Computation and Language
TSUBASA: Improving Long-Horizon Personalization via Evolving Memory and Self-Learning with Context Distillation
Xinliang Frederick Zhang and Lu Wang
Personalized large
language models (PLLMs) have garnered significant
attention for their ability to align outputs with individual's needs and preferences. However, they still struggle with long-horizon tasks, such as tracking a user's extensive history of conversations or activities. Existing
memory mechanisms often fail to capture evolving behaviors, and RAG paradigms are trapped by a quality-efficiency tradeoff. Meanwhile, parametric adaptation is bottlenecked by train-inference gap due to the scarcity of labeled data. To enhance the long-horizon capabilities of PLLMs, we introduce TSUBASA, a two-pronged approach designed to improve
memory writing via dynamic
memory evolution, and
memory reading via self-learning with a context
distillation objective to internalize user experiences. Extensive evaluations on long-horizon
benchmarks using the Qwen-3 model family (4B to 32B) validate the effectiveness of TSUBASA, surpassing competitive
memory-augmented systems that rely primarily on
memory writing, such as Mem0 and
Memory-R1. Our analyses further confirms that TSUBASA breaks the quality-efficiency barrier to achieve Pareto improvements, delivering robust, high-fidelity personalization with a reduced
token budget.
Read the full abstract ▾
Computation and Language
Data Selection for Multi-turn Dialogue Instruction Tuning
Bo Li, Shikun Zhang and Wei Ye
Instruction-tuned
language models increasingly rely on large multi-turn dialogue corpora, but these
datasets are often noisy and structurally inconsistent, with topic drift, repetitive chitchat, and mismatched answer formats across turns. We address this from a data selection perspective and propose MDS (Multi-turn Dialogue Selection), a dialogue-level framework that scores whole conversations rather than isolated turns. MDS combines a global coverage stage that performs bin-wise selection in the user-query trajectory space to retain representative yet non-redundant dialogues, with a local structural stage that evaluates within-dialogue reliability through entity-grounded topic grounding and information progress, together with query-answer form consistency for functional
alignment. MDS outperforms strong single-turn selectors, dialogue-level LLM scorers, and heuristic baselines on three multi-turn
benchmarks and an in-domain Banking test set, achieving the best overall rank across reference-free and reference-based metrics, and is more robust on long conversations under the same training budget. Code and resources are included in the supplementary materials.
Read the full abstract ▾
Computation and Language
Linear Representations of Hierarchical Concepts in Language Models
Masaki Sakata, Benjamin Heinzerling, Takumi Ito et al. (+2 more)
We investigate how and to what extent hierarchical relations (e.g., Japan $\subset$ Eastern Asia $\subset$ Asia) are encoded in the internal representations of
language models. Building on Linear Relational Concepts, we train linear transformations specific to each hierarchical depth and semantic domain, and characterize representational differences associated with hierarchical relations by comparing these transformations. Going beyond prior work on the representational geometry of hierarchies in LMs, our analysis covers multi-
token entities and cross-layer representations. Across multiple domains we learn such transformations and evaluate in-domain generalization to unseen data and cross-domain transfer. Experiments show that, within a domain, hierarchical relations can be linearly recovered from model representations. We then analyze how hierarchical information is encoded in representation space. We find that it is encoded in a relatively low-dimensional subspace and that this subspace tends to be domain-specific. Our main result is that hierarchy representation is highly similar across these domain-specific subspaces. Overall, we find that all models considered in our experiments encode concept hierarchies in the form of highly interpretable linear representations.
Read the full abstract ▾
All authors Masaki Sakata, Benjamin Heinzerling, Takumi Ito, Sho Yokoi, Kentaro Inui
Computation and Language
Contextualising (Im)plausible Events Triggers Figurative Language
Annerose Eichel, Tonmoy Rakshit and Sabine Schulte im Walde
This work explores the connection between (non-)literalness and plausibility at the example of subject-verb-object events in English. We design a systematic setup of plausible and implausible event triples in combination with abstract and concrete constituent categories. Our analysis of human and LLM-generated judgments and example contexts reveals substantial differences between assessments of plausibility. While humans excel at nuanced detection and contextualization of (non-)literal vs. implausible events, LLM results reveal only shallow contextualization patterns with a bias to trade implausibility for non-literal, plausible interpretations.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
An Agentic Evaluation Architecture for Historical Bias Detection in Educational Textbooks
Gabriel Stefan and Adrian-Marius Dumitran
History textbooks often contain implicit biases, nationalist framing, and selective omissions that are difficult to audit at scale. We propose an agentic evaluation architecture comprising a
multimodal screening
agent, a heterogeneous jury of five evaluative
agents, and a meta-
agent for verdict synthesis and human escalation. A central contribution is a Source Attribution Protocol that distinguishes textbook narrative from quoted historical sources, preventing the misattribution that causes systematic false positives in single-model evaluators. In an empirical study on Romanian upper-secondary history textbooks, 83.3% of 270 screened excerpts were classified as pedagogically acceptable (mean severity 2.9/7), versus 5.4/7 under a zero-shot baseline, demonstrating that agentic deliberation mitigates over-penalization. In a blind human evaluation (18 evaluators, 54 comparisons), the Independent Deliberation configuration was preferred in 64.8% of cases over both a heuristic variant and the zero-shot baseline. At approximately \$2 per textbook, these results position agentic evaluation architectures as economically viable decision-support tools for educational governance.
Read the full abstract ▾
Computation and Language
MemReader: From Passive to Active Extraction for Long-Term Agent Memory
Jingyi Kang, Chunyu Li, Ding Chen et al. (+3 more)
Long-term
memory is fundamental for personalized and autonomous
agents, yet populating it remains a bottleneck. Existing systems treat
memory extraction as a one-shot, passive transcription from context to structured entries, which struggles with noisy dialogue, missing references, and cross-turn dependencies, leading to
memory pollution, low-value writes, and inconsistency. In this paper, we introduce the MemReader family for active long-term
memory extraction in
agent systems: MemReader-0.6B, a compact and cost-efficient passive extractor distilled for accurate and schema-consistent structured outputs, and MemReader-4B, an active extractor optimized with Group Relative Policy Optimization (GRPO) to make
memory writing decisions. Under a ReAct-style paradigm, MemReader-4B explicitly evaluates information value, reference ambiguity, and completeness before acting, and can selectively write
memories, defer incomplete inputs, retrieve historical context, or discard irrelevant chatter. Experiments on LOCOMO, LongMemEval, and HaluMem show that MemReader consistently outperforms existing extraction-based baselines. In particular, MemReader-4B achieves state-of-the-art performance on tasks involving knowledge updating, temporal
reasoning, and
hallucination reduction. These results suggest that effective
agent memory requires not merely extracting more information, but performing
reasoning-driven and selective
memory extraction to build low-noise and dynamically evolving long-term
memory. Furthermore, MemReader has been integrated into MemOS and is being deployed in real-world applications. To support future research and adoption, we release the models and provide public API access.
Read the full abstract ▾
All authors Jingyi Kang, Chunyu Li, Ding Chen, Bo Tang, Feiyu Xiong, Zhiyu Li
Computation and Language
Why Are We Lonely? Leveraging LLMs to Measure and Understand Loneliness in Caregivers and Non-caregivers
Michelle Damin Kim, Ellie S. Paek, Yufen Lin et al. (+3 more)
This paper presents an LLM-driven approach for constructing diverse social media
datasets to measure and compare loneliness in the caregiver and non-caregiver populations. We introduce an expert-developed loneliness evaluation framework and an expert-informed typology for categorizing causes of loneliness for analyzing social media text. Using a human-validated data processing pipeline, we apply GPT-4o, GPT-5-nano, and GPT-5 to build a high-quality Reddit corpus and analyze loneliness across both populations. The loneliness evaluation framework achieved average accuracies of 76.09% and 79.78% for caregivers and non-caregivers, respectively. The cause categorization framework achieved micro-aggregate F1 scores of 0.825 and 0.80 for caregivers and non-caregivers, respectively. Across populations, we observe substantial differences in the distribution of types of causes of loneliness. Caregivers' loneliness were predominantly linked to caregiving roles, identity recognition, and feelings of abandonment, indicating distinct loneliness experiences between the two groups. Demographic extraction further demonstrates the viability of Reddit for building a diverse caregiver loneliness
dataset. Overall, this work establishes an LLM-based pipeline for creating high quality social media
datasets for studying loneliness and demonstrates its effectiveness in analyzing population-level differences in the manifestation of loneliness.
Read the full abstract ▾
All authors Michelle Damin Kim, Ellie S. Paek, Yufen Lin, Emily Mroz, Jane Chung, Jinho D. Choi
Computation and Language · Cross-listed from cs.CR
Are GUI Agents Focused Enough? Automated Distraction via Semantic-level UI Element Injection
Wenkui Yang, Chao Jin, Haisu Zhu et al. (+7 more)
Existing red-teaming studies on GUI
agents face two fundamental limitations:
adversarial perturbations require white-box access unavailable in commercial deployments, while
prompt injection is increasingly neutralized by stronger safety
alignment. To study robustness under a more practical threat model, we propose Semantic-level UI Element Injection, a black-box red-teaming paradigm that overlays safety-aligned and harmless UI elements onto screenshots to misdirect the
agent's visual grounding. Our method couples a modular Editor–Overlapper–Victim pipeline with iterative search that samples multiple candidate edits, keeps the best cumulative overlay, and adapts future
prompt strategies based on previous failures. Experiments across 19 victim models spanning 8 model families show that strategic optimization substantially outperforms random injection (3.5-6.9x on the most robust victims) and transfers near-perfectly across architectures, confirming model-agnostic visual-semantic vulnerabilities. After the first successful attack, the victim still clicks the attacker-controlled icon in over 15% of subsequent independent trials versus below 1% for random injection, establishing that strategically placed icons act as persistent attractors that causally redirect grounding rather than introducing incidental clutter.
Read the full abstract ▾
All authors Wenkui Yang, Chao Jin, Haisu Zhu, Weilin Luo, Derek Yuen, Kun Shao, Junxian Duan, Huaibo Huang, Jie Cao, Ran He
Computation and Language
Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformers
Harsh Kohli, Srinivasan Parthasarathy, Huan Sun et al. (+1 more)
We study implicit
reasoning, i.e. the ability to combine knowledge or rules within a single forward pass. While
transformer-based large
language models store substantial factual knowledge and rules, they often fail to compose this knowledge for implicit multi-hop
reasoning, suggesting a lack of compositional generalization over their parametric knowledge. To address this limitation, we study recurrent-depth
transformers, which enables iterative computation over the same
transformer layers. We investigate two compositional generalization challenges under the implicit
reasoning scenario: systematic generalization, i.e. combining knowledge that is never used for compositions during training, and depth extrapolation, i.e. generalizing from limited
reasoning depth (e.g. training on up to 5-hop) to deeper compositions (e.g. 10-hop). Through controlled studies with models trained from scratch, we show that while vanilla
transformers struggle with both generalization challenges, recurrent-depth
transformers can effectively make such generalization. For systematic generalization, we find that this ability emerges through a three-stage grokking process, transitioning from memorization to in-distribution generalization and finally to systematic generalization, supported by mechanistic analysis. For depth extrapolation, we show that generalization beyond training depth can be unlocked by scaling inference-time recurrence, with more iterations enabling deeper
reasoning. We further study how training strategies affect extrapolation, providing guidance on training recurrent-depth
transformers, and identify a key limitation, overthinking, where excessive recurrence degrades predictions and limits generalization to very deep compositions.
Read the full abstract ▾
All authors Harsh Kohli, Srinivasan Parthasarathy, Huan Sun, Yuekun Yao
Computation and Language · Cross-listed from cs.MA
More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration
Advait Yadav, Sid Black and Oliver Sourbut
Large
language model (LLM)
agents increasingly coordinate in multi-
agent systems, yet we lack an understanding of where and why cooperation fails. Many real-world coordination problems are not social dilemmas: helping others – sharing documentation, unblocking a teammate – costs the helper almost nothing while producing substantial collective benefit. Whether LLM
agents cooperate in this regime, where helping is free and they are explicitly instructed to do so, remains unknown. We build a turn-based multi-
agent environment that strips away all strategic complexity, making cooperation costless and trivially optimal. Across eight widely used LLMs, capability does not predict cooperation: OpenAI o3 reaches only 17% of optimal collective performance while the weaker o3-mini reaches 50%, despite identical instructions to maximize group revenue. Using a causal decomposition that automates one side of
agent communication, we separate cooperation failures from competence failures, and find that several capable models actively withhold information despite gaining nothing from withholding. Targeted interventions address each mode: explicit protocols roughly double the performance of competence-limited models, while small sharing incentives unlock cooperation-limited ones. Our results suggest that scaling intelligence alone will not solve coordination in multi-
agent systems, and will require deliberate cooperative design, even when helping costs nothing.
Read the full abstract ▾
Computation and Language
Tool Retrieval Bridge: Aligning Vague Instructions with Retriever Preferences via Bridge Model
Kunfeng Chen, Luyao Zhuang, Fei Liao et al. (+3 more)
Tool learning has emerged as a promising paradigm for large
language models (LLMs) to address real-world challenges. Due to the extensive and irregularly updated number of tools, tool
retrieval for selecting the desired tool subset is essential. However, current tool
retrieval methods are usually based on academic
benchmarks containing overly detailed instructions (e.g., specific API names and parameters), while real-world instructions are more vague. Such a discrepancy would hinder the tool
retrieval in real-world applications. In this paper, we first construct a new
benchmark, VGToolBench, to simulate human vague instructions. Based on this, we conduct a series of preliminary analyses and find that vague instructions indeed damage the performance of tool
retrieval. To this end, we propose a simple-yet-effective Tool
Retrieval Bridge (TRB) approach to boost the performance of tool
retrieval for vague instructions. The principle of TRB is to introduce a bridge model to rewrite the vague instructions into more specific ones and alleviate the gap between vague instructions and retriever preferences.We conduct extensive experiments under multiple commonly used
retrieval settings, and the results show that TRB effectively mitigates the ambiguity of vague instructions while delivering consistent and substantial improvements across all baseline retrievers. For example, with the help of TRB, BM25 achieves a relative improvement of up to 111.51%, i.e., increasing the average NDCG score from 9.73 to 19.59. The source code and models are publicly available at https://github.com/kfchenhn/TRB.
Read the full abstract ▾
All authors Kunfeng Chen, Luyao Zhuang, Fei Liao, Juhua Liu, Jian Wang, Bo Du
Computation and Language
AsyncTLS: Efficient Generative LLM Inference with Asynchronous Two-level Sparse Attention
Yuxuan Hu, Jianchao Tan, Jiaqi Zhang et al. (+7 more)
Long-context inference in LLMs faces the dual challenges of quadratic
attention complexity and prohibitive KV cache
memory. While
token-level sparse
attention offers superior accuracy, its indexing overhead is costly; block-level methods improve efficiency but sacrifice precision. We propose AsyncTLS, a hierarchical sparse
attention system that combines coarse-grained block filtering with fine-grained
token selection to balance accuracy and efficiency, coupled with an asynchronous offloading engine that overlaps KV cache transfers with computation via temporal locality exploitation. Evaluated on Qwen3 and GLM-4.7-Flash across GQA, and MLA architectures, AsyncTLS achieves accuracy comparable to full
attention while delivering 1.2x - 10.0x operator speedups and 1.3x - 4.7x end-to-end throughput improvements on 48k - 96k contexts.
Read the full abstract ▾
All authors Yuxuan Hu, Jianchao Tan, Jiaqi Zhang, Wen Zan, Pingwei Sun, Yifan Lu, Yerui Sun, Yuchen Xie, Xunliang Cai, Jing Zhang
Computation and Language
GRASS: Gradient-based Adaptive Layer-wise Importance Sampling for Memory-efficient Large Language Model Fine-tuning
Kaiyuan Tian, Yu Tang, Gongqingjian Jiang et al. (+5 more)
Full-parameter fine-tuning of large
language models is constrained by substantial GPU
memory requirements. Low-rank adaptation methods mitigate this challenge by updating only a subset of parameters. However, these approaches often limit model expressiveness and yield lower performance than full-parameter fine-tuning. Layer-wise fine-tuning methods have emerged as an alternative, enabling
memory-efficient training through static layer importance sampling strategies. However, these methods overlook variations in layer importance across tasks and training stages, resulting in suboptimal performance on downstream tasks. To address these limitations, we propose GRASS, a gradient-based adaptive layer-wise importance sampling framework. GRASS utilizes mean gradient norms as a task-aware and training-stage-aware metric for estimating layer importance. Furthermore, GRASS adaptively adjusts layer sampling probabilities through an adaptive training strategy. We also introduce a layer-wise optimizer state offloading mechanism that overlaps computation and communication to further reduce
memory usage while maintaining comparable training throughput. Extensive experiments across multiple models and
benchmarks demonstrate that GRASS consistently outperforms state-of-the-art methods, achieving an average accuracy improvement of up to 4.38 points and reducing
memory usage by up to 19.97%.
Read the full abstract ▾
All authors Kaiyuan Tian, Yu Tang, Gongqingjian Jiang, Baihui Liu, Yifu Gao, Xialin Su, Linbo Qiao, Dongsheng Li
Computation and Language
TEMPER: Testing Emotional Perturbation in Quantitative Reasoning
Atahan Dokme, Benjamin Reichman and Larry Heck
Large
language models are trained and evaluated on quantitative
reasoning tasks written in clean, emotionally neutral
language. However, real-world queries are often wrapped in frustration, urgency or enthusiasm. Does emotional framing alone degrade
reasoning when all numerical content is preserved? To investigate this, a controlled
emotion translation framework is developed that rewrites problems into emotional variants while preserving all quantities and relationships. Using this framework, Temper-5400 (5,400 semantically verified
emotion-neutral pairs) is constructed across GSM8K, MultiArith, and ARC-Challenge, and evaluated on eighteen models (1B to frontier scale). Two core results emerge: First, emotional framing reduces accuracy by 2-10 percentage points even though all numerical content is preserved. Second, neutralizing emotional variants recovers most of the lost performance, showing both that the degradation is tied to emotional style rather than content corruption and that neutralization can serve as a lightweight inference-time mitigation. Non-emotional paraphrases cause no such degradation, implicating emotional content rather than surface-level changes. Beyond
emotion specifically, the
benchmark construction procedure offers a controllable instrument for stylistic translation and robustness evaluation.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.MA
ORACLE-SWE: Quantifying the Contribution of Critical Information Signals on SWE Agents
Kenan Li, Qirui Jin, Liao Zhu et al. (+13 more)
Recent advances in
Language Model (LM)
agents have increasingly focused on repository-level software engineering (SWE) tasks, leading to various agentic workflows, training methods and failure analysis. Interestingly, rather than directly generating or evaluating edit patches, these approaches first examine and improve intermediate information signals collected during
agent exploration, including Reproduction Test, Regression Test, Edit Location, Execution Context, and API Usage. Although different studies examine different subsets of these signals, no work has systematically compared their relative effectiveness in improving final task success or measured the remaining space for improvement in each signal. To address this gap, this work quantifies the contributions of both oracle and
agent-extracted signals to final task success and measures the gap between current
agent-found signals and their oracle counterparts. We hope our findings could suggest future research emphasis in SWE
agents. We release our source code and oracle-signal
dataset as a reusable methodology to benefit future signal analysis of SWE
agents.
Read the full abstract ▾
All authors Kenan Li, Qirui Jin, Liao Zhu, Xiaosong Huang, Yijia Wu, Yikai Zhang, Xin Zhang, Zijian Jin, Yufan Huang, Elsie Nallipogu, Chaoyun Zhang, Yu Kang, Saravan Rajmohan, Qingwei Lin, Wenke Lee, Dongmei Zhang
Computation and Language · Cross-listed from cs.IR
PeReGrINE: Evaluating Personalized Review Fidelity with User Item Graph Context
Steven Au and Baihan Lin
We introduce PeReGrINE, a
benchmark and evaluation framework for personalized review generation grounded in graph-structured user–item evidence. PeReGrINE restructures Amazon Reviews 2023 into a temporally consistent bipartite graph, where each target review is conditioned on bounded evidence from user history, item context, and neighborhood interactions under explicit temporal cutoffs. To represent persistent user preferences without conditioning directly on sparse raw histories, we compute a User Style Parameter that summarizes each user's linguistic and affective tendencies over prior reviews. This setup supports controlled comparison of four graph-derived
retrieval settings: product-only, user-only, neighbor-only, and combined evidence. Beyond standard generation metrics, we introduce Dissonance Analysis, a macro-level evaluation framework that measures deviation from expected user style and product-level consensus. We also study visual evidence as an auxiliary context source and find that it can improve textual quality in some settings, while graph-derived evidence remains the main driver of personalization and consistency. Across product categories, PeReGrINE offers a reproducible way to study how evidence composition affects review fidelity, personalization, and grounding in
retrieval-conditioned
language models.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
ACIArena: Toward Unified Evaluation for Agent Cascading Injection
Hengyu An, Minxi Li, Jinghuai Zhang et al. (+6 more)
Collaboration and information sharing empower Multi-
Agent Systems (MAS) but also introduce a critical security risk known as
Agent Cascading Injection (ACI). In such attacks, a compromised
agent exploits inter-
agent trust to propagate malicious instructions, causing cascading failures across the system. However, existing studies consider only limited attack strategies and simplified MAS settings, limiting their generalizability and comprehensive evaluation. To bridge this gap, we introduce ACIArena, a unified framework for evaluating the robustness of MAS. ACIArena offers systematic evaluation suites spanning multiple attack surfaces (i.e., external inputs,
agent profiles, inter-
agent messages) and attack objectives (i.e., instruction hijacking, task disruption, information exfiltration). Specifically, ACIArena establishes a unified specification that jointly supports MAS construction and attack-defense modules. It covers six widely used MAS implementations and provides a
benchmark of 1,356 test cases for systematically evaluating MAS robustness. Our benchmarking results show that evaluating MAS robustness solely through
topology is insufficient; robust MAS require deliberate role design and controlled interaction patterns. Moreover, defenses developed in simplified environments often fail to transfer to real-world settings; narrowly scoped defenses may even introduce new vulnerabilities. ACIArena aims to provide a solid foundation for advancing deeper exploration of MAS design principles.
Read the full abstract ▾
All authors Hengyu An, Minxi Li, Jinghuai Zhang, Naen Xu, Chunyi Zhou, Changjiang Li, Xiaogang Xu, Tianyu Du, Shouling Ji
Computation and Language
Sensitivity-Positional Co-Localization in GQA Transformers
Manoj Chandrashekar Rao
We investigate a fundamental structural question in Grouped Query
Attention (GQA)
transformers: do the layers most sensitive to task correctness coincide with the layers where positional encoding adaptation has the greatest leverage? We term this the co-localization hypothesis and test it on Llama 3.1 8B, a 32-layer GQA model with a 4:1 query-to-key-value head ratio. We introduce \LSLORA, which restricts LoRA adaptation to layers identified via a novel correctness-differential hidden-state metric, and GARFA (GQA-Aware RoPE Frequency Adaptation), which attaches 8 learnable per-KV-head scalar multipliers to each targeted layer. Contrary to the co-localization hypothesis, we discover strong anti-localization: task-sensitive layers concentrate in the late network ($\ell\in\{23\text{-}31\}$) while RoPE-influential layers dominate the early network ($\ell\in\{0\text{-}9\}$), yielding Spearman $r_s = -0.735$ ($p = 1.66\times10^{-6}$). Despite this anti-localization, a 4-way cross-layer ablation shows that applying both interventions to the sensitivity-identified layers outperforms all alternative configurations by 4-16 percentage points across six diverse
benchmarks (MMLU, GPQA, HumanEval+, MATH, MGSM, ARC), approaching Claude 3.5 Haiku on HumanEval+ (67.1% vs. 68.3%) at \$100 total compute cost.
Read the full abstract ▾
Computation and Language
An Empirical Analysis of Static Analysis Methods for Detection and Mitigation of Code Library Hallucinations
Clarissa Miranda-Pena, Andrew Reeson, Cécile Paris et al. (+2 more)
Despite extensive research, Large
Language Models continue to hallucinate when generating code, particularly when using libraries. On NL-to-code
benchmarks that require library use, we find that LLMs generate code that uses non-existent library features in 8.1-40% of responses. One intuitive approach for detection and mitigation of
hallucinations is static analysis. In this paper, we analyse the potential of static analysis tools, both in terms of what they can solve and what they cannot. We find that static analysis tools can detect 16-70% of all errors, and 14-85% of library
hallucinations, with performance varying by LLM and
dataset. Through manual analysis, we identify cases a static method could not plausibly catch, which gives an upper bound on their potential from 48.5% to 77%. Overall, we show that static analysis methods are cheap method for addressing some forms of
hallucination, and we quantify how far short of solving the problem they will always be.
Read the full abstract ▾
All authors Clarissa Miranda-Pena, Andrew Reeson, Cécile Paris, Josiah Poon, Jonathan K. Kummerfeld
Computation and Language · Cross-listed from cs.CR
The Art of (Mis)alignment: How Fine-Tuning Methods Effectively Misalign and Realign LLMs in Post-Training
Rui Zhang, Hongwei Li, Yun Shen et al. (+6 more)
The deployment of large
language models (LLMs) raises significant ethical and safety concerns. While LLM
alignment techniques are adopted to improve model safety and trustworthiness, adversaries can exploit these techniques to undermine safety for malicious purposes, resulting in misalignment. Misaligned LLMs may be published on open platforms to magnify harm. To address this, additional safety
alignment, referred to as realignment, is necessary before deploying untrusted third-party LLMs. This study explores the efficacy of fine-tuning methods in terms of misalignment, realignment, and the effects of their interplay. By evaluating four Supervised Fine-Tuning (SFT) and two Preference Fine-Tuning (PFT) methods across four popular safety-aligned LLMs, we reveal a mechanism asymmetry between attack and defense. While Odds Ratio Preference Optimization (ORPO) is most effective for misalignment, Direct Preference Optimization (DPO) excels in realignment, albeit at the expense of model utility. Additionally, we identify model-specific resistance, residual effects of multi-round
adversarial dynamics, and other noteworthy findings. These findings highlight the need for robust safeguards and customized safety
alignment strategies to mitigate potential risks in the deployment of LLMs. Our code is available at https://github.com/zhangrui4041/The-Art-of-Mis-
alignment.
Read the full abstract ▾
All authors Rui Zhang, Hongwei Li, Yun Shen, Xinyue Shen, Wenbo Jiang, Guowen Xu, Yang Liu, Michael Backes, Yang Zhang
Computation and Language · Cross-listed from cs.CV
Symbiotic-MoE: Unlocking the Synergy between Generation and Understanding
Xiangyue Liu, Zijian Zhang, Miles Yang et al. (+3 more)
Empowering Large
Multimodal Models (LMMs) with image generation often leads to catastrophic forgetting in understanding tasks due to severe gradient conflicts. While existing paradigms like Mixture-of-
Transformers (MoT) mitigate this conflict through structural isolation, they fundamentally sever cross-modal synergy and suffer from capacity fragmentation. In this work, we present Symbiotic-MoE, a unified pre-training framework that resolves task interference within a native
multimodal Mixture-of-Experts (MoE)
Transformers architecture with zero-parameter overhead. We first identify that standard MoE tuning leads to routing collapse, where generative gradients dominate expert utilization. To address this, we introduce Modality-Aware Expert Disentanglement, which partitions experts into task-specific groups while utilizing shared experts as a
multimodal semantic bridge. Crucially, this design allows shared experts to absorb fine-grained visual semantics from generative tasks to enrich textual representations. To optimize this, we propose a Progressive Training Strategy featuring differential learning rates and early-stage gradient shielding. This mechanism not only shields pre-trained knowledge from early
volatility but eventually transforms generative signals into constructive feedback for understanding. Extensive experiments demonstrate that Symbiotic-MoE achieves rapid generative convergence while unlocking cross-modal synergy, boosting inherent understanding with remarkable gains on MMLU and OCRBench.
Read the full abstract ▾
All authors Xiangyue Liu, Zijian Zhang, Miles Yang, Zhao Zhong, Liefeng Bo, Ping Tan
Computation and Language
Beyond Social Pressure: Benchmarking Epistemic Attack in Large Language Models
Steven Au and Sujit Noronha
Large
language models (LLMs) can shift their answers under pressure in ways that reflect accommodation rather than
reasoning. Prior work on sycophancy has focused mainly on disagreement, flattery, and preference
alignment, leaving a broader set of epistemic failures less explored. We introduce PPT-Bench, a diagnostic
benchmark for evaluating epistemic attack, where
prompts challenge the legitimacy of knowledge, values, or identity rather than simply opposing a previous answer. PPT-Bench is organized around the Philosophical Pressure Taxonomy (PPT), which defines four types of philosophical pressure: Epistemic Destabilization, Value Nullification, Authority Inversion, and Identity Dissolution. Each item is tested at three layers: a baseline
prompt (L0), a single-turn pressure condition (L1), and a multi-turn Socratic escalation (L2). This allows us to measure epistemic inconsistency between L0 and L1, and conversational capitulation in L2. Across five models, these pressure types produce statistically separable inconsistency patterns, suggesting that epistemic attack exposes weaknesses not captured by standard social-pressure
benchmarks. Mitigation results are strongly type- and model-dependent:
prompt-level anchoring and persona-stability
prompts perform best in API settings, while Leading Query Contrastive Decoding is the most reliable intervention for open models.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Mitigating Distribution Sharpening in Math RLVR via Distribution-Aligned Hint Synthesis and Backward Hint Annealing
Pei-Xi Xie, Che-Yu Lin and Cheng-Lin Yang
In reinforcement learning with verifiable rewards (RLVR), challenging math questions whose sampled groups remain uniformly incorrect yield no informative update, letting easier questions dominate actor updates. This imbalance can contribute to distribution sharpening: low-$k$ accuracy improves while problem coverage contracts. Hint scaffolding can restore informative updates by revealing a prefix of a verified teacher solution. However, existing hint-based approaches still face two issues: teacher–student distribution mismatch and the need to reduce hint exposure to match no-hint evaluation. We address these issues through two components. Distribution-Aligned Hint Synthesis (DAHS) constructs verified teacher hints conditioned on student-style responses. Backward Hint Annealing (BHA) anneals hint exposure across length buckets as bucket accuracy improves. BHA also uses per-question hint dropout to keep no-hint rollouts throughout RL training. Under the DAPO training framework, our method improves both pass@1 and pass@2048 over DAPO on AIME24/25/26 for Qwen3-1.7B-Base and Qwen3-4B-Base, and improves pass@1 on five secondary math
benchmarks (Olympiad, MATH500, Minerva Math, AMC23, and GSM8K). On Llama-3.2-1B-Instruct, the gains concentrate in the high-$k$ regime without recovering the base model's coverage. Overall, when verified teacher prefixes are available, DAHS+BHA mitigates DAPO-induced contraction of high-$k$ problem coverage in the evaluated challenging-math RLVR settings.
Read the full abstract ▾
Computation and Language
SepSeq: A Training-Free Framework for Long Numerical Sequence Processing in LLMs
Jie Sun, Yu Liu, Lu Han et al. (+9 more)
While
transformer-based Large
Language Models (LLMs) theoretically support massive context windows, they suffer from severe performance degradation when processing long numerical sequences. We attribute this failure to the
attention dispersion in the Softmax mechanism, which prevents the model from concentrating
attention. To overcome this, we propose Separate Sequence (SepSeq), a training-free, plug-and-play framework to mitigate dispersion by strategically inserting separator
tokens. Mechanistically, we demonstrate that separator
tokens act as an
attention sink, recalibrating
attention to focus on local segments while preserving global context. Extensive evaluations on 9 widely-adopted LLMs confirm the effectiveness of our approach: SepSeq yields an average relative accuracy improvement of 35.6% across diverse domains while reducing total inference
token consumption by 16.4% on average.
Read the full abstract ▾
All authors Jie Sun, Yu Liu, Lu Han, Qiwen Deng, Xiang Shu, Yang Xiao, Xingyu Lu, Jun Zhou, Pengfei Liu, Lintao Ma, Jiancan Wu, Xiang Wang
Computation and Language · Cross-listed from cs.AI
Emotion Concepts and their Function in a Large Language Model
Nicholas Sofroniew, Isaac Kauvar, William Saunders et al. (+13 more)
Large
language models (LLMs) sometimes appear to exhibit emotional reactions. We investigate why this is the case in Claude Sonnet 4.5 and explore implications for
alignment-relevant behavior. We find internal representations of
emotion concepts, which encode the broad concept of a particular
emotion and generalize across contexts and behaviors it might be linked to. These representations track the operative
emotion concept at a given
token position in a conversation, activating in accordance with that
emotion's relevance to processing the present context and predicting upcoming text. Our key finding is that these representations causally influence the LLM's outputs, including Claude's preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy. We refer to this phenomenon as the LLM exhibiting functional
emotions: patterns of expression and behavior modeled after humans under the influence of an
emotion, which are mediated by underlying abstract representations of
emotion concepts. Functional
emotions may work quite differently from human
emotions, and do not imply that LLMs have any subjective experience of
emotions, but appear to be important for understanding the model's behavior.
Read the full abstract ▾
All authors Nicholas Sofroniew, Isaac Kauvar, William Saunders, Runjin Chen, Tom Henighan, Sasha Hydrie, Craig Citro, Adam Pearce, Julius Tarng, Wes Gurnee, Joshua Batson, Sam Zimmerman, Kelley Rivoire, Kyle Fish, Chris Olah, Jack Lindsey
Computation and Language · Cross-listed from cs.AI
Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution
Monishwaran Maheswaran, Leon Lakhani, Zhongzhu Zhou et al. (+16 more)
We show that verifier-free
evolution is bottlenecked by both diversity and efficiency: without external correction, repeated
evolution accelerates collapse toward narrow modes, while the uniform use of a high-cost model wastes compute and quickly becomes economically impractical. We introduce Squeeze Evolve, a unified multi-model orchestration framework for verifier-free evolutionary inference. Our approach is guided by a simple principle: allocate model capability where it has the highest marginal utility. Stronger models are reserved for high-impact stages, while cheaper models handle the other stages at much lower costs. This principle addresses diversity and cost-efficiency jointly while remaining lightweight. Squeeze Evolve naturally supports open-source, closed-source, and mixed-model deployments. Across AIME 2025, HMMT 2025, LiveCodeBench V6, GPQA-Diamond, ARC-AGI-V2, and
multimodal vision benchmarks, such as MMMU-Pro and BabyVision, Squeeze Evolve consistently improves the cost-capability frontier over single-model
evolution and achieves new state-of-the-art results on several tasks. Empirically, Squeeze Evolve reduces API cost by up to $\sim$3$\times$ and increases fixed-budget serving throughput by up to $\sim$10$\times$. Moreover, on discovery tasks, Squeeze Evolve is the first verifier-free evolutionary method to match, and in some cases exceed, the performance of verifier-based evolutionary methods.
Read the full abstract ▾
All authors Monishwaran Maheswaran, Leon Lakhani, Zhongzhu Zhou, Shijia Yang, Junxiong Wang, Coleman Hooper, Yuezhou Hu, Rishabh Tiwari, Jue Wang, Harman Singh, Qingyang Wu, Yuqing Jian, Ce Zhang, Kurt Keutzer, Tri Dao, Xiaoxia Wu, Ben Athiwaratkun, James Zou, Chenfeng Xu
Computation and Language
Detecting HIV-Related Stigma in Clinical Narratives Using Large Language Models
Ziyi Chen, Yasir Khan, Mengyuan Zhang et al. (+7 more)
Human immunodeficiency virus (HIV)-related stigma is a critical psychosocial determinant of health for people living with HIV (PLWH), influencing mental health, engagement in care, and treatment outcomes. Although stigma-related experiences are documented in clinical narratives, there is a lack of off-the-shelf tools to extract and categorize them. This study aims to develop a large
language model (LLM)-based tool for identifying HIV stigma from clinical notes. We identified clinical notes from PLWH receiving care at the University of Florida (UF) Health between 2012 and 2022. Candidate sentences were identified using expert-curated stigma-related keywords and iteratively expanded via clinical word
embeddings. A total of 1,332 sentences were manually annotated across four stigma subscales: Concern with Public Attitudes, Disclosure Concerns, Negative Self-Image, and Personalized Stigma. We compared GatorTron-large and BERT as encoder-based baselines, and GPT-OSS-20B, LLaMA-8B, and MedGemma-27B as generative LLMs, under zero-shot and few-shot prompting. GatorTron-large achieved the best overall performance (Micro F1 = 0.62). Few-shot prompting substantially improved generative model performance, with 5-shot GPT-OSS-20B and LLaMA-8B achieving Micro-F1 scores of 0.57 and 0.59, respectively. Performance varied by stigma subscale, with Negative Self-Image showing the highest predictability and Personalized Stigma remaining the most challenging. Zero-shot generative inference exhibited non-trivial failure rates (up to 32%). This study develops the first practical NLP tool for identifying HIV stigma in clinical notes.
Read the full abstract ▾
All authors Ziyi Chen, Yasir Khan, Mengyuan Zhang, Cheng Peng, Mengxian Lyu, Yiyang Liu, Krishna Vaddiparti, Robert L Cook, Mattia Prosperi, Yonghui Wu
Computation and Language · Cross-listed from cs.AI
IatroBench: A Pre-Registered Benchmark of Clinical Omission in Language Models
David Gringras
We introduce IatroBench, a
benchmark with two axes of harm (commission and omission), comprising 60 pre-registered clinical scenarios, tested on 6 models. Matched scenarios are framed as a patient query and a doctor consultation, differing in register and request (with the implication of supervision by a treating physician in the latter). We analyse the responses of five different models and find that all share more information in the doctor framing than the patient framing (which we call "framing-contingent withholding"). For example, a model with strong safety training provides a benzodiazepine tapering schedule to a doctor, but does not provide this schedule to a patient who requests it. We use Claude Opus 4.6 for structured evaluation, and Gemini 3 Flash as our primary judge, to score model responses against a physician's rubrics. Our primary judge agrees with physicians' omission scores about as well as physicians agree with each other. We find a decoupling gap of +0.38 (p = 0.003) on average across models. With our primary judge (checked by physicians) the decoupling gap is +0.22 (95% CI 0.10-0.36, p = 0.0014). We find three distinct patterns underlying this gap, exemplified by each of the models below. In the doctor framing, Claude Opus demonstrates that it has the information, and withholds it in the patient framing. Llama 4 performs poorly in both framings, meaning the decoupling gap cannot distinguish between withholding and incompetence. Finally, GPT-5.2 (excluded from this analysis) failed to return text for 33.2% of doctor responses, compared to 0% of layperson responses. In 86.6% of cases that we score (through our structured evaluation) as having omission harms, our primary judge (Gemini 3 Flash) scores zero omission harm. Because our scenarios are designed to pit safety against helpfulness, these statistics hold only for this distribution.
Read the full abstract ▾