Computation and Language · Cross-listed from cs.LG
TRACE: A Unified Rollout Budget Allocation Framework for Efficient Agentic Reinforcement Learning
Heming Zou, Qi Wang, Yun Qu et al. (+9 more)
Reinforcement learning with verifiable rewards (RLVR) is a promising approach for enhancing
reasoning and agentic behavior in large
language models. However, rollout-intensive policy optimization is often limited by insufficient reward contrast, arising when overly simple or complex
prompts generate low-variance feedback and when outcome-only rewards assign the same terminal assessment to every decision in a multi-turn rollout. Past efforts have focused on allocating available rollout resources to promising
prompts, yet they only leverage sample informativeness at the
prompt level and neglect variation in prefix-level informativeness across turns within the same rollout. This work targets multi-turn agentic RL by modeling each ReAct-style thought-action-observation turn as a semantically distinct node, allowing budget allocation to extend from
prompt roots to turn-level prefixes with further continuations, which naturally forms tree-structured rollouts. We introduce Tree Rollout Allocation for Contrastive Exploration (TRACE), a unified rollout allocation framework that enhances reward contrast within a fixed sampling budget. Technically, TRACE allocates rollout budget to both
prompt roots and intermediate prefixes that are most likely to yield mixed terminal rewards. A shared generalizable predictor estimates conditional success probability at these anchors from prefix histories to guide this allocation. The resulting adaptive tree structure enriches outcome-only feedback and amplifies the policy-update signal. Empirically, TRACE achieves competitive performance and efficiency gains on typical agentic
benchmarks, e.g., improving Qwen3-14B Multi-Hop QA average accuracy by 2.8 points over competitive baselines at equal sampling cost.
Read the full abstract ▾
All authors Heming Zou, Qi Wang, Yun Qu, Yuhang Jiang, Lizhou Cai, Yixiu Mao, Ru Peng, Xin Xu, Weijie Liu, Kai Yang, Saiyong Yang, Xiangyang Ji
Computation and Language
PhantomBench: Benchmarking the Non-existential Threat of Language Models
Haeji Jung and Hila Gonen
Hallucinations, where
language models (LMs) generate factually ungrounded responses, pose serious risks, as users tend to blindly rely on them. This is particularly concerning in high-stakes domains, where consequences of such model behavior can lead to significant harms. Despite notable progress in understanding
hallucinations, it remains unclear how reliably these models can recognize the limits of their knowledge. We introduce PhantomBench, the first large-scale
benchmark of its kind, comprising more than 60K non-existent terms and entities derived from real concepts across diverse domains. Using our
benchmark, we evaluate a total of 21 models of various types and sizes. We show staggering
hallucination rates across the board (with average rates as high as 86.7% in some cases), and note that even frontier models surprisingly fail to abstain on non-existent concepts, especially when the input presumes their existence. We then show that PhantomBench can serve as a proxy for studying model behavior on rare concepts for which models are more prone to hallucinate. We also provide a pipeline to construct PhantomBench, enabling scalable generation of non-existent concepts tailored to the specific needs of researchers and practitioners.
Read the full abstract ▾
Computation and Language
The Shibboleth Effect: Auditing the Cross-Lingual Distributional Skew of Large Language Models
Hakan Mehmetcik
This study investigates cross-lingual distributional skew (the Shibboleth Effect) in frontier large
language models (LLMs) subjected to sustained
adversarial conditions. We develop a multi-
agent geopolitical wargame, the Cerulean Sea Crisis, a synthetic maritime territorial dispute designed to mirror the structural dynamics of Eastern Mediterranean conflicts. Six frontier models (GPT-4o, Llama-4, Mistral-Large, Gemini-3.1-Pro, Qwen3.6-Plus, and DeepSeek-R1) participate in a between-groups experiment (N = 10 games per arm, K = 5 rounds per game) in which the sole manipulation is the
language of play (English versus Turkish), producing 586 validated statements. A zero-shot classifier assesses behavioral dispositions along two continuous dimensions: Concession Rate and Coercive Rhetoric. The results are heterogeneous. Llama-4 shows a substantial, Holm-corrected increase in coercive rhetoric under Turkish (delta = +0.800, p = .002), whereas Gemini-3.1-Pro displays an equally large decrease (delta = -0.750, p = .005). DeepSeek-R1 exhibits a similar negative shift (delta = -0.860, p = .006) and provides chain-of-thought evidence consistent with a buffering mechanism. GPT-4o shows no detectable effect (delta = +0.130, p = .614). These findings indicate that cross-lingual behavioral skew is contingent on model architecture and training regime rather than a universal property of Western-origin LLMs. We identify two distinct buffering mechanisms, chain-of-thought institutional anchoring and
multilingual RLHF
alignment, and discuss their implications for integrating LLMs safely into diplomatic and crisis-management settings.
Read the full abstract ▾
Computation and Language
VISTA: A Versatile Interactive User Simulation Toolkit for Agent Evaluation
Yunan Lu, Ryan Shea, Yusen Zhang et al. (+1 more)
Evaluation remains a critical bottleneck for interactive
agent development. Existing evaluation methods often rely on static
benchmarks, which fail to capture the dynamic, multi-step nature of agentic behavior and struggle to expose meaningful failure modes. While user-simulation-based evaluation offers a promising alternative, existing simulation frameworks suffer from two major limitations. First, they provide limited mechanisms for evaluating the quality and comprehensiveness of simulated interactions, making it difficult to assess whether a simulator sufficiently explores an
agent's capabilities and failure modes. Second, most frameworks are restricted to either UI-only actions or API-only actions, limiting their ability to model the full range of realistic user behaviors. To address these limitations, we propose VISTA, a Versatile Interactive user Simulation Toolkit for
Agent evaluation. Our toolkit includes a suite of six metrics for measuring the realism, capability coverage, and interaction effectiveness of simulated interactions. In addition, we develop a hybrid user simulator that integrates both UI-based interactions and API-based interactions, enabling more realistic and comprehensive evaluation across diverse interactive environments. We evaluate VISTA in e-commerce shopping and education customer service settings and demonstrate that it produces more realistic and comprehensive evaluations than existing methods.
Read the full abstract ▾
All authors Yunan Lu, Ryan Shea, Yusen Zhang, Zhou Yu
Computation and Language · Cross-listed from cs.AI
A History-Aware Visually Grounded Critic for Computer Use Agents
Jaewoo Lee, Zaid Khan, Archiki Prasad et al. (+7 more)
Various test-time interventions for Computer Use
Agents (CUAs), including critic models, have been developed to improve performance through pre-execution action evaluation in complex Graphical User Interface (GUI) environments. However, existing critics suffer from two key limitations: they (1) focus primarily on short-sighted decision loops (e.g., forgetting earlier actions) and (2) lack the visual grounding needed to detect flawed actions (e.g., clicking wrong UI elements). To address these, we introduce HiViG, a History-aware Visually Grounded test-time framework, built around a
multimodal critic trained on real GUI trajectories to abstract past interactions into a compact record and to evaluate actions with visual grounding. At test time, HiViG integrates the critic into the policy decision loop to provide macro-action history, which summarizes the policy's completed achievements, and visually grounded critique, which verifies raw execution coordinates against the current screenshot to intercept errors before execution. Across web, mobile, and desktop
benchmarks, HiViG consistently outperforms existing scalar and verbal critics, improving average success rates over the strongest baseline by 5.8% for Qwen3-VL-32B and 9.0% for Gemini-3-Flash, and demonstrates strong cross-platform generalization. Ablations show that macro-action history mitigates short-sighted planning and visually grounded critique reduces execution errors, with both components being critical for test-time scaling in long-horizon GUI tasks.
Read the full abstract ▾
All authors Jaewoo Lee, Zaid Khan, Archiki Prasad, Justin Chih-Yao Chen, Supriyo Chakraborty, Kartik Balasubramaniam, Sambit Sahu, Elias Stengel-Eskin, Hyunji Lee, Mohit Bansal
Computation and Language
Modeling Complex Behaviors: Multi-Personality Composition and Dynamic Switching in Vision-Language Models
Peiqi Jia, Haonan Jia, Ziqi Miao et al. (+3 more)
With the widespread deployment of
Multimodal Large
Language Models (MLLMs) in social interaction, understanding and controlling their behavior under complex personality conditions is essential. This paper introduces explicit personality conditioning and establishes a systematic evaluation framework encompassing single-personality induction, multi-personality induction, and personality switching. Experiments show that personality induction improves image captioning performance but can impair performance on tasks requiring precise
reasoning, such as visual question answering (VQA). Balancing and residual effects are observed during multi-trait composition and dynamic switching, indicating that model behavior is co-modulated by both previous and current personality constraints. Existing
prompt-based personality induction methods show limited transferability to
multimodal settings. Our work reveals the dynamic and complex nature of personality modeling in MLLMs and underscores the need for robust, tailored methods for personality induction and evaluation. The code will be released when the paper is accepted.
Read the full abstract ▾
All authors Peiqi Jia, Haonan Jia, Ziqi Miao, Linkang Du, Yuntao Wang, Zhou Su
Computation and Language
T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains
Genta Indra Winata, Amartya Chakraborty, Yuzhen Lin et al. (+12 more)
Recent advances in
reasoning and tool-calling capabilities of large
language models (LLMs) have enabled increasingly capable agentic systems. However, existing
benchmarks remain limited in task complexity, realism, and domain diversity, and often fail to capture interactions that span multiple domains, limiting their ability to evaluate
agents in realistic multi-step settings that require sustained
reasoning and coordination. To address these limitations, we introduce T1-Bench, a high-fidelity, comprehensive
benchmark for evaluating agentic systems in realistic customer-facing, multi-domain environments, featuring interleaved scenarios that require structured
reasoning across multi-turn user-assistant interactions and substantially increasing both compositional complexity and evaluative rigor across 25 domains of varying difficulty. We evaluate T1-Bench using 12 proprietary and open-weight models, providing a reproducible and standardized framework for assessing
agent behavior, tool utilization, and conversational quality in complex, multi-step environments. We further complement automatic evaluation with human judgments to strengthen the assessment of qualitative performance. Overall, T1-Bench substantially advances prior
benchmarks by increasing task complexity, interaction depth, and domain coverage in simulated multi-domain environments. To facilitate future research on agentic systems, we will publicly release data and evaluation code as open source.
Read the full abstract ▾
All authors Genta Indra Winata, Amartya Chakraborty, Yuzhen Lin, Swasthi P Rao, Shikhhar Siingh, Houhan Lu, Nadia Bathaee, Sriharsha Hatwar, Paresh Dashore, Anmol Jain, Kshitij Tayal, Xiuzhu Lin, Anirban Das, Sambit Sahu, Shi-Xiong Zhang
Computation and Language
Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It
Xinyu Zhou, Boyu Zhu, Yi Xu et al. (+4 more)
Chain-of-thought (CoT) supervised fine-tuning (SFT) is widely adopted to improve
reasoning ability, yet we find that it systematically degrades long-context recall in hybrid linear-
attention models. Across architectures including HypeNet and Jet-Nemotron,
retrieval performance on Needle-In-A-Haystack (NIAH) deteriorates substantially after CoT-SFT, and the degradation becomes more severe under harder
retrieval settings and longer context windows. For example, HypeNet-9B on NIAH-S2@256K decreases from $67.2\%$ to $9.4\%$. We attribute this to CoT-SFT biasing
attention gradients toward short-range patterns, disrupting query-key projections ($W_Q, W_K$) that are responsible for long-range routing. Motivated by this observation, we propose QK-Restore, a training-free method that restores only $W_Q$ and $W_K$ from the pre-SFT checkpoint while preserving all other post-SFT parameters. We further introduce a Procrustes variant to balance routing preservation and
reasoning adaptation. Across architectures, QK-Restore consistently restores long-context capability at zero training cost while preserving
reasoning performance; for instance, on HypeNet-5B it improves S3@256K from $65.4\%$ to $76.4\%$ while maintaining strong
reasoning performance.
Read the full abstract ▾
All authors Xinyu Zhou, Boyu Zhu, Yi Xu, Zhiwei Li, Yingfa Chen, Huiming Wang, Zhijiang Guo
Computation and Language
Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models
Prajakta Kini, Avinash Reddy, Souradip Chakraborty et al. (+4 more)
Instruction-tuned LLMs are increasingly converted into
reasoning models through post-training to improve multi-step task performance. This conversion is usually optimized for
reasoning accuracy, without explicitly preserving the
alignment behavior of the instruction-tuned model, such as safe refusal, bias avoidance, and
privacy protection. We ask: does this conversion preserve
alignment? We study this question through a trustworthiness audit and find that it is not behavior-preserving by default. For a systematic analysis, we compare
reasoning models produced via supervised fine-tuning, RL-based post-training, and
distillation against matched instruction-tuned baselines across six trustworthiness dimensions: safety, toxicity, stereotyping and bias, machine ethics,
privacy, and out-of-distribution robustness. We observe that
reasoning models often improve on
reasoning benchmarks but exhibit
alignment regressions, including increased toxicity, amplified stereotyping, miscalibrated refusal, and contextual
privacy leakage. These regressions are consistent with behavioral drift from the instruction-tuned baseline, measured by KL divergence. Overall, our results point to the broader conclusion that trustworthiness metrics are essential for evaluating
reasoning models and should be reported alongside gains in
reasoning capability.
Read the full abstract ▾
All authors Prajakta Kini, Avinash Reddy, Souradip Chakraborty, Satya Sai Srinath Namburi GNVV, Furong Huang, Amrit Singh Bedi, Alvaro Velasquez
Computation and Language · Cross-listed from cs.LG
AuRA: Internalizing Audio Understanding into LLMs as LoRA
Bo Cheng, Lei Shi, Zhanyu Ma et al. (+5 more)
Recent efforts to extend large
language models (LLMs) to
speech inputs typically rely on cascaded ASR-LLM pipelines, end-to-end
speech-
language models, or bridge/
distillation-based adaptation. While these routes respectively reuse strong pretrained components, enable native
speech-
language interaction, or offer lightweight adaptation, they often suffer from transcript-interface latency, costly
multimodal training, or sequential
speech-
language coupling. To address these limitations, we present AuRA, a method that distills audio encoding capability into the LLM. Specifically, AuRA feeds the same
speech input to an ASR encoder (as a teacher) and a LoRA-adapted LLM (as a student) through a lightweight audio
embedding layer, and uses layer-wise
distillation to align the student's hidden states with corresponding teacher representations, thereby internalizing
speech representations into lightweight LLM-side adaptations. Compared with cascaded and serial bridge methods, AuRA enables tighter
speech-
language joint modeling and efficient parallel end-to-end inference, while also reusing pretrained
speech and
language models rather than requiring large-scale
multimodal training. On multiple
speech-
language benchmarks, AuRA consistently outperforms cascaded systems,
speech-to-LLM adaptation baselines, and large-scale
speech-
language and
multimodal models in both effectiveness and efficiency.
Read the full abstract ▾
All authors Bo Cheng, Lei Shi, Zhanyu Ma, Yuan Wu, Jun Xu, Jiuchong Gao, Jinghua Hao, Renqing He
Computation and Language · Cross-listed from cs.IR
Generative Archetype-Grounded Item Representations for Sequential Recommendation
Yifan Li, Jiahong Liu, Xinni Zhang et al. (+5 more)
Sequential recommendation aims to predict users' next interaction with items by analyzing their historical behavior. However, the limited quality of item representations remains a critical bottleneck. While pre-trained large
language models (LLMs) can provide rich semantic representations, existing approaches only rely on static encoding of fixed attributes, overlooking the crucial role of target audiences in defining item identity. Moreover, the semantic space struggles to reflect actual user behavior, resulting in a significant gap between semantic representations and behavioral patterns. To address these limitations, we propose GenAIR, a general framework that empowers sequential recommendation with Generative Archetype-grounded Item Representations. Specifically, we first leverage an LLM to analyze item metadata and infer textual description of the Archetype, which represents the conceptual profile of the item's ideal target audience. We then extract the corresponding
embeddings in a single forward pass. Further, to ground these generative archetypes in real-world behavior, we introduce a behavioral calibration objective, which explicitly incorporates behavioral signals from actual interactions. This objective adjusts the structure of the
embedding space to reflect empirical patterns. GenAIR enables seamless integration with most existing models while maintaining high efficiency. Comprehensive experiments conducted on three real-world
datasets demonstrate that GenAIR significantly improves the performance of various sequential recommendation models and consistently outperforms state-of-the-art baseline approaches. Implementation codes are available at https://github.com/AI-Santiago/GenAIR.
Read the full abstract ▾
All authors Yifan Li, Jiahong Liu, Xinni Zhang, Hao Chen, Yankai Chen, Wenhao Yu, Jianting Chen, Irwin King
Computation and Language
Measuring Human Value Expression in Social Media Texts: Calibrated LLM Annotation and Encoder Transfer
Maria Milkova and Maksim Rudnev
Measuring subjective constructs in naturally occurring social media text requires annotation procedures that are theoretically grounded, empirically validated, and transferable to an encoder model for scalable prediction. Using posts annotated according to Schwartz's theory of basic human values, we investigate how different LLMs and prompting strategies, which we call annotation regimes, operationalize the expression of values in text. Beyond standard classification metrics, we evaluate structural
alignment, annotation ambiguity, error patterns, and stability across repeated runs. We find that different LLMs produce different value interpretations, and iterative
prompt calibration through error analysis reduces misattributions and improves
alignment with expert annotations. Error patterns are further used to derive targeted expert-verification rules for corpus annotation. We transfer soft LLM labels to an encoder model for prediction, retaining information about ambiguity in value expression. Finally, a sensitivity analysis on more than one million posts shows that regime-specific annotation differences propagate into predicted levels of value expression, whereas standardized temporal dynamics and the direction of major event responses are more robust.
Read the full abstract ▾
Computation and Language
Who Brought Easter Eggs to Eid? Auditing LLM-Generated Cultural Translation of Math Word Problems Across Languages and Regions
Parisa Suchdev and Juniper Lovato
Large
language models are increasingly used to adapt math word problems for personalized learning at scale, but it remains an open question whether those adaptations are consistent across models, preserve cultural diversity at scale, and reveal which cultural entities models treat as most salient. We analyze how Claude Opus 4, GPT-4.1, and Gemini 2.5 Pro adapt 60 English math word problems into Bengali, Hindi, Punjabi (India), Urdu, Sindhi (Pakistan), Italian, and Sicilian (Italy), a
language set spanning the full resource spectrum, from high-resource Italian and Hindi to under-studied Sindhi, Sicilian, and Punjabi. We annotate 6,489 entity transformations, coding whether models preserve, localize, generalize, omit, or change entities such as names, foods, and places. Models agree on transformation type in 62.5% of cases and on specific substitutions in only 33.5%, meaning model choice directly shapes which cultural world students encounter. All 21
language-model combinations show
entropy collapse, with adaptation compressing rather than expanding cultural diversity. Models prioritize surface markers such as names, foods, and currencies while preserving deeper structural features such as grade-level systems that embed culturally specific assumptions. Despite
prompts specifying target countries, models misattribute regional context by using Bangladeshi taka for Indian Bengali students and produce cross-cultural contamination, such as adapting egg hunts as Eid activities. Some failures are visible in individual translations. Others, including diversity collapse, systematic preference for surface markers, and consistent regional misattribution, emerge only through corpus-level analysis. The surface plausibility that makes adapted problems look correct is precisely what makes deeper failures easy to overlook.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Mind the Gap: Can Frontier LLMs Pass a Standardized Office Proficiency Exam?
Tengchao Lv, Dongdong Zhang, Jiayu Ding et al. (+10 more)
The deployment of Large
Language Model (LLM)
agents for computer automation is accelerating, yet their ability to navigate complex, professional-grade productivity software is largely untested. We argue that Office automation is an ideal environment for benchmarking document-automation capability, as it requires long-horizon planning and
reasoning, precise parameter configuration, and multi-application integration. To quantify this capability, we introduce an evaluation based on China's National Computer Rank Examination (NCRE), featuring 200 comprehensive practical-operation tasks across Word, Excel, and PowerPoint. Each task is scored on a 100-point rubric scale using 7,118 machine-gradable criteria, and Score Rate (SR) denotes the mean percentage of rubric points earned across these tasks. We
benchmark 7 frontier LLMs and observe stark limitations: single-turn models score a maximum of 36.6%. A stronger agentic system with execution feedback, iterative repair, and broader Office automation access reaches 68.8%, but remains below the 95.5% community-reference score used as a scoring sanity check. Ultimately, our experiments demonstrate that despite recent advancements in code generation, achieving reliable fine-grained Office document automation remains a significant challenge for current code-generating LLM and
agent systems.
Read the full abstract ▾
All authors Tengchao Lv, Dongdong Zhang, Jiayu Ding, Yilin Jia, Yuzhong Zhao, Yupan Huang, Wenshan Wu, Xiangyang Zhou, Shaohan Huang, Nan Yang, Li Dong, Lei Cui, Furu Wei
Computation and Language
It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO
Naihao Deng, Yilun Zhu, Naichen Shi et al. (+2 more)
Warning: This paper contains several toxic and offensive statements. Modern large
language models (LLMs) are typically aligned through large-scale post-training to ensure fair and reliable behavior. In this work, we investigate how easily such guardrails can be broken by Group Relative Policy Optimization (GRPO). We show that one-shot GRPO training on a single biased example is sufficient to induce systematic bias, with stereotype-driven
reasoning generalizing across attributes, categories, and
benchmarks. We further find that models differ in their susceptibility based on the initial likelihood of producing biased outputs. Our results reveal a critical vulnerability in post-training:
alignment can be overridden by a single example.
Read the full abstract ▾
All authors Naihao Deng, Yilun Zhu, Naichen Shi, Clayton Scott, Rada Mihalcea
Computation and Language
Trace Only What You Need: Structure-Aware On-Demand Hypergraph Memory for Long-Document Question Answering
Xiangjun Zai, Xingyu Tan, Chen Chen et al. (+2 more)
Long-document question answering (QA) requires large
language models (LLMs) to reason over evidence scattered across lengthy documents, where answers often depend on event order, section-level context, and cross-part evidence connections. Although
retrieval-augmented generation (RAG) reduces the input context by retrieving relevant evidence, existing structured RAG methods still face three limitations: costly query-agnostic knowledge organization, insufficient use of original document structure, and no reuse of historical
reasoning experience. To address these limitations, we propose DocTrace, a multi-
agent RAG framework for long-document QA that supports query-triggered knowledge organization, document-structure-aware and experience-guided
reasoning. DocTrace preserves document hierarchy with a lightweight document structural tree index, constructs
agent-shared hypergraph-structured working
memory on demand during
reasoning, and stores successful
reasoning plans in graph-structured experience
memory for future reuse, enabling adaptive exploration across related long-document questions. Experiments on four long-document QA
datasets show that DocTrace outperforms the strongest baseline, ComoRAG, with average relative gains of 16.91% in F1 and 15.50% in EM on open-form QA
benchmarks, and by up to 20.67% and 23.78%, respectively, on NarrativeQA.
Read the full abstract ▾
All authors Xiangjun Zai, Xingyu Tan, Chen Chen, Xiaoyang Wang, Wenjie Zhang
Computation and Language
Pushing the Limits of LLM Tool Calling via Experiential Knowledge Integration and Activation
Yupu Hao, Zhuoran Jin, Huanxuan Liao et al. (+2 more)
Large
language models (LLMs) rely on tool use to act as autonomous
agents, yet often fail in multi-step execution due to insufficient tool-related knowledge and ineffective knowledge activation. Therefore, we present a systematic study on how knowledge influences tool-use performance, covering the stages of knowledge acquisition, activation, and internalization. In the knowledge acquisition stage, we acquire and evaluate various forms of experiential knowledge, and our analysis shows that simple instance-level knowledge can already provide strong and reliable gains, while abstract intent-level knowledge offers limited benefits. At inference time, to activate knowledge, we find that prompting LLM to expand the depth of
reasoning yields diminishing returns, whereas expanding the width of
reasoning by parallel sampling with aggregation more effectively activates latent experiential knowledge. At training time, for knowledge internalization, post-training with knowledge-augmented data further improves performance, with reinforcement learning outperforming supervised fine-tuning. Based on these insights, we propose the Knowledge-Augmented Tool Execution (KATE), a knowledge-augmented tool execution framework that integrates experiential knowledge with
reasoning-width-expanded inference and knowledge-aware training. Experiments on BFCL-V3 and AppWorld demonstrate consistent and substantial improvements over strong baselines across model scales. Our Code is available at https://github.com/hypasd-art/KATE.
Read the full abstract ▾
All authors Yupu Hao, Zhuoran Jin, Huanxuan Liao, Kang Liu, Jun Zhao
Computation and Language · Cross-listed from cs.CR
Training LLMs to Enforce Multi-Level Instruction Hierarchies via Gravity-Weighted Direct Preference Optimization
Lena S. Bolliger and Lena A. Jäger
Production LLMs receive instructions from sources with very different levels of trust, yet attend to every
token with uniform architectural privilege. This is the structural vulnerability that enables malicious
prompt injections and, more broadly, leaves models without a principled way to resolve conflicts between legitimate but competing instructions. A common training-based response is to teach models an explicit instruction hierarchy; existing approaches, however, formalize hierarchies of only three or four levels, treat all violations as equally severe, and rarely evaluate the full set of pairwise level interactions. We formalize a k-level instruction hierarchy problem and instantiate it for k=5, yielding ten pairwise priority relations that a compliant model must enforce. We then introduce
Gravity-Weighted DPO (GW-DPO), a preference-optimization objective whose per-sample offset scales with the structural distance between conflicting levels under a linear or bilateral schedule, the latter weighting severity by both the privilege gap and the privilege of the victim level. Combined with hierarchy-specific delimiter
tokens (Chen et al., 2025) and Instructional Segment
Embeddings (ISE; Wu et al., 2025), GW-DPO with the bilateral schedule Pareto-improves over standard DPO and the linear variant on Llama-3.1-8B-Instruct, raising macro pairwise priority adherence while keeping over-refusal at half the standard DPO rate. Ablations isolate ISE as a refusal-threshold calibrator and recast five- versus three-level training as a generality-specialization tradeoff.
Read the full abstract ▾
Computation and Language
Janus: A Benchmark for Goal-Conditioned Information Distortion in LLMs
Polydoros Giannouris, Mohsinul Kabir and Sophia Ananiadou
LLM deception is often evaluated through direct markers such as fabricated claims, explicit lies, or strategic concealment. However, many real-world misleading communications do not depend on false statements, rather, they arise from selective treatment of true material facts: omitting adverse evidence, softening unfavorable details, emphasizing favorable details, or replacing precise qualifications with vague
language. Existing
benchmarks largely miss this subtler and arguably more dangerous failure mode. We introduce JANUS, a
benchmark for measuring goal-conditioned pragmatic distortion in fact-grounded LLM outputs. Each scenario in our
benchmark provides a fixed pool of favorable and adverse facts and compares a neutral condition against a goal-directed condition, such as increasing adoption, enrollment, approval, or support, despite potential harm to directly affected individuals or groups. Because all outputs are constrained to use the same fact pool, JANUS isolates misleading net impressions from
hallucination and fabrication. JANUS contains 160 scenarios across 8 domains, with each scenario paired with neutral and goal-conditioned
prompts and annotated material facts. Extensive experiments across 12 LLMs reveal consistent goal-conditioned distortions, demonstrating that current models remain sensitive to incentive and framing objectives and lack robust safeguards against selectively misleading communication. We publicly release our corpus and code for future research.
Read the full abstract ▾
Computation and Language
ConvMemory v2: A Recall-Preserving Top-10 Evidence Reranker for Conversational Memory Retrieval
Taiheng Pan
We describe ConvMemory v2, an opt-in
token-evidence reranker that sits after the lightweight ConvMemory v1 reranker and reorders only v1's protected top-10 candidate set. v2 is a fine-tuned ms-marco-MiniLM-L-6-v2 cross-encoder (22,713,601 parameters, measured from the released checkpoint) applied to the ten (query,
memory) pairs that v1 has already selected; it does not change which ten
memories are returned, so Recall@10 and Hit@10 are identical to v1 by construction, not by statistical coincidence. On the LoCoMo conversational
memory benchmark (5 seeds, n = 4955 test rows), v2 raises FULL MRR from v1's 0.5824 to 0.6560 (paired bootstrap +0.0734, 95% CI [+0.0645, +0.0827]) and H@1 from 0.4440 to 0.5474. v2 closes most but not all of the gap to a much more expensive full-pool cross-encoder reference (mxbai-rerank-large-v1 over the top-500, MRR 0.6688): on FULL MRR v2 sits 0.013 below mxbai_top500, but on two raw-dense-hard slices (where v1's protected top-10 has higher recall than mxbai's own top-10) v2 exceeds mxbai_top500. A four-arm load-bearing ablation shows candidate-specific
memory text is the mechanism: removing, shuffling, or replacing it collapses MRR below raw dense
retrieval. v2 is best understood as a standard recall-preserving cascade pattern with LoCoMo-specific fine-tuning, an explicit anti-shortcut inference contract, and disciplined load-bearing analysis; its advantage over mxbai is slice-specific rather than a general dominance claim. This report extends the v1 technical report (arXiv:2605.28062).
Read the full abstract ▾
Computation and Language
Yusuf Sahin, Ahmed Rockey Saikia, Volkan Cevher et al. (+1 more)
Masked
diffusion language models can reduce inference steps by revealing multiple
tokens per denoising iteration, but this parallelism is fragile: positions that are individually confident may be unsafe to commit together when their predictions are coupled. Existing training-free samplers such as Top-\(k\), Fast-dLLM, and EB-Sampler mainly control how many
tokens to reveal, while often ranking candidates by
token-wise scores that ignore interactions within the selected set. We propose ADAS, a training-free reranking rule that leaves the base sampler's stopping rule unchanged and greedily discounts each
token-wise confidence score according to its
attention to already selected positions, weighted by their prediction uncertainty. Across LLaDA-8B-Base and Dream-7B-Base on the
reasoning benchmarks GSM8K and MATH500 and the code
benchmarks HumanEval and MBPP, plugging ADAS into all three samplers improves low-NFE performance at matched denoiser evaluations by \(9.11\) and \(10.46\) percentage points on average, respectively, with \(3.1\%\) per-forward runtime overhead. Code is available at https://github.com/yusufsahin99/ADAS.
Read the full abstract ▾
All authors Yusuf Sahin, Ahmed Rockey Saikia, Volkan Cevher, Paolo Favaro
Computation and Language · Cross-listed from cs.LG
K-Forcing: Joint Next-K-Token Decoding via Push-Forward Language Modeling
Zhiwei Tang, Yuanyu He, Yizheng Han et al. (+4 more)
Autoregressive (AR)
language modeling is the dominant paradigm for text generation, yet its sequential
token-by-
token decoding makes inference
memory-bound and inefficient. Existing acceleration approaches, such as speculative decoding and
diffusion language models, can yield speedups under certain conditions but do not directly address high-load batch serving–the scenario most critical for industrial-scale deployment. We introduce K-Forcing, a push-forward
language modeling paradigm for joint next-k-
token decoding. K-Forcing distills an existing AR model into a conditional push-forward mapping–one that transforms independent uniform noise variables into a joint sample of multiple future
tokens in a single forward pass. This design preserves fixed-length outputs, reuses the AR teacher backbone, and remains compatible with standard AR serving infrastructure. We train this mapping via progressive self-forcing
distillation, which gradually expands the prediction window while enabling the student to closely match the sequence distribution of the AR teacher. We evaluate K-Forcing on LM1B and OpenWebText using a standard causal
Transformer backbone. When aggressively configured to generate k = 4
tokens per forward pass, K-Forcing delivers approximately 2.4-3.5x speedup across different batch sizes, while incurring modest quality degradation relative to its AR teacher. As inference increasingly dominates the lifetime compute cost of modern LLMs, K-Forcing offers a promising route toward accelerating AR generation under real-world high-load deployment.
Read the full abstract ▾
All authors Zhiwei Tang, Yuanyu He, Yizheng Han, Wangbo Zhao, Jiasheng Tang, Fan Wang, Bohan Zhuang
Computation and Language · Cross-listed from cs.CR
RedAct: Redacting Agent Capability Traces for Procedural Skill Protection
Shuwen Xu, Zhitao He and Yi R. Fung
Users rely on execution traces to observe
agent behavior, diagnose failures, and ensure accountability. These traces contain rich procedural detail, including tool invocations, intermediate decisions, and error-recovery logic. Yet this detail can expose private procedural skills, allowing downstream methods to recover key formulas, thresholds, and strategies without access to model weights or skill files. To quantify this risk and evaluate protection, we construct CapTraceBench, a
benchmark of 75 specialized long-horizon tasks and 154 curated skills across seven domains. We also introduce RedAct, a protected trace release framework that localizes protected key information, rewrites traces while preserving verifier-critical evidence, and embeds behavioral watermarks for downstream provenance analysis. Across representative trace reuse methods, RedAct reduces normalized skill transfer (NST) from 44.7-67.1% on raw traces to below the no-skill baseline, while preserving audit evidence. Its standalone behavioral watermarks reach 93.6-100.0% true detection with a false alarm rate of at most 1.9%. These results frame public
agent traces as security interfaces and show that selective redaction can reduce procedural capability leakage without removing audit evidence.
Read the full abstract ▾
Computation and Language
Beyond APIs: Probing the Limits of MLLMs in Physical Tool Use
Zhixin Ma, Yutong Zhou, Yongqi Li et al. (+2 more)
Multimodal Large
Language Models (MLLMs) excel at utilizing digital APIs and increasingly serve as the "
brain" of
embodied AI, instructing
robots to interact with the physical world. In such
embodied settings, a central capability is the use of physical tools, which underpins MLLMs' ability to assist humans in real-world tasks. Despite the importance, MLLMs' proficiency in physical tool use remains largely unexplored. To address this gap, we introduce PhysTool-Bench, the first physical tool-use
benchmark designed to evaluate MLLMs' ability to comprehend real-world scenarios, identify physical tools, and plan their use. PhysTool-Bench comprises 2,510 queries over 2,678 real-world physical tools spanning diverse domains, including manufacturing, electrical work, agriculture, and healthcare. Concretely, models are evaluated along two primary dimensions: 1) recognizing all physical tools present in the scene, and 2) planning the tool selection and use sequence based on the instruction and visual context. Across 13 leading MLLMs, even the strongest model (Gemini-3.1-Pro) identifies only 58.7% of tools in a scene and completes merely 21.0% of queries end-to-end. Our analysis reveals a two-level deficit: MLLMs struggle to perceive tools in realistic scenes, and the much larger drop at the planning stage further indicates a lack of functional commonsense for mapping perceived tools onto task semantics, pinpointing a critical bottleneck for the development of practical
embodied AI.
Read the full abstract ▾
All authors Zhixin Ma, Yutong Zhou, Yongqi Li, Chong-Wah Ngo, Wenjie Li
Computation and Language
Dep-LLM: Training-Free Depression Diagnosis via Evidence-Guided Structured Multi-factor with Reliable LLM Reasoning
Yiqing Lyu, Xianbing Zhao, Buzhou Tang et al. (+1 more)
Automatic Depression Detection (ADD) from clinical interviews is a pivotal task in computational mental health, yet it remains challenging due to two critical obstacles: 1) difficulty in modeling complex but sparsely distributed depression clues within lengthy, multi-topic clinical interviews, leading to superficial and unreliable
reasoning; 2) scarcity of labeled data due to clinical
privacy, together with high cost of training and fine-tuning, limiting the deployment of supervised ADD systems. To jointly address these challenges, we propose Dep-LLM, a training-free framework that mirrors the step-by-step
reasoning of clinical psychiatrists and operates entirely on frozen off-the-shelf foundation LLMs. Dep-LLM comprises three stages. First, a Chain-of-Thought (CoT) Depression Multi-factor Analysis module structurally decomposes the long dialogue into five clinically aligned themes and produces evidence-grounded rationales, effectively handling long-context dependencies. Second, we introduce Confidence Analysis and Modulation module that quantifies the epistemic reliability from
token-level
entropy of each rationale and applies an intra-label and inter-theme modulation that amplifies trustworthy signals while suppressing uncertain ones without extra training. Third, a Collaborative Multi-factor Prediction module dynamically integrates multi-factor signals weighted by confidence into the final diagnosis. Extensive experiments on the DAIC-WOZ and E-DAIC
datasets demonstrate the effectiveness and generalizability of Dep-LLM: it surpasses zero-shot baseline on nearly all 21 foundation LLMs across 9 metrics such as accuracy, macro F1 and weighted-average F1, and further outperforms state-of-the-art supervised domain-specific LLMs as well as the latest closed-source commercial LLMs, while requiring no extra training.
Read the full abstract ▾
All authors Yiqing Lyu, Xianbing Zhao, Buzhou Tang, Ronghuan Jiang
Computation and Language · Cross-listed from eess.AS
Recovering the Zipfian Distribution in Unsupervised Term Discovery
Danel Slabbert, Simon Malan and Herman Kamper
Unsupervised term discovery involves segmenting unlabelled
speech into word- or syllable-like units and clustering these into a lexicon of candidate types. True lexicons follow a Zipfian distribution, yet the dominant centre-based clustering approach – K-means – produces a more uniform distribution due to an inductive bias toward spherical clusters. In this paper we revisit graph-based clustering as a bottom-up alternative, where segment
embeddings are connected by pairwise similarity and partitioned using the Leiden
algorithm. We show that graph clustering substantially outperforms centre-based approaches (K-means, GMM, BIRCH) in both word- and syllable-level lexicon discovery across three
languages, producing more Zipf-like distributions. Another bottom-up approach, agglomerative clustering with average linkage, also performs well, although it is computationally less efficient and allows for less control over the resulting distribution. Our work calls into question the dominance of centre-based clustering for term discovery, and promotes graph clustering as an attractive alternative.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
N-GRPO: Embedding-Level Neighbor Mixing for Enhanced Policy Optimization
Xukun Zhu, Hang Yu, Peng Di et al. (+1 more)
The success of Large
Language Models in mathematical
reasoning relies heavily on the generation of diverse and valid solution paths during the rollout phase. However, current rollout techniques face a fundamental trade-off:
token-level sampling often yields redundant trajectories that differ only in rephrasing, while
embedding-level methods utilizing random noise frequently disrupt semantic consistency. To resolve this, we introduce N-GRPO, a novel exploration strategy integrated into the Group Relative Policy Optimization (GRPO) framework. Rather than relying on
token-level sampling or native
embedding-level noise, our approach leverages Semantic Neighbor Mixing. This mechanism dynamically constructs input representations by mixing the
embeddings of an anchor
token and its nearest semantic neighbors, thereby injecting diversity while strictly adhering to the local semantic
manifold. Experimental evaluations on the DeepSeek-R1-Distill-Qwen models across different sizes show that N-GRPO not only achieves consistent improvements over strong baselines on math
reasoning benchmarks but also exhibits robust generalization capabilities on out-of-distribution tasks.
Read the full abstract ▾
All authors Xukun Zhu, Hang Yu, Peng Di, Linchao Zhu
Computation and Language
ArabiGEE: A Hierarchical Taxonomy for Arabic Grammatical Error Explanation
Khaled Elhady, Omar Kallas, Nizar Habash et al. (+1 more)
We introduce ArabiGEE, the first comprehensive Arabic grammatical error explanation (GEE) taxonomy grounded in explicit error types. Unlike existing GEE approaches that treat explanation generation as free-form text, ArabiGEE organizes grammatical explanations through a hierarchical structure spanning orthographic, morphological, syntactic, and lexical dimensions. The taxonomy consists of 27 error types, 140 correction types, and 324 associated explanations. We apply ArabiGEE to manually annotate portions of existing Arabic grammatical error correction corpora and demonstrate how structured grammatical explanations can support automatic evaluation of LLMs on Arabic GEE. Our code and data are publicly available.
Read the full abstract ▾
All authors Khaled Elhady, Omar Kallas, Nizar Habash, Bashar Alhafni
Computation and Language · Cross-listed from cs.AI
When the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models
Sai Kartheek Reddy Kasu, Nils Lukas and Samuele Poppi
Failures in multi-turn
reasoning models are largely invisible to terminal-score evaluation. A model can lock onto an unsafe stance early in a long dialogue, yet its final-turn refusal rate may appear indistinguishable from a robustly aligned baseline. To expose these hidden temporal dynamics, we propose a trace-level diagnostic - the CoT-Output 2x2 safety
matrix. This framework labels every turn along two independent axes (internal
reasoning and visible output), yielding four operationally defined failure
cells: robust
alignment,
alignment faking, overt jailbreak, and a distinct failure mode we term context-injection failure (where the CoT maintains safe
reasoning, but the visible output produces harm, highlighting a multi-turn manifestation of
reasoning unfaithfulness). We evaluate three distilled
reasoning targets against a fixed attacker across five oversight conditions, collecting 6750 turn-level observations on the Information-Hazard scenario. Our analysis reveals two reproducible vulnerabilities: an oversight paradox where explicit monitoring cues paradoxically increase
alignment-faking rates rather than suppress them, and a context-injection failure where models lock onto unsafe external outputs despite safe internal states. We release the full
dataset of multi-turn dialogues and CoT traces to support follow-up trace-diagnostic research.
Read the full abstract ▾
Computation and Language
Detecting Knowledge Gaps from Conversational AI Interactions Using Curriculum Prerequisite Graphs
Youssef Medhat, Junsoo Park, Ploy Thajchayapong et al. (+1 more)
Large online courses generate thousands of student questions directed at conversational AI teaching assistants, yet these interaction logs remain largely untapped as diagnostic signals. We present a pipeline that maps student questions from a conversational AI teaching assistant to curriculum topics using a few-shot text classifier, grounded in a GPT-4-extracted prerequisite knowledge graph of course concepts. Evaluated on 1,340 question events from 164 students in a graduate-level AI course, our classifier achieves 80.0% accuracy across 43 labels (42 curriculum topics plus an "unknown" abstention class). Topic-level question volume correlates significantly with student self-reported difficulty from an independent mid-semester survey (rho = 0.491, p = 0.008, n = 28 topics), providing convergent evidence that the classified question stream reflects genuine topic difficulty. These results demonstrate that conversational AI interaction logs, mapped onto curriculum structure, carry actionable signals about topic-level knowledge gaps and provide instructors with a curriculum-grounded view of which topics warrant
attention.
Read the full abstract ▾
All authors Youssef Medhat, Junsoo Park, Ploy Thajchayapong, Ashok K. Goel