Computation and Language · Cross-listed from cs.LG
Skip-Connected Policy Optimization for Implicit Advantage
Fengwei Teng, Jinyi Bai, Xinhao Yao et al. (+3 more)
Group Relative Policy Optimization (GRPO) has proven effective in RLVR by using outcome-based rewards. While fine-grained dense rewards can theoretically improve performance, we reveal that under practical sampling budgets, Monte Carlo estimation yields high-variance and sign-inconsistent advantages for early
reasoning tokens, paradoxically underperforming outcome-only GRPO. We propose Skip-Connected Optimization (SKPO), which decomposes
reasoning into upstream and downstream phases: upstream receives dense rewards from downstream Monte Carlo sampling with single-stream optimization; downstream maintains group-relative optimization, where a skip connection concatenates the upstream segment with the original problem, enabling the model to leverage helpful upstream
reasoning while preserving the freedom to bypass flawed
reasoning through direct problem access. Experiments demonstrate improvements of 3.91% and 6.17% relative gains over the strongest baselines on Qwen2.5-Math-7B and Llama-3.2-3B respectively across mathematical
benchmarks and out-of-domain tasks including general
reasoning and code generation. Further analysis reveals an implicit advantage: SKPO generates trajectories with higher intermediate-step quality even when matched for final correctness.
Read the full abstract ▾
All authors Fengwei Teng, Jinyi Bai, Xinhao Yao, Demi Ruohan Wang, Jiahao Zhao, Zhijiang Guo
Computation and Language · Cross-listed from cs.LG
PRAGMA: Revolut Foundation Model
Maxim Ostroukhov, Ruslan Mikhailov, Vladimir Iashin et al. (+11 more)
Modern financial systems generate vast quantities of transactional and event-level data that encode rich economic signals. This paper presents PRAGMA, a family of foundation models for banking event sequences. Our approach pre-trains a
Transformer-based architecture with masked modelling on a large-scale, heterogeneous banking event corpus using a self-supervised objective tailored to the discrete, variable-length nature of financial records. The resulting model supports a wide range of downstream tasks such as credit scoring, fraud detection, and lifetime value prediction: strong performance can be achieved by training a simple linear model on top of the extracted
embeddings and can be further improved with lightweight fine-tuning. Through extensive evaluation on downstream tasks, we demonstrate that PRAGMA achieves superior performance across multiple domains directly from raw event sequences, providing a general-purpose representation layer for financial applications.
Read the full abstract ▾
All authors Maxim Ostroukhov, Ruslan Mikhailov, Vladimir Iashin, Artem Sokolov, Andrei Akshonov, Vitaly Protasov, Andrey Goncharov, Dmitrii Beloborodov, Vince Mullin, Roman Yokunda Enzmann, Georgios Kolovos, Jason Renders, Pavel Nesterov, Anton Repushko
Computation and Language
EXAONE 4.5 Technical Report
Eunbi Choi, Kibong Choi, Sehyun Chun et al. (+55 more)
This technical report introduces EXAONE 4.5, the first open-weight
vision language model released by LG AI Research. EXAONE 4.5 is architected by integrating a dedicated visual encoder into the existing EXAONE 4.0 framework, enabling native
multimodal pretraining over both visual and textual modalities. The model is trained on large-scale data with careful curation, particularly emphasizing document-centric corpora that align with LG's strategic application domains. This targeted data design enables substantial performance gains in document understanding and related tasks, while also delivering broad improvements across general
language capabilities. EXAONE 4.5 extends context length up to 256K
tokens, facilitating long-context
reasoning and enterprise-scale use cases. Comparative evaluations demonstrate that EXAONE 4.5 achieves competitive performance in general
benchmarks while outperforming state-of-the-art models of similar scale in document understanding and Korean contextual
reasoning. As part of LG's ongoing effort toward practical industrial deployment, EXAONE 4.5 is designed to be continuously extended with additional domains and application scenarios to advance AI for a better life.
Read the full abstract ▾
All authors Eunbi Choi, Kibong Choi, Sehyun Chun, Seokhee Hong, Junwon Hwang, Hyojin Jeon, Ahra Jo, Hyunjik Jo, Yeonsik Jo, Joonkee Kim, Seonghwan Kim, Soyeon Kim, Sunkyoung Kim, Yireun Kim, Yongil Kim, Changhun Lee, Haeju Lee, Jinsik Lee, Kyungmin Lee, Sangha Park, Kwangrok Ryoo, Minju Seo, Sejong Yang, Heuiyeen Yeen, Hwan Chang, Stanley Jungkyu Choi, Yejin Choi, Kyubeen Han, Joonwon Jang, Kijeong Jeon, Geunyeong Jeong, Gerrard Jeongwon Jo, Jiyeon Jung, Daeseong Kim, Dohoon Kim, Dohyun Kim, Hyunseo Kim, Minu Kim, Myoungshin Kim, Youchul Kim, Byungoh Ko, Christopher Lee, Edward Hwayoung Lee, Honglak Lee, Jiyoung Lee, Sangeun Lee, Seungwon Lim, Woohyung Lim, Jueun Mun, Jaewoo Park, Jimin Park, Jinho Park, Yongmin Park, Wooseok Seo, Yongwoo Song, Sihyuk Yi, Kyungjae Yoo, Sangyeon Yoon
Computation and Language · Cross-listed from cs.CV
Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts
Haolei Xu, Haiwen Hong, Hongxing Li et al. (+7 more)
Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on
vision-
language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent
reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic
alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant
reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three
multimodal MoE models across six
benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual
reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures.
Read the full abstract ▾
All authors Haolei Xu, Haiwen Hong, Hongxing Li, Rui Zhou, Yang Zhang, Longtao Huang, Hui Xue, Yongliang Shen, Weiming Lu, Yueting Zhuang
Computation and Language · Cross-listed from cs.CV
AVGen-Bench: A Task-Driven Benchmark for Multi-Granular Evaluation of Text-to-Audio-Video Generation
Ziwei Zhou, Zeyuan Lai, Rui Wang et al. (+6 more)
Text-to-Audio-
Video (T2AV) generation is rapidly becoming a core interface for media creation, yet its evaluation remains fragmented. Existing
benchmarks largely assess audio and
video in isolation or rely on coarse
embedding similarity, failing to capture the fine-grained joint correctness required by realistic
prompts. We introduce AVGen-Bench, a task-driven
benchmark for T2AV generation featuring high-quality
prompts across 11 real-world categories. To support comprehensive assessment, we propose a multi-granular evaluation framework that combines lightweight specialist models with
Multimodal Large
Language Models (MLLMs), enabling evaluation from perceptual quality to fine-grained semantic controllability. Our evaluation reveals a pronounced gap between strong audio-visual aesthetics and weak semantic reliability, including persistent failures in text rendering,
speech coherence, physical
reasoning, and a universal breakdown in musical pitch control. Code and
benchmark resources are available at http://aka.ms/avgenbench.
Read the full abstract ▾
All authors Ziwei Zhou, Zeyuan Lai, Rui Wang, Yifan Yang, Zhen Xing, Yuqing Yang, Qi Dai, Lili Qiu, Chong Luo
Computation and Language · Cross-listed from cs.CV
OpenVLThinkerV2: A Generalist Multimodal Reasoning Model for Multi-domain Visual Tasks
Wenbo Hu, Xin Chen, Yan Gao-Tian et al. (+3 more)
Group Relative Policy Optimization (GRPO) has emerged as the de facto Reinforcement Learning (RL) objective driving recent advancements in
Multimodal Large
Language Models. However, extending this success to open-source
multimodal generalist models remains heavily constrained by two primary challenges: the extreme variance in reward
topologies across diverse visual tasks, and the inherent difficulty of balancing fine-grained perception with multi-step
reasoning capabilities. To address these issues, we introduce Gaussian GRPO (G$^2$RPO), a novel RL training objective that replaces standard linear scaling with non-linear distributional matching. By mathematically forcing the advantage distribution of any given task to strictly converge to a standard normal distribution, $\mathcal{N}(0,1)$, G$^2$RPO theoretically ensures inter-task gradient equity, mitigates vulnerabilities to heavy-tail outliers, and offers symmetric update for positive and negative rewards. Leveraging the enhanced training stability provided by G$^2$RPO, we introduce two task-level shaping mechanisms to seamlessly balance perception and
reasoning. First, response length shaping dynamically elicits extended
reasoning chains for complex queries while enforce direct outputs to bolster visual grounding. Second,
entropy shaping tightly bounds the model's exploration zone, effectively preventing both
entropy collapse and
entropy explosion. Integrating these methodologies, we present OpenVLThinkerV2, a highly robust, general-purpose
multimodal model. Extensive evaluations across 18 diverse
benchmarks demonstrate its superior performance over strong open-source and leading proprietary frontier models.
Read the full abstract ▾
All authors Wenbo Hu, Xin Chen, Yan Gao-Tian, Yihe Deng, Nanyun Peng, Kai-Wei Chang
Computation and Language
Demystifying OPD: Length Inflation and Stabilization Strategies for Large Language Models
Feng Luo, Yu-Neng Chuang, Guanchu Wang et al. (+4 more)
On-policy
distillation (OPD) trains student models under their own induced distribution while leveraging supervision from stronger teachers. We identify a failure mode of OPD: as training progresses, on-policy rollouts can undergo abrupt length inflation, causing truncated trajectories to dominate the training data. This truncation collapse coincides with abrupt repetition saturation and induces biased gradient signals, leading to severe training instability and sharp degradation in validation performance. We attribute this problem to the interaction between student-induced data collection and the
distillation objective, which implicitly favors long and repetitive rollouts. To address this issue, we propose StableOPD, a stabilized OPD framework that combines a reference-based divergence constraint with rollout mixture
distillation. These together mitigate repetition-induced length inflation and further stabilize OPD training. Across multiple math
reasoning datasets, our approach prevents truncation collapse, stabilizes training dynamics, and improves performance by 7.2% on average.
Read the full abstract ▾
All authors Feng Luo, Yu-Neng Chuang, Guanchu Wang, Zicheng Xu, Xiaotian Han, Tianyi Zhang, Vladimir Braverman
Computation and Language · Cross-listed from cs.AI
Ads in AI Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest
Addison J. Wu, Ryan Liu, Shuyue Stella Li et al. (+2 more)
Large
language models (LLMs) are trained to align with user preferences through methods like reinforcement learning. Yet models are beginning to be deployed not solely to satisfy users, but to generate revenue for the companies that created them through advertisements. This creates the potential for LLMs to face conflicts of interest, where the most beneficial response to a user may not be aligned with the company's incentives. For instance, a sponsored product may be more expensive but otherwise equal to another; here, what does (and should) the LLM recommend to the user? In this paper, we provide a framework for categorizing the ways in which conflicting incentives might change how LLMs interact with users, inspired by literature from linguistics and advertising regulation. We then present a suite of evaluations to examine how current models handle these tradeoffs. A majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations, including recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchasing process (GPT 5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%). Behaviors vary strongly with levels of
reasoning and users' inferred socio-economic status. Our results highlight some hidden risks to users that can emerge when companies begin to subtly incentivize advertisements in chatbots.
Read the full abstract ▾
All authors Addison J. Wu, Ryan Liu, Shuyue Stella Li, Yulia Tsvetkov, Thomas L. Griffiths
Computation and Language · Cross-listed from cs.LG
What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal
Stephen Cheng, Sarah Wiegreffe and Dinesh Manocha
Applying steering vectors to large
language models (LLMs) is an efficient and effective model
alignment technique, but we lack an interpretable explanation for how it works–specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-
token activation patching framework and discover that different steering methodologies leverage functionally interchangeable
circuits when applied at the same layer. These
circuits reveal that steering vectors primarily interact with the
attention mechanism through the OV
circuit while largely ignoring the QK
circuit. Freezing all
attention scores during steering drops performance by only 8.83% across three model families. A mathematical decomposition of the steered OV
circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 85-96% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.
Read the full abstract ▾
Computation and Language
ClawBench: Can AI Agents Complete Everyday Online Tasks?
Yuxuan Zhang, Yubo Wang, Yipeng Zhu et al. (+25 more)
AI
agents may be able to assist with emails and documents, but can they reliably complete everyday online workflows on real websites? Everyday online tasks offer a realistic yet unsolved testbed for evaluating the next generation of AI
agents. To this end, we introduce ClawBench, an evaluation framework comprising 153 everyday online tasks that people need to accomplish regularly in their lives and work, spanning 144 platforms across 15 categories, from completing purchases and booking appointments to submitting job applications. These tasks require capabilities beyond existing
benchmarks, such as obtaining relevant information from user-provided documents, navigating multi-step workflows across diverse platforms, and write-heavy operations like filling in many detailed forms correctly. Unlike existing
benchmarks that evaluate
agents in offline sandboxes with static pages, ClawBench operates on production websites, preserving the full complexity, dynamic nature, and interaction challenges of real-world web environments. An interception layer captures and blocks the final submission request, ensuring safe evaluation without real-world side effects. Our evaluations of 8 frontier models show that both proprietary and open-source models complete only a small portion of these tasks. For example, Claude Sonnet 4.6 achieves only 33.3%, which exposes gaps in current AI
agents. Progress on ClawBench brings us closer to AI
agents that can function as general-purpose assistants.
Read the full abstract ▾
All authors Yuxuan Zhang, Yubo Wang, Yipeng Zhu, Penghui Du, Junwen Miao, Xuan Lu, Zhuofeng Li, Xingwei Qu, Zhengkang Guo, Yuanzhe Shen, Dingjie Song, Han Zhou, Tuney Zheng, Xian Wu, Hao Yu, Songcheng Cai, Yi Lu, Yunzhuo Hao, Minyi Lei, Liang Chen, Kai Zou, Huifeng Yin, Wendong Xu, Dongfu Jiang, Ping Nie, Jiaheng Liu, Wenhu Chen, Kelsey R. Allen
Computation and Language
Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts
Jiayuan Ye, Vitaly Feldman and Kunal Talwar
Large
language models (LLMs) can struggle to memorize factual knowledge in their parameters, often leading to
hallucinations and poor performance on knowledge-intensive tasks. In this paper, we formalize fact memorization from an information-theoretic perspective and study how training data distributions affect fact accuracy. We show that fact accuracy is suboptimal (below the capacity limit) whenever the amount of information contained in the training data facts exceeds model capacity. This is further exacerbated when the fact frequency distribution is skewed (e.g. a power law). We propose data selection schemes based on the training loss alone that aim to limit the number of facts in the training data and flatten their frequency distribution. On semi-synthetic
datasets containing high-
entropy facts, our selection method effectively boosts fact accuracy to the capacity limit. When pretraining
language models from scratch on an annotated Wikipedia corpus, our selection method enables a GPT2-Small model (110m parameters) to memorize 1.3X more entity facts compared to standard training, matching the performance of a 10X larger model (1.3B parameters) pretrained on the full
dataset.
Read the full abstract ▾
Computation and Language
What do Language Models Learn and When? The Implicit Curriculum Hypothesis
Emmy Liu, Kaiser Sun, Millicent Li et al. (+4 more)
Large
language models (LLMs) can perform remarkably complex tasks, yet the fine-grained details of how these capabilities emerge during pretraining remain poorly understood. Scaling laws on validation loss tell us how much a model improves with additional compute, but not what skills it acquires in which order. To remedy this, we propose the Implicit Curriculum Hypothesis: pretraining follows a compositional and predictable curriculum across models and data mixtures. We test this by designing a suite of simple, composable tasks spanning
retrieval, morphological transformations, coreference, logical
reasoning, and mathematics. Using these tasks, we track emergence points across four model families spanning sizes from 410M-13B parameters. We find that emergence orderings of when models reach fixed accuracy thresholds are strikingly consistent ($\rho = .81$ across 45 model pairs), and that composite tasks most often emerge after their component tasks. Furthermore, we find that this structure is encoded in model representations: tasks with similar function vector representations also tend to follow similar trajectories in training. By using the space of representations derived from our task set, we can effectively predict the training trajectories of simple held-out compositional tasks throughout the course of pretraining ($R^2 = .68$-$.84$ across models) without previously evaluating them. Together, these results suggest that pretraining is more structured than loss curves reveal: skills emerge in a compositional order that is consistent across models and readable from their internals.
Read the full abstract ▾
All authors Emmy Liu, Kaiser Sun, Millicent Li, Isabelle Lee, Lindia Tjuatja, Jen-tse Huang, Graham Neubig
Computation and Language · Cross-listed from stat.ML
Differentially Private Language Generation and Identification in the Limit
Anay Mehrotra, Grigoris Velegkas, Xifan Yu et al. (+1 more)
We initiate the study of
language generation in the limit, a model recently introduced by Kleinberg and Mullainathan [KM24], under the constraint of differential
privacy. We consider the continual release model, where a generator must eventually output a stream of valid strings while protecting the
privacy of the entire input sequence. Our first main result is that for countable collections of
languages,
privacy comes at no qualitative cost: we provide an $\varepsilon$-differentially-private
algorithm that generates in the limit from any countable collection. This stands in contrast to many learning settings where
privacy renders learnability impossible. However,
privacy does impose a quantitative cost: there are finite collections of size $k$ for which uniform private generation requires $\Omega(k/\varepsilon)$ samples, whereas just one sample suffices non-privately. We then turn to the harder problem of
language identification in the limit. Here, we show that
privacy creates fundamental barriers. We prove that no $\varepsilon$-DP
algorithm can identify a collection containing two
languages with an infinite intersection and a finite set difference, a condition far stronger than the classical non-private characterization of identification. Next, we turn to the
stochastic setting where the sample strings are sampled i.i.d. from a distribution (instead of being generated by an adversary). Here, we show that private identification is possible if and only if the collection is identifiable in the
adversarial model. Together, our results establish new dimensions along which generation and identification differ and, for identification, a separation between
adversarial and
stochastic settings induced by
privacy constraints.
Read the full abstract ▾
All authors Anay Mehrotra, Grigoris Velegkas, Xifan Yu, Felix Zhou
Computation and Language · Cross-listed from cs.DL
sciwrite-lint: Verification Infrastructure for the Age of Science Vibe-Writing
Sergey V Samsonau
Scientific papers make claims about prior work backed by citations. Verifying those citations at scale (that each cited paper exists, says what the citation claims, and is itself reliable) is structurally beyond what human review can deliver: a typical paper has dozens of citations, and a careful reviewer reads at most a handful end-to-end. AI-assisted writing makes this gap even more urgent: LLMs hallucinate references and may fill in plausible details from titles or abstracts of papers they never read, worse for the smaller local-weights models that
privacy-aware researchers must use. sciwrite-lint applies the linting paradigm from software engineering to citation verification: it runs entirely on the researcher's machine (free public databases, a single consumer GPU, and open-weights models), is fast enough to re-lint between revisions so authors catch problems at the source while drafting, and serves journals and reviewers as an automated first pass. The pipeline checks reference existence, metadata accuracy, retraction status, and claim support, traverses one level into cited papers' bibliographies, and produces per-reference reliability scores. We evaluate on 30 unseen papers (arXiv and bioRxiv) with error injection and LLM-adjudicated false-positive analysis. The same linting workflow extends to internal consistency: numbers in text vs. tables, abstract vs. body, figure captions vs. content, statistical results vs. their verbal interpretation, plus structural cross-references (dangling cites, orphan references). As a separate experimental contribution we also propose SciLint Score: citation-chain integrity combined with a contribution component operationalizing five philosophy-of-science frameworks (Popper, Lakatos, Kitcher, Laudan, Mayo).
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CR
PIArena: A Platform for Prompt Injection Evaluation
Runpeng Geng, Chenlong Yin, Yanting Wang et al. (+2 more)
Prompt injection attacks pose serious security risks across a wide range of real-world applications. While receiving increasing
attention, the community faces a critical gap: the lack of a unified platform for
prompt injection evaluation. This makes it challenging to reliably compare defenses, understand their true robustness under diverse attacks, or assess how well they generalize across tasks and
benchmarks. For instance, many defenses initially reported as effective were later found to exhibit limited robustness on diverse
datasets and attacks. To bridge this gap, we introduce PIArena, a unified and extensible platform for
prompt injection evaluation that enables users to easily integrate state-of-the-art attacks and defenses and evaluate them across a variety of existing and new
benchmarks. We also design a dynamic strategy-based attack that adaptively optimizes injected
prompts based on defense feedback. Through comprehensive evaluation using PIArena, we uncover critical limitations of state-of-the-art defenses: limited generalizability across tasks, vulnerability to adaptive attacks, and fundamental challenges when an injected task aligns with the target task. The code and
datasets are available at https://github.com/sleeepeer/PIArena.
Read the full abstract ▾
All authors Runpeng Geng, Chenlong Yin, Yanting Wang, Ying Chen, Jinyuan Jia
Computation and Language · Cross-listed from cs.CV
What They Saw, Not Just Where They Looked: Semantic Scanpath Similarity via VLMs and NLP metric
Mohamed Amine Kerkouri, Marouane Tliba, Bin Wang et al. (+3 more)
Scanpath similarity metrics are central to eye-movement research, yet existing methods predominantly evaluate spatial and temporal
alignment while neglecting semantic equivalence between attended image regions. We present a semantic scanpath similarity framework that integrates
vision-
language models (VLMs) into eye-tracking analysis. Each fixation is encoded under controlled visual context (patch-based and marker-based strategies) and transformed into concise textual descriptions, which are aggregated into scanpath-level representations. Semantic similarity is then computed using
embedding-based and lexical NLP metrics and compared against established spatial measures, including MultiMatch and DTW. Experiments on free-viewing eye-tracking data demonstrate that semantic similarity captures partially independent variance from
geometric alignment, revealing cases of high content agreement despite spatial divergence. We further analyze the impact of contextual encoding on description fidelity and metric stability. Our findings suggest that
multimodal foundation models enable interpretable, content-aware extensions of classical scanpath analysis, providing a complementary dimension for gaze research within the ETRA community.
Read the full abstract ▾
All authors Mohamed Amine Kerkouri, Marouane Tliba, Bin Wang, Aladine Chetouani, Ulas Bagci, Alessandro Bruno
Computation and Language · Cross-listed from cs.IT
Formalizing building-up constructions of self-dual codes through isotropic lines in Lean
Jae-Hyun Baek and Jon-Lark Kim
The purpose of this paper is two-fold. First, we show that, after a specified form isometry, the two-coordinate reduction in the binary
Hilbert-symbol realization of Chinburg and Zhang is inverse to Kim's building-up construction, up to permutation equivalence. Second, for $q\equiv1\pmod4$, we develop a $q$-ary analogue of this reduction-and-extension mechanism. The identity $c^2=-1$ yields the isotropic line governing the split construction. For every fixed ordered pairing of the coordinates, we obtain a universal rank-$r$ boxed normal form, where $r$ is the dimension of the intersection with the product of these isotropic lines. Applications include optimal self-dual $[6,3,4]$ and $[8,4,4]$ codes over $\mathbb F_{5}$, optimal self-dual $[8,4,5]$ and $[10,5,6]$ codes over $\mathbb F_{13}$, and a self-dual $[12,6,6]$ code over $\mathbb F_{13}$. We also give an exact repeated boxed realization of self-dual $[18,9,8]$ and $[20,10,10]$ codes over $\mathbb F_{13}$, in which the split-boxed parent and its building-up child occur in one complete generator
matrix. The algebraic core is formalized in Lean 4.
Read the full abstract ▾
Computation and Language
AI generates well-liked but templatic empathic responses
Emma S. Gueorguieva, Hongli Zhan, Jina Suh et al. (+4 more)
Recent research shows that greater numbers of people are turning to Large
Language Models (LLMs) for emotional support, and that people rate LLM responses as more empathic than human-written responses. We suggest a reason for this success: LLMs have learned and consistently deploy a well-liked template for expressing empathy. We develop a taxonomy of 10 empathic
language "tactics" that include validating someone's feelings and paraphrasing, and apply this taxonomy to characterize the
language that people and LLMs produce when writing empathic responses. Across a set of 2 studies comparing a total of n = 3,265 AI-generated (by six models) and n = 1,290 human-written responses, we find that LLM responses are highly formulaic at a discourse functional level. We discovered a template – a structured sequence of tactics – that matches between 83–90% of LLM responses (and 60–83% in a held out sample), and when those are matched, covers 81–92% of the response. By contrast, human-written responses are more diverse. We end with a discussion of implications for the future of AI-generated empathy.
Read the full abstract ▾
All authors Emma S. Gueorguieva, Hongli Zhan, Jina Suh, Javier Hernandez, Tatiana Lau, Junyi Jessy Li, Desmond C. Ong
Computation and Language · Cross-listed from cs.AI
SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions
Ashima Suvarna, Kendrick Phan, Mehrab Beikzadeh et al. (+2 more)
Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved
reasoning in formal domains such as mathematics and code, but extending these gains beyond STEM remains challenging. Extending RLVR beyond STEM is fundamentally constrained by the lack of high-quality verifiable training data. In this work, we introduce
SUPERNOVA, a framework for curating RLVR data from natural instruction
datasets, which are a rich source of expert-annotated data but are underexplored for RLVR training. Through 100+ controlled RL experiments, we systematically study how to utilize these
dataset for RLVR and how data curation decisions affect downstream
reasoning performance . In particular, we investigate three data designs: (a) source task selection, (b) task mixing, and (c) synthetic interventions. Our analysis reveals that source task selection has a significant impact on downstream
reasoning performance. Moreover, selecting tasks based on their performance for individual target tasks outperforms strategies based on overall average performance and synthetic interventions do not improve
reasoning. Guided by these insights, we construct
SUPERNOVA, a high-quality RLVR
dataset of 25K instances curated from natural instruction
datasets. We show that training Qwen3-0.6B on
SUPERNOVA outperforms the base Qwen3-0.6B, yielding a relative gain of 64.4pp on BigBench Extra Hard (BBEH), a challenging
benchmark comprising 23 complex
reasoning tasks. Importantly, we find that gains from
SUPERNOVA generalize to unseen
benchmarks, larger model scales, and newer model families. Overall, our findings provide practical insights for curating human-annotated resources to extend RLVR to general
reasoning. Models, Data, Code at https://github.com/asuvarna31/
supernova.
Read the full abstract ▾
All authors Ashima Suvarna, Kendrick Phan, Mehrab Beikzadeh, Hritik Bansal, Saadia Gabriel
Computation and Language · Cross-listed from cs.CV
Marcel Gröpl, Jaewoo Jung, Seungryong Kim et al. (+2 more)
Despite rapid progress, pretrained
vision-
language models still struggle when answers depend on tiny visual details or on combining clues spread across multiple regions, as in documents and compositional queries. We address this by framing grounding as test-time evidence
retrieval: given a query, the model should actively identify where to look next to resolve ambiguity. To this end, we propose a training-free, model-intrinsic grounding method that uses uncertainty as supervision. Specifically, we compute the
entropy of the model's next-
token distribution and backpropagate it to the visual
token embeddings to obtain an
entropy-gradient relevance map, without auxiliary detectors or
attention-map heuristics. We then extract and rank multiple coherent regions to support multi-evidence queries, and introduce an iterative zoom-and-reground procedure with a spatial-
entropy stopping rule to avoid over-refinement. Experiments on seven
benchmarks across four VLM architectures demonstrate consistent improvements over existing methods, with the largest gains on detail-critical and high-resolution settings, while also producing more interpretable evidence localizations.
Read the full abstract ▾
All authors Marcel Gröpl, Jaewoo Jung, Seungryong Kim, Marc Pollefeys, Sunghwan Hong
Computation and Language
Lilian Wanzare, Cynthia Amol, Ezekiel Maina et al. (+14 more)
AfriVoices-KE is a large-scale
multilingual speech dataset comprising approximately 3,000 hours of audio across five Kenyan
languages: Dholuo, Kikuyu, Kalenjin, Maasai, and Somali. The
dataset includes 750 hours of scripted
speech and 2,250 hours of spontaneous
speech, collected from 4,777 native speakers across diverse regions and demographics. This work addresses the critical underrepresentation of African
languages in
speech technology by providing a high-quality, linguistically diverse resource. Data collection followed a dual methodology: scripted recordings drew from compiled text corpora, translations, and domain-specific generated sentences spanning eleven domains relevant to the Kenyan context, while unscripted
speech was elicited through textual and image
prompts to capture natural linguistic variation and dialectal nuances. A customized mobile application enabled contributors to record using smartphones. Quality assurance operated at multiple layers, encompassing automated signal-to-noise ratio validation prior to recording and human review for content accuracy. Though the project encountered challenges common to low-resource settings, including unreliable infrastructure, device compatibility issues, and community trust barriers, these were mitigated through local mobilizers, stakeholder partnerships, and adaptive training protocols. AfriVoices-KE provides a foundational resource for developing inclusive automatic
speech recognition and text-to-
speech systems, while advancing the digital preservation of Kenya's linguistic heritage.
Read the full abstract ▾
All authors Lilian Wanzare, Cynthia Amol, Ezekiel Maina, Nelson Odhiambo, Hope Kerubo, Leila Misula, Vivian Oloo, Rennish Mboya, Edwin Onkoba, Edward Ombui, Joseph Muguro, Ciira wa Maina, Andrew Kipkebut, Alfred Omondi Otom, Ian Ndung'u Kang'ethe, Angela Wambui Kanyi, Brian Gichana Omwenga
Computation and Language · Cross-listed from cs.LG
KV Cache Offloading for Context-Intensive Tasks
Andrey Bocharnikov, Ivan Ermakov, Denis Kuznedelev et al. (+2 more)
With the growing demand for long-context LLMs across a wide range of applications, the key-value (KV) cache has become a critical bottleneck for both latency and
memory usage. Recently, KV-cache offloading has emerged as a promising approach to reduce
memory footprint and inference latency while preserving accuracy. Prior evaluations have largely focused on tasks that do not require extracting large amounts of information from the context. In this work, we study KV-cache offloading on context-intensive tasks: problems where the solution requires looking up a lot of information from the input
prompt. We create and release the Text2JSON
benchmark, a highly context-intensive task that requires extracting structured knowledge from raw text. We evaluate modern KV offloading on Text2JSON and other context-intensive tasks and find significant performance degradation on both Llama 3 and Qwen 3 models. Our analysis identifies two key reasons for poor accuracy: low-rank projection of keys and unreliable landmarks, and proposes a simpler alternative strategy that significantly improves accuracy across multiple LLM families and
benchmarks. These findings highlight the need for a comprehensive and rigorous evaluation of long-context compression techniques.
Read the full abstract ▾
All authors Andrey Bocharnikov, Ivan Ermakov, Denis Kuznedelev, Vyacheslav Zhdanovskiy, Yegor Yershov
Computation and Language · Cross-listed from cs.AI
Learning Who Disagrees: Demographic Importance Weighting for Modeling Annotator Distributions with DiADEM
Samay U. Shetty, Tharindu Cyril Weerasooriya, Deepak Pandita et al. (+1 more)
When humans label subjective content, they disagree, and that disagreement is not noise. It reflects genuine differences in perspective shaped by annotators' social identities and lived experiences. Yet standard practice still flattens these judgments into a single majority label, and recent LLM-based approaches fare no better: we show that prompted large
language models, even with chain-of-thought
reasoning, fail to recover the structure of human disagreement. We introduce DiADEM, a neural architecture that learns "how much each demographic axis matters" for predicting who will disagree and on what. DiADEM encodes annotators through per-demographic projections governed by a learned importance vector $\boldsymbol{\alpha}$, fuses annotator and item representations via complementary concatenation and Hadamard interactions, and is trained with a novel item-level disagreement loss that directly penalizes mispredicted annotation variance. On the DICES conversational-safety and VOICED political-offense
benchmarks, DiADEM substantially outperforms both the LLM-as-a-judge and neural model baselines across standard and perspectivist metrics, achieving strong disagreement tracking ($r{=}0.75$ on DICES). The learned $\boldsymbol{\alpha}$ weights reveal that race and age consistently emerge as the most influential demographic factors driving annotator disagreement across both
datasets. Our results demonstrate that explicitly modeling who annotators are not just what they label is essential for NLP systems that aim to faithfully represent human interpretive diversity.
Read the full abstract ▾
All authors Samay U. Shetty, Tharindu Cyril Weerasooriya, Deepak Pandita, Christopher M. Homan
Computation and Language
Synthetic Data for any Differentiable Target
Tristan Thrush, Sung Min Park, Herman Brunborg et al. (+5 more)
What are the limits of controlling
language models via synthetic training data? We develop a reinforcement learning (RL) primitive, the
Dataset Policy Gradient (DPG), which can precisely optimize synthetic data generators to produce a
dataset of targeted examples. When used for supervised fine-tuning (SFT) of a target model, these examples cause the target model to do well on a differentiable metric of our choice. Our approach achieves this by taking exact data attribution via higher-order gradients and using those scores as policy gradient rewards. We prove that this procedure closely approximates the true, intractable gradient for the synthetic data generator. To illustrate the potential of DPG, we show that, using only SFT on generated examples, we can cause the target model's LM head weights to (1) embed a QR code, (2) embed the pattern $\texttt{67}$, and (3) have lower $\ell^2$ norm. We additionally show that we can cause the generator to (4) rephrase inputs in a new
language and (5) produce a specific UUID, even though neither of these objectives is conveyed in the generator's input
prompts. These findings suggest that DPG is a powerful and flexible technique for shaping model properties using only synthetic training examples.
Read the full abstract ▾
All authors Tristan Thrush, Sung Min Park, Herman Brunborg, Luke Bailey, Marcel Roed, Neil Band, Christopher Potts, Tatsunori Hashimoto
Computation and Language · Cross-listed from cs.AI
Verify Before You Commit: Towards Faithful Reasoning in LLM Agents via Self-Auditing
Wenhao Yuan, Chenchen Lin, Jian Chen et al. (+3 more)
In large
language model (LLM)
agents,
reasoning trajectories are treated as reliable internal beliefs for guiding actions and updating
memory. However, coherent
reasoning can still violate logical or evidential constraints, allowing unsupported beliefs repeatedly stored and propagated across decision steps, leading to systematic behavioral drift in long-horizon agentic systems. Most existing strategies rely on the consensus mechanism, conflating agreement with faithfulness. In this paper, inspired by the vulnerability of unfaithful intermediate
reasoning trajectories, we propose Self-Audited Verified
Reasoning (SAVeR), a novel framework that enforces verification over internal belief states within the
agent before action commitment, achieving faithful
reasoning. Concretely, we structurally generate persona-based diverse candidate beliefs for selection under a faithfulness-relevant structure space. To achieve
reasoning faithfulness, we perform
adversarial auditing to localize violations and repair through constraint-guided minimal interventions under verifiable acceptance criteria. Extensive experiments on six
benchmark datasets demonstrate that our approach consistently improves
reasoning faithfulness while preserving competitive end-task performance.
Read the full abstract ▾
All authors Wenhao Yuan, Chenchen Lin, Jian Chen, Jinfeng Xu, Xuehe Wang, Edith Cheuk Han Ngai
Computation and Language
A GAN and LLM-Driven Data Augmentation Framework for Dynamic Linguistic Pattern Modeling in Chinese Sarcasm Detection
Wenxian Wang, Xiaohu Luo, Junfeng Hao et al. (+4 more)
Sarcasm is a rhetorical device that expresses criticism or emphasizes characteristics of certain individuals or situations through exaggeration, irony, or comparison. Existing methods for Chinese sarcasm detection are constrained by limited
datasets and high construction costs, and they mainly focus on textual features, overlooking user-specific linguistic patterns that shape how opinions and
emotions are expressed. This paper proposes a Generative
Adversarial Network (GAN) and Large
Language Model (LLM)-driven data augmentation framework to dynamically model users' linguistic patterns for enhanced Chinese sarcasm detection. First, we collect raw data from various topics on Sina Weibo. Then, we train a GAN on these data and apply a GPT-3.5 based data augmentation technique to synthesize an extended sarcastic comment
dataset, named SinaSarc. This
dataset contains target comments, contextual information, and user historical behavior. Finally, we extend the BERT architecture to incorporate multi-dimensional information, particularly user historical behavior, enabling the model to capture dynamic linguistic patterns and uncover implicit sarcastic cues in comments. Experimental results demonstrate the effectiveness of our proposed method. Specifically, our model achieves the highest F1-scores on both the non-sarcastic and sarcastic categories, with values of 0.9138 and 0.9151 respectively, which outperforms all existing state-of-the-art (SOTA) approaches. This study presents a novel framework for dynamically modeling users' long-term linguistic patterns in Chinese sarcasm detection, contributing to both
dataset construction and methodological advancement in this field.
Read the full abstract ▾
All authors Wenxian Wang, Xiaohu Luo, Junfeng Hao, Xiaoming Gu, Xingshu Chen, Zhu Wang, Haizhou Wang
Computation and Language · Cross-listed from cs.AI
SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
Ziyu Ma, Shidong Yang, Yuxiang Ji et al. (+5 more)
Large
language model (LLM)
agents such as OpenClaw rely on reusable skills to perform complex tasks, yet these skills remain largely static after deployment. As a result, similar workflows, tool usage patterns, and failure modes are repeatedly rediscovered across users, preventing the system from improving with experience. While interactions from different users provide complementary signals about when a skill works or fails, existing systems lack a mechanism to convert such heterogeneous experiences into reliable skill updates. To address these issues, we present SkillClaw, a framework for collective skill
evolution in multi-user
agent ecosystems, which treats cross-user and over-time interactions as the primary signal for improving skills. SkillClaw continuously aggregates trajectories generated during use and processes them with an autonomous evolver, which identifies recurring behavioral patterns and translates them into updates to the skill set by refining existing skills or extending them with new capabilities. The resulting skills are maintained in a shared repository and synchronized across users, allowing improvements discovered in one context to propagate system-wide while requiring no additional effort from users. By integrating multi-user experience into ongoing skill updates, SkillClaw enables cross-user knowledge transfer and cumulative capability improvement, and experiments on WildClawBench show that limited interaction and feedback, it significantly improves the performance of Qwen3-Max in real-world
agent scenarios.
Read the full abstract ▾
All authors Ziyu Ma, Shidong Yang, Yuxiang Ji, Xucong Wang, Yong Wang, Yiming Hu, Tongwen Huang, Xiangxiang Chu
Computation and Language · Cross-listed from cs.AI
Don't Overthink It: Inter-Rollout Action Agreement as a Free Adaptive-Compute Signal for LLM Agents
Khushal Sethi
Inference-time compute scaling has emerged as a powerful technique for improving the reliability of large
language model (LLM)
agents, but existing methods apply compute uniformly: every decision step receives the same budget regardless of its difficulty. We introduce TrACE (Trajectorical Adaptive Compute via agrEement), a training-free controller that allocates LLM calls adaptively across
agent timesteps by measuring inter-rollout action agreement. At each step, TrACE samples a small set of candidate next actions and measures how consistently the model commits to the same action. High agreement signals an easy decision; the controller commits immediately. Low agreement signals uncertainty; the controller samples additional rollouts up to a configurable cap before committing to the plurality action. No learned components, no external verifier, and no human labels are required. We evaluate TrACE against greedy decoding and fixed-budget self-consistency (SC-4, SC-8) on two
benchmarks spanning single-step
reasoning (GSM8K, n=50) and multi-step household
navigation (MiniHouse, n=30), using a Qwen 2.5 3B Instruct model running on CPU. TrACE-4 matches SC-4 accuracy while using 33% fewer LLM calls on GSM8K and 39% fewer on MiniHouse. TrACE-8 matches SC-8 accuracy with 55% fewer calls on GSM8K and 65% fewer on MiniHouse. We further show that inter-rollout agreement is a reliable signal of step-level success, validating the core hypothesis that the model's own output consistency encodes difficulty information that can be exploited without training. TrACE is the first training-free, per-timestep adaptive-compute controller for LLM
agents to be evaluated on multi-step sequential decision tasks.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
SOLAR: Communication-Efficient Model Adaptation via Subspace-Oriented Latent Adapter Reparametrization
Seyed Mahmoud Sajjadi Mohammadabadi, Xiaolong Ma, Lei Yang et al. (+2 more)
Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, enable scalable adaptation of foundation models by injecting low-rank adapters. However, their communication and storage costs remain a major bottleneck in resource-constrained settings. We propose
SOLAR (Subspace-Oriented Latent Adapter Reparameterization), a post-training compression framework that substantially reduces the communication cost (i.e., the number of parameters to transmit or store) of PEFT adapters.
SOLAR expresses each PEFT update as a linear combination of basis vectors formed from the foundation model's singular vectors with controlled random perturbations. By exploiting the subspace similarity (the
alignment of principal directions) between the foundation model and task-specific fine-tuned updates,
SOLAR decouples the adapter size from PEFT structure and ensures compact yet expressive representations. It is model-agnostic and compatible with existing PEFT methods, including LoRA, AdaLoRA, and other adapter modules. We theoretically establish a bound on the reconstruction error. Experiments on
language and
vision tasks using LLaMA, GPT, and ViT models demonstrate that
SOLAR preserves task performance while significantly reducing model representation sizes, offering an effective and communication-efficient solution for deployment in distributed systems and edge devices.
Read the full abstract ▾
All authors Seyed Mahmoud Sajjadi Mohammadabadi, Xiaolong Ma, Lei Yang, Feng Yan, Junshan Zhang
Computation and Language
Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces
Jiawei Chen, Ruoxi Xu, Boxi Cao et al. (+12 more)
The emergence of Large
Language Models (LLMs) has illuminated the potential for a general-purpose user simulator. However, existing
benchmarks remain constrained to isolated scenarios, narrow action spaces, or synthetic data, failing to capture the holistic nature of authentic human behavior. To bridge this gap, we introduce OmniBehavior, the first predictive user simulation
benchmark constructed entirely from real-world data, integrating long-horizon, cross-scenario, and heterogeneous behavioral patterns into a unified framework. Based on this
benchmark, we first provide empirical evidence that previous
datasets with isolated scenarios suffer from tunnel
vision, whereas real-world decision-making relies on long-term, cross-scenario causal chains. Extensive evaluations of state-of-the-art LLMs reveal that current models struggle to accurately predict these complex behaviors, with performance plateauing even as context windows expand. Crucially, a systematic comparison between simulated and authentic behaviors uncovers a fundamental structural bias: LLMs tend to converge towards a positive average person, exhibiting hyper-activity, persona homogenization, and a utopian bias. This results in the loss of individual differences and long-tail behaviors, highlighting critical directions for future high-fidelity simulation research.
Read the full abstract ▾
All authors Jiawei Chen, Ruoxi Xu, Boxi Cao, Ruotong Pan, Yunfei Zhang, Yifei Hu, Yong Du, Tingting Gao, Han Li, Yaojie Lu, Yingfei Sun, Xianpei Han, Le Sun, Xiangyu Wu, Hongyu Lin