers, regularized updates outperform both reuse strategies in median at 64 target questions per attribute. A separate survey comparison finds that selecting the more responsive checkpoint can worsen human fit; responsiveness is confounded with training status, and temperature calibration largely removes this cost but not errors in group ordering. Within the tested gain representation, apparent transfer can come from target calibration; source relationships must add predictive value beyond calibration and regularization. Code and data are available at https://github.com/thzva/persona-gain
Read the full abstract ▾
All authors Yufan Zhou, Yuxuan Liu, Enze Ma, Lyumanshan Ye, Zhongqi Yue, Robin De Croon, Yucheng Jin, Katrien Verbert, Zhao Wang
Computation and Language · Cross-listed from cs.AI
Adaptive Consistency Graph for Long-Horizon Agents
Jiecong Wang, Hao Peng and Zhanyi Wang
Large
language model
agents can often make reasonable local decisions on short tasks, yet their performance degrades when success requires long sequences of dependent actions and tool calls. During execution, task requirements, historical evidence, and the current execution state may gradually become disconnected, so later decisions can drift from the original objective. We study this problem by introducing the Adaptive Consistency Graph (ACG) for long-horizon execution. ACG incrementally organizes execution evidence and its provenance in a persistent graph, then constructs a temporary requirement-centered view for each decision under a bounded context budget. Rather than replacing the base
agent's planner or tool executor, ACG provides a structured and traceable context view for each decision. In the matched evaluation, ACG improves GPT-5.6-luna's average success from 44.5% with ReAct to 50.2%, with the largest gain on BrowseComp-Plus (73.5% versus 62.4%). We further analyze trajectory structure and inference cost to characterize this improvement. Our code is available at https://github.com/yunsaijc/Adaptive-Consistency-Graph.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Retrospective Distillation Attribution via Normalized Response Similarity
Minwoo Jang, Jaechang Kim, Minhyeon Oh et al. (+2 more)
Model
distillation transfers capabilities through supervised fine-tuning (SFT) on teacher responses, often collected from commercial APIs, raising questions of model provenance. Existing
distillation attribution methods have been largely evaluated on students immediately after the SFT step. However, a distilled model may undergo further SFT, preference optimization, or reinforcement learning before release, while an auditor may lack access to the pre-
distillation checkpoint required by reference-based attribution. To close this gap, we propose SCOUT, an output-only method that aggregates recurring *syntactic patterns* into candidate profiles, filters low-contrast patterns, and calibrates student–candidate distances against inter-candidate distances. SCOUT supports attribution and abstention using only current texts, without model weights,
token likelihoods, or historical checkpoints. Auditing publicly released descendants of distilled models spanning diverse post-training objectives, SCOUT consistently identifies the
distillation source. Furthermore, tracing teacher-associated *syntactic signatures* along training trajectories reveals that they emerge during
distillation and persist through subsequent preference optimization and reinforcement learning.
Read the full abstract ▾
All authors Minwoo Jang, Jaechang Kim, Minhyeon Oh, Jeongyeon Hwang, Jungseul Ok
Computation and Language · Cross-listed from cs.CV
Gradient-Guided Decoupled Adaptation for Geospatial Vision-Language Models
Dongdong Wang, Deepak Balakrishnan, Ravi Srinivasan et al. (+1 more)
Existing geospatial
vision-
language models (Geo-VLMs) typically optimize diverse geospatial tasks through a unified multi-task adaptation paradigm without explicitly accounting for the heterogeneous optimization characteristics. Our empirical observations reveal heterogeneous gradient characteristics across tasks, including
vision-
language differences, intra-branch gradient relationships, and task interference, which hinder effective multi-task optimization. Motivated by these observations, we propose Gradient-Guided Decoupled Adaptation (G2DA), a gradient-aware optimization framework for multi-task Geo-VLM learning. G2DA first partitions tasks into
vision- and
language-centric groups through gradient-guided cross-modal decoupling. It then constructs modality-specific curricula based on task gradient similarity and employs bidirectional rehearsal to mitigate the recency effects introduced by sequential optimization. We evaluate G2DA on three Geo-VLM
benchmarks using six InternVL3 and Qwen3.5-VL variants, along with GeoChat and GeoLLaVA. Across all 24
benchmark-model combinations, G2DA consistently outperforms representative baselines, improving over the strongest competitor by 3.08, 4.30, and 2.81 percentage points on UrBench-MCQ, XLRS-Bench-Lite, and VRS-Bench-VQA, respectively. These results demonstrate the effectiveness of gradient-guided task organization for Geo-VLM adaptation.
Read the full abstract ▾
All authors Dongdong Wang, Deepak Balakrishnan, Ravi Srinivasan, Shenhao Wang
Computation and Language
LLM Alignment–Utility Asymmetry under Semantic-Preserving Transformations
Mohan Li, Chengyu Yu, Francesco Sovrano et al. (+2 more)
Large
Language Model (LLM)
alignment is intended to ensure that models remain helpful and safe, but its stability under input distributional shift is not yet fully understood. Prior work shows that aligned models can fail under jailbreak
prompts, alternative encodings, and cross-lingual transfer, yet these failures are usually studied as attacks rather than controlled probes of
alignment generalization. Moreover, existing evidence is largely grounded in natural
language variation already represented during pretraining, leaving unresolved whether
alignment generalizes with semantic content or remains tied to superficial surface patterns. In this paper, we study this question using synthetic semantic-preserving transformations that are rule-based and invertible, preserving task-relevant meaning while shifting inputs beyond standard linguistic variation. Across four open-weight and four commercial models, under both fine-tuning and in-context learning, we use these transformations as a probe of
alignment generalization and identify an empirical pattern we term
Alignment–utility asymmetry: once models can operate effectively on transformed inputs, task utility is often substantially retained while
alignment failure increases more sharply. For example, adapted GPT-4.1 mini shows only limited utility degradation under transformation while its harmful rate rises from 13.3 to 74.3; Gemini 3 Flash similarly retains near-original utility while its harmful rate increases from 2.3 to 43.0. Taken together, these results suggest that semantic-preserving distribution shifts can expose a recurring gap in how utility and
alignment generalize in current LLMs.
Read the full abstract ▾
All authors Mohan Li, Chengyu Yu, Francesco Sovrano, Marc Langheinrich, Martin Gjoreski
Computation and Language · Cross-listed from cs.AI
MassAlloc Attention: Let Attention Allocate Its Own Compute
Jingze Shi, Zhangyang Peng, Xianduo Li et al. (+6 more)
FullAttn often assigns negligible normalized mass to much of the causal score space, yet dense kernels execute the complete post-score path after forming each QK tile. We introduce MALA, a fused
attention primitive that preserves score access to every legal causal interaction and uses normalized contribution to allocate post-score computation. Forward uses its evolving online-softmax normalizer, while backward reuses the finalized normalizer to derive nested retained support using only standard
attention state. A common tolerance governs training and inference, allowing for adaptive retention of the work. MALA reduces low-contribution post-score computation. A matched-work study at 8K isolates the benefit of distribution-adaptive allocation: under exactly matched total post-score work, MALA approaches a per-instance reference-mass oracle, with mean omitted mass of 0.0188% versus 0.0182%. Across context lengths from 1K to 32K
tokens, the same tolerance maintains low output and gradient errors relative to the reference. Across a broader controlled associative-recall comparison, MALA closely tracks FullAttn as context grows, reaching 89.67% accuracy at 8K compared with 89.97% for FullAttn. In an
attention-operator
benchmark at 128K
tokens with tensor parallelism, MALA reduces forward and backward latency during training by 2.2x and 3.0x and decoding latency during inference by 1.6x relative to FullAttn. Across scaling-law training from 0.6B to 14B parameters, MALA closely tracks FullAttn in perplexity while reducing total training FLOPs. The resulting 14B models and 32B models from separate continued training achieve comparable knowledge,
reasoning, and long-context
retrieval scores to FullAttn. These results indicate that allocating post-score computation according to normalized
attention contributions can retain the evaluated capabilities of FullAttn while reducing
attention computation.
Read the full abstract ▾
All authors Jingze Shi, Zhangyang Peng, Xianduo Li, Yanlin Qi, Xiaotian Lin, Haoxian Chen, Liangdong Wang, Guang Liu, Yuyu Luo
Computation and Language · Cross-listed from cs.AI
Despite Instructions: Frontier Agents Improvise Covert Channels at Test Time
Jacob Dineen, Silei Ren, Muhao Chen et al. (+2 more)
In security-sensitive applications,
language-model
agents are often required to coordinate without disclosing confidential information. Yet repeated interactions may also let ordinary messages acquire shared private meaning. We study a repeated game with pairs of models in which the sender model observes one of four secret states and selects one of four summaries of the same public report, while the receiver model tries to infer the secret state. We find that model pairs can learn to communicate the secret using only one bit of feedback indicating whether the receiver inferred it correctly. This learning occurs during inference with fixed parameters and no supplied codebook or encoding examples. The effect also persists when
agents generate their own free-form updates in a simulated incident-response task. Across ten independent games, pairs of GPT-5.6 Sol
agents reach 98.8% final accuracy, compared with 25% chance, despite explicit instructions prohibiting disclosure and a monitor that screens each message without access to the
agents' interaction histories. The same interactions that help
agents cooperate can therefore allow confidential information to pass through messages intended for legitimate coordination.
Read the full abstract ▾
All authors Jacob Dineen, Silei Ren, Muhao Chen, Dan Roth, Ben Zhou
Computation and Language · Cross-listed from cs.AI
IGSD: Environment-Verified Hindsight Self-Distillation for Search Agents
Angqing Jiang, Gaoming Zhang, Chaoqun Zhang et al. (+5 more)
On-policy self-
distillation densifies
agent training without external teachers: a policy conditioned on privileged hindsight provides step-level guidance for its own unprivileged rollouts. For search
agents, however, hindsight can make the teacher prefer a query that does not improve
retrieval from the student's state. Existing methods either distill this preference directly or filter it with model-internal scores, but neither strategy verifies the query's executed
retrieval consequence. We propose Information-Gain-Gated Self-
Distillation (IGSD), which verifies on-policy
token proposals with environment feedback before distilling them. Treating each query
token as a micro-action, IGSD completes the teacher's
token proposal and the student's sampled
token into matched queries and executes both from the same failed state with the same retriever. Shared counterfactual controls account for query-conditioned shifts in answer likelihood, so their difference, the executed paired information gain, provides a relative utility contrast for the retrieved documents. IGSD uses this contrast as a positive-only soft weight for candidate-pair
distillation, while leaving the GRPO objective unchanged and confining verification to training. Across seven single-hop and multi-hop QA
benchmarks, IGSD reaches macro-average exact-match accuracies of 42.8% and 47.0% with 3B and 7B policies, respectively, without inference-time verification. These results support environment-verified hindsight as an effective approach to reliable action-level supervision for search
agents.
Read the full abstract ▾
All authors Angqing Jiang, Gaoming Zhang, Chaoqun Zhang, Jianchun Song, Liyuan Kong, Kena Qi, Wei Lin, Defu Lian
Computation and Language
Focusing Condition: Inference-Time Self-Contrastive Steering Elicits Better Conditional Text Embeddings in LLMs
Zifeng Cheng, Lingyun Qian, Zhiwei Jiang et al. (+5 more)
Extracting conditional text
embeddings from large
language models (LLMs) is a promising paradigm, as it requires neither additional data nor fine-tuning. Existing methods incorporate conditions into
prompts to guide LLMs to focus on specific aspects and elicit conditional text
embeddings. However, relying solely on
prompts often fails to produce high-quality conditional text
embeddings, as they remain entangled with general text
embeddings, ultimately degrading their quality. To this end, we propose an inference-time, plug-and-play Self-Contrastive Steering (SCS) method that constructs unconditional general text
embeddings and uses them to refine conditional text
embeddings, making them more focused on the target condition. Specifically, we modify the
attention mask and positional encodings to mask the condition, thereby obtaining unconditional text
embeddings and intervening in the multi-head self-
attention computation process. Notably, our method is highly efficient, requiring only a single additional multi-head self-
attention computation at inference time. Extensive experiments on clustering, Semantic Textual Similarity, and triplet
alignment datasets demonstrate that our method can seamlessly improve the performance of existing
prompt-based methods across different LLMs in a training-free and plug-and-play manner. Our code will be released at https://github.com/zifengcheng/SCS
Read the full abstract ▾
All authors Zifeng Cheng, Lingyun Qian, Zhiwei Jiang, Cong Wang, Yafeng Yin, Fei Shen, Ao Zhou, Qing Gu
Computation and Language · Cross-listed from cs.LG
The GUI Is Not the State: Diagnosing State Aliasing in GUI World Models
Dongsheng Liu, Chao Jin, Wenkui Yang et al. (+7 more)
GUI World Models (GUI-WMs) are increasingly used to predict future states for
agent planning and simulation, yet most existing formulations condition only on the current GUI observation and action. We identify state aliasing, where the vis- ible interface omits transition-relevant environment state, so identical observable conditions can correspond to different valid futures. To diagnose this failure mode, we introduce StateAliasBench, a diagnostic
benchmark that explicitly isolates such ambiguities via strict pairing. We further propose lightweight predictive- state recovery that infers structured state from history and augments otherwise frozen GUI-WMs through a deterministic state interface. Family-specific special- ists provide state recovery across heterogeneous state types, and multi-teacher dis- tillation consolidates them into a single unified
estimator. Experiments show that existing GUI-WMs exhibit systematic failures under observation-only condition- ing, while predictive-state augmentation substantially restores state-sensitive pre- diction across evaluated WMs, preserves generative fidelity, and improves down- stream performance of GUI
agents on AndroidWorld. These results suggest that reliable GUI world modeling should account not only for what is visible, but also for the hidden transition state that determines what happens next.
Read the full abstract ▾
All authors Dongsheng Liu, Chao Jin, Wenkui Yang, Hejin Wang, Junwei Yang, Zeren Zhang, Ziwei Chen, Huaibo Huang, Jie Cao, Ran He
Computation and Language · Cross-listed from cs.LG
Distance-KV: Exploiting Relative Distance for Efficient Long-Context Inference
Xianpeng Shang, Canbin Huang, Jiang Li et al. (+4 more)
The
memory usage and decoding latency of LLM inference grow rapidly with context length. To reduce these costs, key-value (KV) cache compression methods selectively retain cached states based on
token importance or differences in
attention patterns across heads. However, we discover that
retrieval capability varies substantially with relative distance, even within the same
attention head. To exploit this structure, we introduce Distance-KV, which learns a static KV retention pattern over the joint space of layers,
attention heads, and relative distances. The pattern is learned offline with the
language model frozen and reused across inputs to prune and compact the KV cache without online importance scoring. Across three backbone models and four long-context
benchmarks, Distance-KV consistently achieves the best overall performance among competing KV cache compression methods, exceeding the strongest compression baseline by up to 9.3 points on RULER at 128K. On Llama-3.1-8B-Instruct at 128K, Distance-KV reduces KV cache
memory by 65.4% and achieves a $1.66\times$ decoding speedup relative to Dense. Together, these results identify relative distance as an important structural dimension for understanding how LLMs retrieve information over long contexts and for designing more efficient inference methods.
Read the full abstract ▾
All authors Xianpeng Shang, Canbin Huang, Jiang Li, Tian Lan, Qianyi Cai, Xiaojun Quan, Xiangdong Su
Computation and Language
From Knowing to Abstaining: Bridging the Representation-Action Gap in Vision-Language Models
Jialuo He and Huangxun Chen
The ability of
vision-
language models (VLMs) to abstain from unanswerable questions is as important as their ability to answer answerable ones accurately. Recently, several
benchmarks have emerged to evaluate and improve VLM abstention, but they have substantial limitations. First, samples often contain shortcut cues in images or questions that reveal answerability, while an explicit "unanswerable" option further prevents accurate assessment of spontaneous abstention. Second, as training data, they generally provide only binary labels without fine-grained explanations for deeper supervision. To address these limitations, we introduce Visual Answerability Diagnosis with Rationales (VAD-R), a
benchmark constructed through a two-stage pipeline of shortcut filtering and quality verification to prevent answerability leakage. Each example is annotated with step-by-step rationales and causal evidence-gap labels. Evaluation of state-of-the-art open- and closed-source VLMs on VAD-R reveals limited spontaneous abstention, with average recall rates of only 11.4% and 16.3%, respectively. Probing analyses show that hidden-state representations in certain layers can effectively distinguish answerability, yet this distinction fails to manifest in final responses. Motivated by this observation, we introduce Rep2Act, a representation-to-action
alignment method that translates latent answerability awareness into explicit abstention decisions. Rep2Act improves action accuracy on VAD-R from 56.67% to 86.33% for Qwen2.5-VL-3B and from 59.33% to 88.67% for Qwen2.5-VL-7B. On the out-of-distribution TUBench, Rep2Act achieves an average F1 score of 53.3% with only a 3B model, surpassing the closed-source GPT-4 Turbo and GPT-4o by 16.2% and 1.1%, respectively.
Read the full abstract ▾
Computation and Language
ExpVoyager: Direct Experience Navigation for Dynamic Agent Skill Synthesis
Kwangwook Seo and Dongha Lee
Learning from experience in LLM
agents has become a key paradigm for developing self-evolving
agents that continuously learn and expand their capabilities. Within this paradigm, synthesizing the
agent skill has emerged as a promising solution for transforming accumulated experience into reusable procedural knowledge, serving as an important layer for the harness system that supplies
agents at runtime. Despite its potential, existing approaches largely abstract past experience into fixed procedural knowledge before downstream demands are known, which risks discarding knowledge that later becomes critical while retaining instance-specific details irrelevant to future tasks. In this paper, we reframe
agent skill synthesis as a dynamic navigation problem over past experience, where
agents actively explore accumulated trajectories on demand for the current task with targeted and fine-grained access to experience knowledge. To this end, we propose ExpVoyager, a novel framework in which a skill curator navigates raw experience across different views and resolutions, continually identifying reusable procedural knowledge from what it observes while tracking remaining knowledge needs that guide where to navigate next. Extensive experiments demonstrate both the effectiveness and versatility of ExpVoyager, showing consistent improvements in downstream task performance, continual gains as the experience space scales, and practical compatibility with existing skills under efficient experience access.
Read the full abstract ▾
Computation and Language
MixDetect: Word-Level Localization and Quantification of AI Editing
Hongrui Bao, Yubing Ren, Zhendong Pan et al. (+3 more)
Large
language models are increasingly used to edit human-written text rather than generate entire texts from scratch. Conventional AI-text detectors mainly distinguish human-written from fully AI-generated text, while recent methods for AI-edited text typically provide only a text-level label or editing-degree score. We introduce MixDetect, a word-level framework for localizing and quantifying AI editing. MixDetect separately predicts whether each word has been edited and, conditional on editing, how substantial the edit is, allowing editing scope and editing intensity to be estimated separately. During training, source–edited pairs are aligned to construct word-level supervision, while inference requires only the input text. Experiments show that MixDetect accurately localizes AI-edited words, reflects differences in editing intensity, and reveals different scope–intensity patterns across editing degrees and operations. The overall AI editing magnitude increases under additional AI editing, decreases when AI-generated text is edited by humans, and remains nearly unchanged under ordinary human-to-human editing. The aggregated text-level predictions also perform well on binary and ternary AI-text classification and remain effective under domain and generator shifts. These results show that AI editing can be analyzed beyond a single authorship label or editing-degree score by identifying both where AI editing occurs and how substantial the edits are.
Read the full abstract ▾
All authors Hongrui Bao, Yubing Ren, Zhendong Pan, Fang Fang, Shi Wang, Yanan Cao
Computation and Language
Does CoT-Pass@k Really Check the CoT? A Multilingual Mathematical Audit
Tar{\i}k Tuna Taşalt{\i}, Burcu Hüdaverdi and David Semedo
Pass@k measures whether a model reaches a correct answer under repeated sampling, but never how: a lucky guess counts the same as sound
reasoning. CoT-Pass@k was proposed to close that gap, adding an LLM-as-judge that must assess a solution's
reasoning chain before it counts. Its value rests entirely on one assumption: that the judge catches flawed
reasoning. That assumption has never been tested inside the metric that depends on it, and never outside English, though the metric's claims concern models used in many
languages. We report the first audit of that verification step, run under the metric's own protocol on a multilingual suite of five mathematical
benchmarks in English, Turkish and Portuguese, two of them natively written. We corrupt correct solutions with deterministic edits that damage the chain and the final answer separately. We observe that all three judges accept corrupted chains almost as often as clean ones. V4-Flash and Qwen3.6 reject a solution sharply only when its final answer is wrong and accept a wrong answer more readily when the chain agrees with it; the metric's own judge accepts most wrong answers as well. Our study shows that chain-answer agreement dominates the two larger judges' verdicts and that all three fail to reliably detect the tested
reasoning errors. Consequently the difference Pass@k - CoT-Pass@k averages 19.7 points on an earlier solver generation but only 4.1 on the current one. What little remains depends on the
token budgets on both sides and on the generation mode; raising the generation budget moves Pass@64 by more than fifty points while the difference stays at zero. We close with two checks any judged
reasoning metric should pass before its numbers are read as evidence about
reasoning.
Read the full abstract ▾
Computation and Language
The Alignment Paradox: How Post-Training Amplifies Confident Hallucinations in Language Models
Qingjia Huang, Yakai Li, Jianguo Wu et al. (+5 more)
Large
language models (LLMs) can produce factually incorrect answers with high confidence, undermining their reliability and limiting the effectiveness of uncertainty-based error detection. While prior research attributes confident
hallucinations to factors such as missing knowledge in training data,
reasoning errors, or
stochastic decoding, we uncover that post-training
alignment itself is a primary driver of these errors, a phenomenon we call the
Alignment Paradox. Across five model families evaluated on factual
benchmarks, unaligned base models produce few high-confidence errors on long-tail factual queries, whereas instruction-tuned models multiply high-confidence errors ($p \ge 0.95$) by more than an order of magnitude (10$\times$ to 35$\times$). Layer-wise probing with the Logit Lens reveals that this overconfidence emerges in late layers, where wrong-answer margins expand past 4.0 points after remaining near zero across early and intermediate layers. These findings motivate limiting margin growth during post-training. We implement this principle through an
entropy-dependent margin bound in direct preference optimization (DPO). In multi-epoch experiments with Mistral-7B, the bounded objective reduces high-confidence errors by up to 35.3% relative to standard DPO while maintaining performance on evaluated general
reasoning benchmarks. These results show that bounded margins mitigate confident
hallucinations during post-training.
Read the full abstract ▾
All authors Qingjia Huang, Yakai Li, Jianguo Wu, Qihang Zhou, Aimin Yu, Xiaoqi Jia, Luping Ma, Weijuan Zhang
Computation and Language
KV-Lingo: Learning KV-Cache Translators with Distillation
Valérie Castin, Keitaro Sakamoto, Anastasiia Filippova et al. (+3 more)
Large
language models represent context with a key-value (KV) cache. Caches are model-specific: for the same text, models with different architectures or weights produce incompatible representations. This makes it costly to switch models over a shared context: although the context has already been processed by one model, the incoming model must process it again to build its own cache. We introduce KV-Lingo, a method for translating the KV cache of a source model into one that can be read by a target model. KV-Lingo consists of a collection of linear maps, typically one per layer of the target model, that are applied independently on all
tokens' key and value representations. We train these maps using
distillation, minimising the divergence between the target model's predictions from its native cache and those from the translated cache. We consider several model pairs spanning multiple sizes and architectures, training one translator per pair on a generic text corpus. The resulting translators preserve strong downstream performance in both small-to-large and large-to-small transfers. Since a switch then costs a linear map and a single decoding step instead of a prefill, replacing re-prefill with cache translation reduces the time to first
token after a model switch by 9.6x already on a 64-
token prompt for Qwen models on an Apple M3 Ultra, and by up to 29x at 32k context length on an H100. These gains make KV-Lingo particularly useful for dynamic model routing: a context can be processed by one model and handed off to another only when needed, without re-prefilling the shared prefix. We finally show that KV-Lingo can be used for seamless model switching, staying close to re-prefill across repeated switches in our multi-turn evaluations.
Read the full abstract ▾
All authors Valérie Castin, Keitaro Sakamoto, Anastasiia Filippova, João Monteiro, Marco Cuturi, Pierre Ablin
Computation and Language · Cross-listed from cs.CV
Retrieved but Not Delivered: Multimodal Memory Delivery for Long-Term Agents
Yuhang Jiang, Qingwei Liao, Kaize Yin et al. (+3 more)
Work on
memory for
multimodal agents optimizes what is written, updated and retrieved. Between
retrieval and the answer, however, is a stage that
multimodal memory evaluations do not isolate: what of the retrieved
memory reaches the model, and in what form. We call it delivery, and a controlled decomposition on MemLens locates the remaining room there. With the retrieved evidence set exactly fixed, delivering the original pixels instead of withholding them raises accuracy by 13.87 points on an 8B backbone, whereas making
retrieval perfect on those same messages improves it by 2.31. Delivery is the larger term on all three MemLens backbones and grows with backbone strength;
retrieval grows too, without closing the gap. We propose DeliverMem, an instantiation of delivery as three decisions: keep the original modality, give each item a readable identity, and state when it was seen, with a
retrieval-side adapter for the one property delivery cannot supply. Each is measured against a delivery-matched control that alters only its own variable. DeliverMem leads the strongest published
memory agent on MemLens at all four context lengths, and beats DMV-Bench's own strongest method at every setting on both backbones. On MemLens it does this on a tenth to a seventieth of the input. Each decision helps only where the question lacks what it supplies, and is null elsewhere. A single fixed configuration nonetheless leads both
benchmarks, without training any component or modifying the stored records. Project page: https://avalon-s.github.io/DeliverMem/
Read the full abstract ▾
All authors Yuhang Jiang, Qingwei Liao, Kaize Yin, Xingling Liu, Luca Cuomo, Silvio Bacci
Computation and Language
HERO-MoE: Historical Expert Routing with Scale-Preserving Fusion
Junxiang Qiu, Zhengsu Chen, Xinting Hu et al. (+6 more)
Mixture-of-Experts (MoE) architectures have become a standard way to scale model capacity while keeping computation sparse, yet routing remains a key determinant of MoE quality and training behavior. Prior empirical studies suggest that MoE routing reflects input semantics and upstream computation across depth, but standard routers do not explicitly use the routing distributions produced by preceding layers. We propose HERO-MoE, Historical Expert ROuting with Scale-Preserving Fusion, a routing framework that injects historical routing priors into MoE routers by reusing detached, dense routing distributions collected from preceding MoE layers. The key idea is simple: HERO-MoE preserves the original
token-conditioned routing branch and adds a residual historical routing contribution before the standard softmax and top-$k$ dispatch. To stabilize this historical signal, HERO-MoE introduces a scale-preserving fusion mechanism that matches the magnitude of historical routing
memory to the current hidden representation and accounts for the number of visible historical layers, without introducing an auxiliary routing loss or a fusion-specific tuning parameter. By reusing routing distributions already computed by preceding MoE layers, HERO-MoE improves training-loss reduction with modest end-to-end overhead. The resulting router remains compatible with standard sparse dispatch, including top-$k$ and group-limited routing, and can be inserted into existing MoE backbones with minimal architectural changes. Experiments on an approximately 8B-parameter MoE model with 0.5B active parameters, trained from scratch on 100B
tokens, show that HERO-MoE reduces the final loss from 1.6393 to 1.6184, while peak
memory and FLOPs increase by only 0.44% and 0.64%, respectively.
Read the full abstract ▾
All authors Junxiang Qiu, Zhengsu Chen, Xinting Hu, Shuo Wang, Hengheng Zhang, Shaofeng Zhang, Changcheng Li, Boyu Shi, Qi Tian
Computation and Language
Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL
Jinhao Dong, Liang Zhao, Zihao Yue et al. (+7 more)
Reinforcement learning (RL) for code
agents often uses executable tests to provide binary rewards. With these rewards, Group Relative Policy Optimization (GRPO) assigns identical advantages to test-passing trajectories within each rollout group, overlooking differences in implementation quality and adherence to task requirements. This leaves the policy without a learning signal that favors clean, targeted implementations over those containing unnecessary or out-of-scope changes. We introduce GAGAR, a framework for quality-aware credit redistribution in code
agent RL. Built on dynamic sampling that retains groups containing both passing and failing trajectories, GAGAR places all trajectories from each group in a shared workspace, where an SFT-trained agentic grader jointly inspects them and ranks the test-passing candidates. Based on this ranking, we downweight lower-ranked trajectories and proportionally rescale the advantages of all test-passing trajectories to restore their original sum. This sum-preserving redistribution retains the relative weights established by quality-based downweighting while shifting credit toward higher-quality implementations. We evaluate GAGAR at industrial scale using pre-RL SFT checkpoints of
MiMo-V2.6-Flash (310B total parameters) and
MiMo-V2.6-Pro (1.02T total parameters). Controlled code-only Flash experiments show improved code
agent performance, reduced trajectory-length growth, and more stable training. We further apply GAGAR in large-scale mixed-task RL with both Flash and Pro. Our results support combining test-based verification with groupwise agentic grading to improve the quality and stability of code
agent RL.
Read the full abstract ▾
All authors Jinhao Dong, Liang Zhao, Zihao Yue, Wenhan Ma, Linghao Zhang, Lei Li, Shicheng Li, Yifan Song, Bowen Ye, Fuli Luo
Computation and Language · Cross-listed from cs.AI
CUE-Mem: Benchmarking Long-Term User Memory via Implicit Cues in Multimodal Conversations
Yulin Hu, Yanyan Zhao, Zimo Long et al. (+6 more)
Long-term
memory is essential for
multimodal agents that interact with users across sustained conversations. However, user
memories are not always explicitly stated: they may also be implied by recurring background objects in images, ambient sounds in audio, or other peripheral
multimodal cues. Existing
benchmarks largely focus on text-only
memory or explicit
multimodal evidence, leaving implicit
multimodal cues underexplored. We introduce CUE-Mem, a text-image-audio
benchmark for evaluating long-term user
memory from implicit cues. CUE-Mem contains 2,674 questions across explicit and implicit evidence settings and covers four tasks: Entity Recall, Long Pattern, Personalized Recommendation, and Answer Refusal. Across textualized
memory systems, implicit performance remains far below oracle evidence, locating the main bottleneck in preserving and retrieving subtle cues rather than question answerability. Increasing caption detail recovers more of this evidence, but brings uneven gains and rapidly growing
token costs, motivating native
multimodal access. Yet native access does not uniformly resolve the bottleneck: evidence use depends strongly on the backbone, while
multimodal indexing introduces substantial
retrieval noise. CUE-Mem provides a testbed for
memory systems that selectively retain, retrieve, and use subtle
multimodal evidence.
Read the full abstract ▾
All authors Yulin Hu, Yanyan Zhao, Zimo Long, Xing Fu, Mengtong Ji, Weixiang Zhao, Yutai Hou, Qianchao Wang, Dandan Tu
Computation and Language · Cross-listed from cs.CV
Attribution Gaps in Zero-Training LLM+OVOD Pipelines: A Fine-Grained Analysis of the CAAP–SNAP Discrepancy
Yu-Feng Yen
LAOD and similar zero-training LLM+open-vocabulary-detector (OVOD) pipelines score two things separately: class-agnostic localization accuracy (CAAP) and semantic naming accuracy (SNAP). The two consistently diverge, and nobody has asked why. This paper asks why, on the full 5,000-image COCO-Val split (27,273 detections) rather than the small subset the original work evaluated on. Object visual complexity turns out not to be the driver – small and occluded objects are, if anything, localized better than large ones. Vocabulary novelty is: once the LLM's wording falls outside the detector's native category set, localization accuracy falls from 80.9% to 31.6%. That drop is not spread evenly across unfamiliar phrasing, though. Almost all of it comes from cases where the novel wording actually names a different object than the one COCO annotated (true synonyms still score 89.3%; semantically unrelated "noise" labels score 12.0%). A closer look at a further failure subset tells a similar story: 78-88% of what looks like complete localization failure is really the model correctly finding a real object that COCO's non-exhaustive 80-category scheme simply never labeled, not
hallucination. Swap the detector backbone (YOLO-World for Grounding DINO) or the LLM (Gemma-3 for Qwen2.5-VL) and both the effect and its rough size hold up, so this looks like a general property of the pipeline family rather than a quirk of one model pairing. The upshot is that a large share of the apparent CAAP–SNAP gap traces back to closed-category annotation limits rather than a real grounding failure, which matters for how we detect
hallucination, analyze failure modes, and design evaluation for grounded
multimodal systems meant to work in the open world.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CR
Reading Is Not Leaking: Local, Auditable Measurement and Reduction of Inference Exposure from Public Footprints
Mahmudul Faisal Al Ameen
Anyone with a public footprint leaks facts that were never stated, and
language models make the inference cheap. We present a framework for measuring and reducing this inference exposure that runs on the owner's own CPU with no
language model at analysis time, instantiated on organisations and on individuals. It starts from a measurement result: scoring an inference system against the target's private truth conflates how well the system reads the record with how much the record leaks. On a 128-question instrument over sixteen synthetic firms, almost half of the questions are never answered correctly by any of six readers, four of them
language models, and a majority-class guess accounts for most of every reader's score. We therefore separate reading accuracy from leakage rate and introduce an injection protocol that creates
cells with known support. Our analyser combines rules, statistical solvers and a 106M-parameter encoder trained from scratch that marks verbatim evidence and never generates text; every answer carries a graded certificate whose recorded proof replays. Its certified answers are correct in 93% of resolved cases, against 49-73% for the
language models' quote-backed answers, whose citations are produced alongside the answer rather than deriving it; with plain-prose articles in the record, 70% of its evidence-bearing answers rest on evidence that establishes them, against 18-56% for the models. A constrained defence that rewrites each fact's carrier as a true but coarser statement hides every single-carrier fact from four
language-model adversaries at 40% lower edit cost than deletion. On sixteen synthetic people the guessing term is larger still, and a decoy planner with no
language model halves the correct answers of the
estimator it targets without transferring to a second.
Read the full abstract ▾
Computation and Language
Husein Zolkepli
Whisper is still what runs in production: one permissively licensed checkpoint, 99
languages, no per-
language tuning. But it writes sentences nobody said. On 42 clips of pure room tone, whisper-large-v3 emits words on 61.9% of them and emits something on 100%. The usual response is to distil a student from Whisper pseudo-labels, filtering the
hallucinations out of the corpus first, which deletes the evidence while leaving the behaviour in place: the student is fitted to the subset where the teacher was right, and inherits a failure mode absent from its own training data. The second response, adding non-
speech audio so the model learns to stay quiet, already ships and works, but teaches suppression without discrimination. The checkpoint best on every non-
speech arm here is also the one that recovers the fewest genuinely spoken phrases and deletes 58% of repeated
speech. The teacher has to be fixed, with both halves of the signal. Our
benchmark scores both at once: 11,852 clips over eight arms, 6,267 synthetic positives, 8,296 clips of real audio that made a production model fail, and FLEURS in 58
languages. We collect 40,891
hallucination phrases in 100
languages, choose a text-to-
speech system by measurement, and synthesise those phrases as positives, so the model meets the same text both as something to suppress and as something to transcribe. Across 33 matched pairs of fine-tunes differing only by those positives, adding them lowers word emission on real voice-free audio in 31 pairs and raises phrase recovery in 32; among the 30 pairs that hold English accuracy it is 30 out of 30. The best checkpoint takes
hallucination on silence from 61.9% to 2.4% and words over real voice-free audio from 99.9% to 47.8%, while raising phrase recovery from 69.8% to 82.7%. The
benchmark, lexicon and synthetic corpus are released at https://huggingface.co/
datasets/Scicom-intl/Whisper-
Hallucination
Read the full abstract ▾
Computation and Language
Language as an Independent Information Layer: A Conceptual Model of Communication, Cognition and Decision-Making
Anastasiia Alifanova and Elena Benderskaya
Based on an analysis of the role of
language in thought and communication, this article proposes a new concept for designing corporate knowledge bases. The concept integrates the probabilistic vector space of a corporate vocabulary, reflect-ing industry specifics, subject focus, terminology, and culture, with traditional ontological modeling. This combination enables the efficient extraction of knowledge from accumulated corporate documents while strictly accounting for specific business processes. Consequently, this concept bridges statistical and semantic (cause-and-effect) methodologies. Furthermore, analyzing the projec-tions of probabilistic spaces and causal relationships can help identify bottlenecks in business logic. As a dynamic system,
language functions as a separate, inde-pendent layer within the overall information architecture. Introducing a dynamic component into the probabilistic space of word distribution allows it to be mod-eled as a multidimensional solution space for various problem formulations. In this context, input data defining the problem conditions serve as control parame-ters for dynamic transformations.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Shared Autoregressive Context Can Distort Relationships in Synthetic Data
Thomas S. Robinson
Large
language models can generate several records within one autoregressive completion, making earlier answers available as context for later records. This paper shows that such shared-completion batching can distort relationships among variables in the resulting synthetic data, using controlled tests on synthetic survey respondents. In a matched experiment on 2,000 European Social Survey profiles, generating ten rather than one respondent per request increases mean absolute error in within-country correlations by 48-58% for Qwen3.8-27B and 114-127% for Llama-3.3-70B-Instruct across three seeds, holding profiles, examples, questions and decoding parameters fixed. The distortion primarily reflects exaggerated relationship strength, while retaining substantial agreement with the human ordering of correlations. Controlled interventions establish answer history as a causal channel: re-pairing the same preceding values, with profiles and marginal distributions fixed, changes correlations among subsequently generated responses. Hiding preceding answers reduces correlation error in the tested settings but worsens marginal accuracy. Exploratory corrections across social-attitude, health and economic data likewise show that lower correlation error can coexist with worse marginal distributions and regression estimates. Request construction is therefore part of the data-generating process, and synthetic-data validity must be evaluated against the analyses the generated data are intended to support.
Read the full abstract ▾