paperlist.ai

AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

Super Library Agent: Joint Generation and Maintenance of Multiple Applications Beyond the Single Codebase

Daegyu Sung, Yukyeong Lee, Geon Park, Yumin Choi, Sung Ju Hwang

Organizations often develop and maintain portfolios of related applications: independently deployable codebases that share substantial domain logic, interface patterns, or operational conventions. As LLM coding agents are increasingly used to generate and maintain such software, a naive application-by-application workflow duplicates shared logic across codebases and allows prolonged agentic maintenance to accumulate verbosity, dead code, and structural erosion. We introduce the Super Library Agent problem, where an agent sequentially generates a portfolio of N related applications while maintaining a shared Super Library of reusable cross-application components. A minimal sequential scaffold can in principle extract shared code and migrate applications to the evolving library, but in practice suffers from low extraction recall and fragile dependency migration. We address these failures with candidate-guided extraction over code chunk summaries, pre-extraction codebase consolidation, and context-aware migration using extraction traces and call-graph information. Across WebGen-Bench and PaperBench, our method preserves application functionality while significantly reducing redundancy and token footprint (verbosity, token length) over zero-shot, and avoiding the structural erosion introduced by naive library construction, with additional reductions in LOC and MDL. Our code is available at https://github.com/sbigstar0310/super-library-agent.

arXiv

Detecting and Repairing Hallucinations in Retrieval-Augmented Generation

Sai Krishna Reddy Mulakkayala, Niki van Stein, Aske Plaat

Language models increasingly answer questions by consulting retrieved documents rather than memory alone, a design now common in search assistants and enterprise knowledge tools. Grounding a model in retrieved text reduces unsupported statements but does not eliminate them, and a reader cannot tell a grounded sentence from an invented one. Most research on this problem stops at detection, yet flagging a faulty answer changes nothing for the person reading it, and little is known about which action should follow. Using RAGTruth, a benchmark whose unsupported passages are annotated by hand, we split each flagged answer into individual factual claims, check each against the retrieved source, and compare leaving the answer untouched with three repair strategies of increasing richness: deleting an unsupported claim, replacing it with source text, and rewriting it. Three language models from different families judge the 916 repaired answers. Every strategy reduces the proportion of answers judged to contain unsupported content, and all three judges agree on the ordering. Deletion achieves the largest reduction while retaining least of the original answer, at 64.3% of the text, whereas rewriting retains 80.1% and reduces least. Repair is not confined to faulty answers: 83.5% of answers annotated clean are edited too. The strategies occupy different points on a grounding preservation trade-off rather than forming a quality ranking, and choosing between them needs evidence about answer usefulness that automatic metrics cannot supply.

arXiv

Learning Simple Test-Time Environments for LLM Web Agents

Junxuan Li, Zijun Liu, Ziyi Huang, Peng Li, Yuzhou Liu, Ming Yan, Yang Liu

Large language model (LLM) agents have demonstrated remarkable proficiency in manually constructed environments, yet their performance frequently collapses when transitioned to complex real-world settings. Existing research largely attribute this degradation to the compositional generalization gaps in LLMs on combinations of multiple simple, well-structured environments. In this work, we propose that LLM web agents can learn simple environment observations at test time. Specifically, we introduce trial steps for agents to decompose a complex environment observation into sub-modules, and implement a label-free learning method, Test-Time Environment Decomposition (TTED), to adapt agent behaviors with experience during inference. Our empirical evaluations demonstrate the framework's efficacy across both synthetic and realistic benchmarks, showing (1) experience gains acquired within simpler sub-environments can be effectively composed to improve performance in the full one, and (2) test-time training on sub-environments can significantly enhance the compositional generalization of agents in real-world web automation tasks. We also provide key insights in the design of the label-free learning algorithm. As more complex environments are accessed by LLM agents, we believe learning environment decomposition skills at test time will be critical for robust real-world deployment.

arXiv

Cloud and On-Premises Deployment of Uzbek Legal RAG via Targeted Retriever Fine-Tuning

Tatul Danielyan, Mariam Avetisyan, Hrant Davtyan

Deploying large language models for legal question answering raises challenges that general-purpose leaderboards do not capture, particularly for low-resource languages and under hard operational constraints. We report on building and operating a retrieval-augmented (RAG) legal assistant for Uzbek that must run in two regimes: a managed cloud service that maximizes answer quality within a per-token cost ceiling, and an on-premises deployment for clients whose legal data may not leave their infrastructure, restricting us to open-weight models on limited local hardware under latency constraints. Because no evaluation existed for this setting, we build two domain benchmarks: a retrieval benchmark of 178 expert-annotated legal queries with gold provision spans, and an end-to-end benchmark of 504 expert-curated question--answer pairs scored by an LLM judge whose ratings we validate against human judgments and against an independent-family judge. Applying these benchmarks under each regime, we find the open-versus-proprietary gap is small and cheaply closed by fine-tuning. Therefore, we train UTE-1, which is a state-of-the-art text embedder among open models for Uzbek. We also demonstrate that closing the performance gap via fine-tuning is both impractical due to the intensive hardware demands of long-context legal Q\&A and unnecessary, given that legal acts change frequently. We support this by reporting a negative result from a QLoRA experiment. We distill practical guidance for similar deployments, drawn from a system serving real users in production. We release our benchmarks, evaluation code and the fine-tuned embedder (UTE-1) \href{https://metric-ai-lab.github.io/Uzbek-Legal-RAG/}{at this https URL} to support future work on low-resource legal NLP.

arXiv

Modality Fault Lines: Structural Corruptions Reveal Fragile Omni-Modal Reasoning

Zhaolu Kang, Meixin Wu, Yu Xue, Yingjie He, Qiming Shi, Lei Wei, Yidi Wang, Richeng Xuan, Zhichao Hu

Omni-modal large language models are increasingly evaluated on clean text--vision--audio inputs, where every channel is present, synchronized, and readily interpretable. Such scores are often taken as evidence of robust cross-modal fusion, but clean evaluation cannot tell whether success depends on stable cross-modal structure or on cues sufficient only in intact inputs. To address this gap, we define a modality fault line: a boundary at which model behavior becomes unstable when a modality remains present and human-interpretable, but its internal evidence structure is perturbed. We introduce SCEval (Structure-Corruption Evaluation) a diagnostic evaluation protocol that keeps the question, answer space, and modality channels fixed while applying controlled structural corruptions to text, vision, and audio individually and jointly. Built from $273$ human-verified tri-modal examples from Social-IQ, OmniBench, and VALOR, SCEval evaluates $15$ proprietary and open-source omni-modal systems. The results show that structural corruption lowers clean accuracy, text--vision damage forms the most stable shared fault line, and multi-modal degradation is non-additive rather than a simple function of the number of corrupted modalities. Clean omni-modal accuracy therefore does not establish that a model will remain reliable when cross-modal evidence becomes structurally unreliable.

arXiv

SHADOWBENCH: Toward Reliable Automatic Evaluation of Semantic Alignment in Autoformalization

Hojae Han, Jongyoon Kim, Sanghyuk Park, Sanghyeok Park, Dongwook Cheon, Yeachan Park, Myungjae Jeon, Sunjong Choi, Myung Jae Jeon, Sunjong Choe, Soonho Kong, Wonseok Heo, Wonseok Hur, Seung-won Hwang, Donghoon Hyeon

Autoformalization translates informal mathematical theorems into code for proof assistants such as Lean. A central challenge is that current evaluation metrics can accept type-correct but misaligned statements or reject correct statements written in a different formulation. Inspired by Pass@$k$, we propose SA-Pass (*Semantic Alignment Pass*), which tests formal statements using auxiliary statements called *shadows* that characterize the intended statement. A generated statement receives full credit only when it compiles, implies each shadow (forward check), and is implied by their conjunction (backward check). We instantiate SA-Pass in ShadowBench, a Lean 4 full autoformalization benchmark of 178 postgraduate- to research-level problems spanning eight mathematical areas. Claude Code (Opus 4.8) with Numina-Lean-Agent reaches $61.8\%$ compile rate and $11.2\%$ SA-Pass. Across outputs generated by six agentic configurations, SA-Pass achieves $98.8\%$ binary agreement with expert judgments. An early version of ShadowBench served as the benchmark for Track 4 of the ICML 2026 AI4Math Challenge.

arXiv

Large Language Models Systematically Favor Popular Options: Evidence and Mitigation Across MCQs

Abdelrahman Abdallah, Mohammed Ali, Bhawna Piryani, Mahmoud Abdalla, Adam Jatowt

Multiple-choice questions (MCQs) are a standard format for evaluating large language models (LLMs), yet the popularity of answer options can confound evaluation. Modern LLMs systematically prefer popular but incorrect options over less popular correct ones, a vulnerability we call \textbf{popularity bias}. This pattern aligns with confidence miscalibration: model confidence remains high even as accuracy collapses for popular options. To systematically isolate this phenomenon, we introduce \textbf{PopMCQ}, a benchmark with six controlled strategies that vary option popularity while keeping the correct answer fixed. In our most adversarial setting, where all distractors are more popular than the correct option, models choose popular but wrong answers 66\% of the time. To mitigate this bias, we propose \textbf{PopDebias}, a lightweight inference-time correction that estimates and removes a popularity prior from model predictions. It requires no fine-tuning, is label-free at test time (using only a small calibration split for parameter fitting), and adds negligible computational cost. Experiments on 22 open-source LLMs (0.5B to 32B parameters) show consistent improvements, with accuracy gains up to 54.1 percentage points under strong popularity pressure. The code and data are available https://github.com/DataScienceUIBK/PopMCQ

arXiv

Validating FKG.in: Soundness Assessment in LLM-Augmented Indian Food Knowledge

Saransh Kumar Gupta, Armaan Shah, Lipika Dey, Partha Pratim Das, Ramesh Jain

The online culinary ecosystem is increasingly populated by recipe content generated, modified, or summarized by Large Language Models (LLMs). While often plausible, such outputs may contain hallucinated ingredients, misrepresented quantities, or culturally implausible combinations, limiting their suitability for downstream applications and knowledge graph construction. In this paper, we present a semi-automated soundness assessment workflow for validating structured recipe data extracted and augmented by LLMs from informal culinary sources. Developed as part of FKG(.in), a knowledge graph of Indian food, the pipeline identifies and addresses common failure modes, including structural inconsistencies, semantic and logical incoherence, and deviations from the source text, through a multi-stage process combining formal grammars, vocabulary-based checks, statistical heuristics, Set Transformer-based coherence modeling, and retrieval-based verification. Although evaluated on Indian recipes, the proposed methods are applicable to broader multilingual and multicultural culinary domains. We provide a practical, auditable, and application-agnostic framework for validating LLM-augmented recipe data, thereby strengthening the foundations of machine-readable food knowledge infrastructures in the era of LLM-generated content.

arXiv

When Patients Cut In: Extending Clinical Conversational AI Safety to Interruptions

Zachary Ellis, Spencer Hazel, Adam Brandt, Yajie Vera He, Ernest Lim, Jared Joselowitz

Clinical voice agents are now deployed in routine care, where real patients do not wait their turn: they interrupt. These systems typically use a cascaded architecture (speech-to-text -> LLM -> text-to-speech), so when a patient cuts the agent off mid-utterance, clinically required content can be lost even when the model handles cooperative transcripts well. Yet clinical conversational-AI benchmarks almost universally assume patients wait for the agent to finish, missing interruption-induced loss of required content. We present a transcript-based evaluation of interruption recovery, adapting conversation-analytic overlap categories into three operational types (recognitional, competitive, transitional sub-unit) and testing four deployment-oriented, non-reasoning LLM configurations across four cells spanning history-taking (information gathering) and FAQ (information provision), scored on whether the agent preserves the clinically required content. In the gathering cells, target-question failure varied across models; in the provision cells, where arms are directly comparable, failure rose for every model. Rankings differ across cells, and competitive FAQ interruption produced 30/30 provision-coverage failures for all four models (Wilson 95% CI: 88.6-100.0%; baseline 0/30 for three, 4/30 for Llama). A brief apology marker ("sorry to interrupt") shifts recovery by tens of percentage points, inconsistently across models, and for one it reduces recovery. Interruption robustness therefore cannot be a single score: evaluation must be content-grounded, reported per cell, and matched to the deployment's interruption profile.

arXiv

Anchoring Speech with Semantics: A Multimodal Adapter Mechanism for Automatic Speech Recognition in Low-Resource Languages

Kuan-Tang Huang, Cheng-Yeh Yang, Chien-Chun Wang, Hung-Shin Lee, Hsin-Min Wang, Berlin Chen

Low-resource ASR remains difficult because scarce transcripts provide limited supervised evidence for target-side generation. To address this gap, we propose SAMA-ASR, a lightweight adapter mechanism that augments the decoder with semantic anchors from auxiliary translations and an acoustic anchor from speech; in principle, the mechanism can be applied to similar encoder--decoder multitask speech models. Through cross-modal adaptation, SAMA-ASR conditions decoder states on translation-derived semantic embeddings and a speech embedding, combining utterance-level meaning with speech-grounded evidence before token prediction. At evaluation time, these semantic anchors can be generated automatically by an upstream speech-to-text translator rather than supplied as oracle translations. Experiments on two 30-hour datasets covering the low-resource Sinitic varieties Taiwanese Hokkien and Hakka show that SAMA-ASR improves over acoustic, prior prompt-based, and semantic-only translation-guided baselines and remains effective in practical automatic semantic-anchor settings; translator-capacity analyses show that useful semantic anchors can be produced by a compact ST model.

arXiv

Attribute-Based Activation Steering of LLMs for Group-Specific Explanation Generation

Leandra Fichtel, Janek Prange, Henning Wachsmuth

To effectively enable people to understand new topics, explanations should be tailored to their backgrounds and abilities. Prompting alone has been shown to be insufficient for creating such explanations, and other computational methods are missing so far. Therefore, this paper investigates whether LLMs can be steered to generate explanations that are tailored to a specific group of people. To this end, we propose an approach that first identifies group-specific attributes in terms of explanatory style and knowledge of a specific target group. Building on activation engineering, it then computes attribute-based steering vectors and adds them to the internal activations of an LLM during inference to enable a fine-grained steering. In our experiments, we assess the steering effectiveness in terms of specificity and factuality of the generated explanations. Additionally, we evaluate the explanations in a study with human experts from different target groups. Compared to prompting and state-of-the-art steering baselines, our approach tailors the explanations significantly better to the target group while maintaining the best specificity-factuality balance.

arXiv

Toward Cultural Alignment: Human-Centered Evaluation of Multimodal AI Stories Across Five African Communities

Millicent Ochieng, Felermino D. M. A. Ali, Elizabeth A. Ankrah, Najeeb Gambo Abdulhamid, Migisha Boyd, Stephanie Nyairo, Mercy Muchai, Samuel Chege Maina, Aditya Vashistha, Anja Thieme, Jacki O'Neill

In this paper, we examine how well AI-generated multimodal stories align with the lived practices, relationships, language, values, and visual expectations of the communities they represent. We conduct a community-grounded mixed-methods evaluation with 19 culture representatives across five African communities, combining quantitative annotations with qualitative focus group discussions. We find that cultural alignment depends not simply on recognizable cultural markers, but on how those markers fit social, linguistic, procedural, and visual context. From these evaluations, we develop a taxonomy of cultural alignment comprising five broader cultural marker categories and eight recurring mechanisms of misalignment. We additionally evaluate five multimodal LLM judges to examine whether automated evaluation can approximate community-grounded judgments at scale. Judge reliability and score calibration vary substantially across communities, with no single judge performing consistently across all five settings. These findings motivate community-calibrated evaluation pipelines in which automated judges are validated against community judgments to determine where they can be trusted and where human review remains necessary.

arXiv

How Identity and Opinion Shape Political Sycophancy in LLMs

Li-Ni Fu, Chang Chih Meng, Chien-Hua Chen, Hen-Hsen Huang, I-Chen Wu

As Large Language Models (LLMs) increasingly encourage users to disclose personal profiles for tailored assistance, measuring their political alignment becomes increasingly important. However, many existing benchmarks for assessing political behavior rely on closed-ended questions and do not fully capture how a model's stance may adapt to user-provided context during interaction. We introduce a framework that disentangles two distinct triggers of political sycophancy: opinion (aligning with explicit narratives) and identity (stereotyping based on demographic labels). Using 450 manually-checked political dilemmas as controlled probes, we evaluate 13 instruction-tuned LLMs. We uncover a dissociation: a model's susceptibility to explicit opinions does not necessarily predict its susceptibility to identity cues, and vice versa. When both signals are present, their effects are generally sub-additive rather than simply additive. Additionally, system-level personas primarily shift a model's baseline stance while having limited effect on the stance shift caused by user opinion or identity. Ultimately, our results suggest that LLM political stance is interactively and steerably vulnerable rather than being a fixed trait, highlighting how personalization may amplify identity- or opinion-conditioned shifts in the model's behaviors.

arXiv

Rate-Coding Bundle Memory: A Unified Model of Memory and Control for Symbolic Computation in the Brain

Teun van Gils, Rowan P. Sommers, Markus Ostarek, Peter Hagoort

We propose a neurobiologically plausible model of cognition that combines the advantages of connectionist and symbolic systems, and that can explain a wide range of cognitive phenomena. This model, called Rate-Coding Bundle Memory (RCBM), is based on the Symbolic Subsystem Hypothesis, which posits that the brain implements a symbolic subsystem within its fundamentally connectionist nature. RCBM is a hybrid model that uses rate coding to represent symbols in a continuous space, and it uses a bundle memory system to store and retrieve these symbols. The model is capable of solving a wide range of cognitive phenomena, including one-shot learning, pattern separation, and the binding problem. We argue that RCBM provides a promising framework for understanding the nature of cognition, and that it can be used to develop more sophisticated models of cognition in the future.

arXiv

Locked at the Entrance, Open Inside: Where RLVR Narrows the Solution Space

Qiancheng Zhou, Ruizhe Li

Reinforcement learning with verifiable rewards (RLVR) substantially improves single-sample accuracy (pass@1) but causes the policy's solution space to contract, diminishing the returns of test-time scaling. In this work, we investigate where inside a reasoning trajectory this breadth is lost: does the policy fail to access a valid solution family, or does it fail to execute computation once initiated? To disentangle access from execution, we analyze the Countdown task, whose solution space can be exhaustively enumerated into discrete entrance families defined by the first operand and operator, across PPO on Qwen2.5-3B and GRPO on Qwen2.5-3B-Instruct. Across both training setups, solution coverage falls by up to 67%, halving even on problems solved across all checkpoints. We show that this contraction is heavily concentrated at the entrance: per-token likelihood shifts are 11x--16x larger prior to the first arithmetic operation than during downstream reasoning. Supplying only an unselected entrance prefix restores completion rates in low-access families by over an order of magnitude (0.018 -> 0.212 under PPO), demonstrating that alternative solutions remain executable but are no longer initiated. Guided by this localization, we find that while surface prompting fails to recover diversity, entrance-targeted interventions succeed: late-layer parameter interpolation with early checkpoints increases solution coverage by 37% at no loss in pass@1. Finally, we show that early-step entropy collapse recurs across six math benchmarks with 7B and 14B models, but is not an inevitable byproduct of reasoning optimization: an SFT baseline preserves more than double the coverage, and staged SFT--DPO--RLVR pipelines retain early-step entropy. In summary, reasoning breadth is lost at the door, not inside the room. Code: https://github.com/ershiyidian/early-branch-locking.

arXiv

Toward a Cross-Lingual Romanization Ecosystem for Sinitic Languages: A Paired Mandarin-Cantonese Case Study

Zijie Zhang, Tan Lee, Yong Cao, Benyou Wang

This paper proposes the Sinitic Romanization Ecosystem, a cross-lingual Sinitic romanization design framework with supporting digital infrastructure and a community-driven open-source workflow. The design framework addresses the lack of systematic cross-lingual romanization alignment among Sinitic languages through four design principles: phonetic correspondence for representing similar sounds with similar romanized symbols, historical-phonological correspondence for aligning cognate romanization strings, one-phoneme-one-symbol, and basic Latin-letter use, with a balancing consideration recognizing trade-offs among these principles. For the main paired case study, we devel-op CantRomZJ1 and MandRomZJ1, Cantonese and Manda-rin romanization schemes following the design framework, respectively. We also develop schemes for several other Sinitic languages, including Meixian Hakka, Shanghai Wu, and Nanjing Jianghuai Mandarin, following the same de-sign framework. To bring the romanization schemes into practical use, we develop open-source infrastructure for structured romanization storage, conversion, parsing, dic-tionary construction, and input-method generation. Finally, we evaluate the design framework through speech-to-romanization experiments based on Meta's Massively Mul-tilingual Speech (MMS) fine-tuning. Compared with the Pinyin+Jyutping baseline, our Man-dRomZJ1+CantRomZJ1 condition reduces Cantonese WER and CER by 7.80% and 10.61%, respectively. These results suggest that cross-lingual romanization alignment can improve transfer in low-resource Sinitic speech technology.

arXiv

Quantifying Error Tolerance in Synthetic Data: An Atomic-level Operand vs. Operator Perturbation Study

Jiaxiang Liu, Chenhao Yuan, Shuwen Xu, Boxuan Xing, Xiusheng Huang, Yinhao Xu, Hao Liu, Wenhao Teng, Xiangwen Liao, Pengfei Cao, Jun Zhao, Kang Liu

Synthetic data generation has become a cornerstone for advancing large language models. However, the lack of the quantitative analysis for error tolerance became a critical bottleneck. Consequently, current filtering strategies fluctuate between two extremes: they are either overly aggressive, risking the exclusion of potentially valuable samples, or overly permissive, failing to eliminate erroneous samples effectively. To bridge this gap, this paper introduces Atomic Tree Operation Modeling (ATOM), a framework that decomposes data into functional units ($f(x)\rightarrow y$). ATOM distinguishes benign Operand $x$ perturbations from fatal Operator $f$ perturbations. The former are needlessly discarded by aggressive filtering, while the latter slip through permissive filtering. Our experiments reveal a double dissociation: models are robust to operand perturbations but collapse under operator perturbations. By prioritizing operator over aggressive operand precision, our ATOM-synthesized data outperforms rigorous baselines (e.g., +3.1% gain over LIMA), suggesting that operator diversity matters more than operand precision. Our code is available at https://github.com/Lut-hub/ATOM.

arXiv

AI Historian: Helping historians organize and verify person-centred temporal clues from dispersed historical narratives

Yifeng Lu, Zijie Yang, Jie Li, Qingkai Min, Yue Zhang

History is not preserved in complete, continuous form. Accounts of a person's activities, relationships and historical contexts are scattered across texts, chapters and narrative perspectives; historians must retrieve, identify and compare these materials to reconstruct temporal sequences and verify them against sources. Here we present AI Historian (AIH), an AI agent system that helps historians organize person-time evidence from dispersed biographical narratives. It takes source sentences as evidence units, identifies people and temporal cues, verifies candidate cross-text associations and infers comparable temporal ranges while preserving traceable source-text evidence. We evaluated AIH on six Shiji cases concerning Liu Bang, Xiang Yu and Xiao He. AIH Agent achieved a temporal-localization MicroIoU of 86.2%, compared with 81.3% for human-only annotation and 17.1% for direct large-language-model prompting; it required about 14 min, versus 1 h 32 min for human-only annotation. We further applied AIH to the Twenty-Four Histories and other ancient Chinese histories, ancient Japanese and Korean histories, and modern and contemporary historical materials, and released the results through Westlake Historian. These results indicate that AIH can reduce the cost of organizing historical materials at scale while turning connections obscured by chapter-based narration into traceable, revisable research questions for collaborative testing.

arXiv

HEAR Who Said What: Unlocking Speaker-Attributed Reasoning via Counterfactual Voice Grounding

Dongwook Lee, Sangkwon Park, Eunwoo Song, Che Hyun Lee, Youngho Cho, Junho Kim, June Young Yi, Heeseung Kim, Sungroh Yoon

Speech Language Models (SLMs) are increasingly deployed in multi-speaker environments, yet their ability to attribute speech to the correct speaker and reason over speaker identities remains unclear. Hence, we introduce HEAR, a conceptually hierarchical benchmark diagnosing the foundational capabilities of speaker-attributed reasoning, comprising 2.4K human-verified samples from 887 diverse multi-party audio clips. Evaluating 20 leading SLMs on HEAR reveals they struggle with these foundational tasks, often relying on semantic priors rather than actual vocal cues. To address this, we present A2R, a 30B model optimized on Counterfactual Audio with Speaker-level Hard negatives (CASH), a dataset designed to guide the model to prioritize acoustic vocal cues over linguistic signals. A2R achieves strong performance on HEAR and exhibits zero-shot generalization to diverse multi-speaker downstream tasks, demonstrating that learned speaker attribution unlocks the model's latent capacity for speaker-aware reasoning. All resources are available at https://attributetoreason.github.io/AttributeToReason/

arXiv

Emergent Misalignment Is Not Magical

Mingxuan Li, Qirun Dai, Heran Wang, Chenhao Tan

Fine-tuning large language models (LLMs) on narrowly harmful datasets can lead to misalignment broadly, a phenomenon known as emergent misalignment (EM). EM poses a challenge for AI safety and our understanding of LLMs. Prior work often frames EM as an unexpected behavior, and explains it by appealing to general misalignment directions or anthropomorphizing it as acquiring an evil persona. However, the mechanisms behind these framings remain obscure. In this work, we show that EM is a predictable and data-dependent generalization phenomenon. By examining the base model's representation of EM training data and evaluation prompts, we find that evilness after EM training is highly predictable from representational distance: the closer an evaluation prompt is to training data centroid, the more evilness it elicits from EM models after training (with an average Spearman correlation of -0.73 across 12 model-dataset settings). Building upon this analysis, we further demystify EM by showing that (1) its effectiveness changes significantly based on training data format; (2) there is not a general misalignment direction that transfers across different EM models; (3) the effect of EM is fundamentally different from persona changes. Furthermore, we extend the EM generalization metric from a scalar distance to a dataset-specific generalization direction, which robustly predicts EM models' evilness under semantics-preserving prompt perturbations including appending random tokens and paraphrasing, where other methods do not reliably generalize.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.