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

What Makes a Sale? Simulating End-to-End Seller--Buyer Retail Dynamics with LLM Agents

Jeonghwan Choi, Jibin Hwang, Gyeonghun Sun, Minjeong Ban, Taewon Yun, Hyeonjae Cheon, Hwanjun Song

Evaluating retail strategies before deployment is difficult, as outcomes are determined across multiple stages, from seller-side persuasion through buyer-seller interaction to purchase decisions. However, existing retail simulators capture only partial aspects of this process and do not model cross-stage dependencies, making it difficult to assess how early decisions affect downstream outcomes. We present RetailSim, an end-to-end retail simulation framework that models this pipeline in a unified environment, explicitly designed for simulation fidelity through diverse product spaces, persona-driven agents, and multi-turn interactions. We evaluate RetailSim with a dual protocol comprising human evaluation of behavioral fidelity and meta-evaluation against real-world economic regularities, showing that it successfully reproduces key patterns such as demographic purchasing behavior, the price-demand relationship, and heterogeneous price elasticity. We further demonstrate its practical utility via decision-oriented use cases, including persona inference, seller-buyer interaction analysis, and sales strategy evaluation, showing RetailSim's potential as a controlled testbed for exploring retail strategies.

arXiv

VLMs Need Words: Vision Language Models Ignore Visual Detail In Favor of Semantic Anchors

Haz Sameen Shahgir, Xiaofu Chen, Yu Fu, Erfan Shayegani, Nael Abu-Ghazaleh, Yova Kementchedjhieva, Yue Dong

Vision-language models (VLMs) have achieved impressive performance across a wide range of multimodal tasks. However, they often fail on tasks that require fine-grained visual perception, even when the required information is still present in their internal representations. Prior work has attributed this ``hidden-in-plain-sight'' gap to the language model, but the cause remains unexplained. In this work, we demonstrate that this gap arises from the language model's lack of semantic labels for fine-grained visual details: when visual entities can be mapped to known concepts, VLMs bypass visual comparison and reason through language; when they cannot, VLMs resort to brittle and hallucinated descriptions. We verify this across semantic correspondence, synthetic shape matching, and face matching, and find that VLMs perform much better when the relevant entities are nameable than when they are unnamable. Mechanistically, Logit Lens analysis confirms that VLMs explicitly recover semantic labels for nameable entities and surface more unique tokens compared to unnameable entities. Furthermore, we show that this limitation can be addressed: teaching completely arbitrary names for unknown entities improves performance. More importantly, task-specific finetuning yields even stronger generalization without relying on language priors, i.e., through real visual perception. Our findings suggest that current VLM failures on visual tasks reflect a learned shortcut rather than a fundamental limitation of multimodal reasoning. Code and datasets are available at https://github.com/Patchwork53/VLMs-Need-Words-COLM2026.

arXiv

LogitScope: A Framework for Analyzing LLM Uncertainty Through Information Metrics

Farhan Ahmed, Yuya Jeremy Ong, Chad DeLuca

Understanding and quantifying uncertainty in large language model (LLM) outputs is critical for reliable deployment. However, traditional evaluation approaches provide limited insight into model confidence at individual token positions during generation. To address this issue, we introduce LogitScope, a lightweight framework for analyzing LLM uncertainty through token-level information metrics computed from probability distributions. By measuring metrics such as entropy and varentropy at each generation step, LogitScope reveals patterns in model confidence, identifies potential hallucinations, and exposes decision points where models exhibit high uncertainty, all without requiring labeled data or semantic interpretation. We demonstrate LogitScope's utility across diverse applications including uncertainty quantification, model behavior analysis, and production monitoring. The framework is model-agnostic, computationally efficient through lazy evaluation, and compatible with any HuggingFace model, enabling both researchers and practitioners to inspect LLM behavior during inference.

arXiv

Pingala: Prosody-Aware Decoding for Sanskrit Poetry Generation

Manoj Balaji Jagadeeshan, Atul Singh, Nallani Chakravartula Sahith, Amrith Krishna, Pawan Goyal

Poetry generation in Sanskrit typically requires the verse to be semantically coherent and adhere to strict prosodic rules. In Sanskrit prosody, every line of a verse is typically a fixed length sequence of syllables adhering to prescribed binary patterns of syllable weights. We observe that instead of treating a verse as a monolithic sequence, segmenting them as grouped-lines leads to significant improvement in semantic coherence by 10\% with comparable metrical adherence. Specifically, Pingala, our proposed decoding approach is designed to encourage every line to have well-formed words and our token selection biases the model towards it by preferring longer tokens. Writing in Sanskrit follows phonemic orthography, hence using a phonetically aware transliteration scheme, SLP1, increased the metrical alignment by 46\% with comparable semantic similarity, for a instruction fine-tuned large language models like Phi-4. We also introduce a new approach for reference-free evaluation using cross-encoders which achieved better alignment with true poetry instances.

arXiv

Word Recovery in Large Language Models Enables Character-Level Tokenization Robustness

Zhipeng Yang, Shu Yang, Lijie Hu, Di Wang

Large language models (LLMs) trained with canonical tokenization exhibit surprising robustness to non-canonical inputs such as character-level tokenization, yet the mechanisms underlying this robustness remain unclear. We study this phenomenon through mechanistic interpretability and identify a core process we term word recovery. We first introduce a decoding-based method to detect word recovery, showing that hidden states reconstruct canonical word-level token identities from character-level inputs. We then provide causal evidence by removing the corresponding subspace from hidden states, which consistently degrades downstream task performance. Finally, we conduct a fine-grained attention analysis and show that in-group attention among characters belonging to the same canonical token is critical for word recovery: masking such attention in early layers substantially reduces both recovery scores and task performance. Together, our findings provide a mechanistic explanation for tokenization robustness and identify word recovery as a key mechanism enabling LLMs to process character-level inputs.

arXiv

Quantifying Hallucinations in Language Language Models on Medical Textbooks

Brandon C. Colelough, Davis Bartels, Dina Demner-Fushman

Hallucinations, the tendency for large language models to provide responses with factually incorrect and unsupported claims, is a serious problem within natural language processing for which we do not yet have an effective solution to mitigate against. Existing benchmarks for medical QA rarely evaluate this behavior against a fixed evidence source. We ask how often hallucinations occur on textbook-grounded QA and how responses to medical QA prompts vary across models. We conduct two experiments, the first experiment to determine the prevalence of hallucinations for a prominent open source large language model (LLaMA-70B-Instruct) in medical QA given closed-source zero-shot prompts, and the second experiment to determine the prevalence of hallucinations and clinician preference to model responses. We observed, in experiment one, with the passages provided, LLaMA-70B-Instruct hallucinated in 19.7\% of answers (95\% CI 18.6 to 20.7) even though 98.8\% of prompt responses received maximal plausibility, and observed in experiment two, across models, lower hallucination rates aligned with higher usefulness scores ($ρ=-0.71$, $p=0.058$). Clinicians produced high agreement (quadratic weighted $κ=0.92$) and ($τ_b=0.06$ to $0.18$, $κ=0.57$ to $0.61$) for experiments 1 and 2 respectively. Our findings indicate that, across all scales and architectures tested, current large language models remain unfit for unsupervised clinical deployment, and that human expert oversight is both necessary and the dominant cost driver.

arXiv

Suffix-Constrained Greedy Search Algorithms for Causal Language Models

Ayoub Hammal, Pierre Zweigenbaum, Caio Corro

Large language models (LLMs) are powerful tools that have found applications beyond human-machine interfaces and chatbots. Beside free-form generation, there has been an interest in constrained generation, a setting where LLMs are constrained to generate well-formed outputs with respect to the language defined by a formal grammar. Although appealing, this setting may be over restrictive for downstream applications. For example, many LLM tasks require the model to reason freely before generating its final response in a specific format. In this work, we introduce suffix-constrained generation, a constrained generation setting in which only the end of the response is constrained by a grammar, a scenario that is not supported by existing constrained generation methods. We introduce several suffix-constrained generation algorithms that are based on greedy search. We experiment on several datasets, and show that our approach allows to guarantee suffix constraints without having a negative impact on results, and even improving them in many settings.

arXiv

ChiEngMixBench: Evaluating Large Language Models on Expert-Style Chinese-English Terminology Mixing

Qingyan Yang, Tongxi Wang, Yunsheng Luo

Large language models increasingly mediate multilingual professional communication, where useful generation requires adapting to community conventions about which expressions are retained, translated, or mixed. Existing benchmarks rarely isolate such community-conditioned choices. We introduce ChiEngMixBench, a controlled benchmark for Chinese AI/CS discourse, where Chinese frames routinely incorporate established English technical terms. Built from public technical discussions, it contains 1,706 source-derived candidate pairs covering 1,344 non-empty normalized terms, including a 1,167-pair strict subset that fixes the Chinese prefix and syntactic position while varying only the terminology form. The benchmark combines paired likelihood comparisons with a transparent reference-profile diagnostic for open-ended responses. Across nine open-weight models, Chinese equivalents receive higher likelihood on most pairs, revealing a gap between source-attested usage and model preference. Specialized terms show a small directional lift that is not robust after frequency and length controls and multiple-comparison correction. Human evaluation and baseline analyses show that reference-profile conformity is informative under the intended mixed-style rubric but does not reliably predict holistic response preference. ChiEngMixBench provides a reusable testbed for community-specific multilingual conventions with explicit diagnostic boundaries.

arXiv

Where Knowledge Collides: A Mechanistic Study of Intra-Memory Knowledge Conflict in Language Models

Minh Vu Pham, Hsuvas Borkakoty, Yufang Hou

In language models (LMs), intra-memory knowledge conflict arises when inconsistent information about the same subject is encoded within the model's parametric knowledge. Prior work has primarily focused on resolving conflicts between a model's internal knowledge and external sources, which is known as context-memory knowledge conflict, through approaches such as fine-tuning or knowledge editing, while the understanding of conflicts that arise internally remains largely unexplored. In this work, we design a framework to identify where internal conflicting knowledge is encoded within LMs. We test our framework on four LMs using both synthetic and real-world knowledge conflicts. We find that internal conflicts often arise and are resolved in the final layers across all models, but that interventions are markedly less effective on real-world knowledge conflicts. Targeted attention-head interventions outperform layer-wise ones, and a filtering analysis shows that heads specialized for a single competing fact are far more common in synthetic conflicts, helping explain this gap. Finally, we find no evidence of a single universal circuit for handling knowledge conflict. Instead, our results suggest that distinct circuits may separately encode competing pieces of knowledge, giving rise to conflict. Our results offer a first mechanistic account of intra-memory conflict resolution and highlight a substantial gap between synthetic and real-world settings.

arXiv

STAGE: A Full-Screenplay Benchmark for Reasoning over Evolving Stories

Qiuyu Tian, Yiding Li, Yingce Xia, Fengyi Chen, Youyong Kong, Fan Guo, Yuyao Li, Jinjing Shen, Zhijing Xie, Yiyun Luo, Xin Zhang, Zequn Liu

Movie screenplays are a demanding testbed for long-form narrative understanding, as characters' goals, beliefs, knowledge, and relationships evolve continuously across scenes. However, existing benchmarks primarily evaluate isolated facts from the completed screenplay, leaving unassessed whether models can track the evolving state of characters as the story unfolds. We introduce STAGE, a benchmark over 151 English and Chinese full-length screenplays, built on a provenance-linked narrative backbone that recovers the state and epistemic access of each character at every point along its timeline. Three tasks derived from the backbone jointly probe whether models can maintain, explain, and act on evolving narrative state: Character Development Tracking updates a focal character's state between checkpoints, Cross-Scene Narrative Evolution Reasoning targets cross-scene state transitions, and In-Script Character Role-Playing requires responses bounded by the character's state and knowledge at a specified point. We identify three failure modes of current LLMs: silent forgetting under recursive state updating, limited cross-scene reasoning even when all relevant evidence is supplied, and a trade-off in role-playing where stylistic character fidelity and screenplay-grounded memory faithfulness are optimized by different memory-access strategies. STAGE thus provides a unified framework for diagnosing how current models fail to track, reason about, and enact story evolution.

arXiv

Self-Guided Adaptive Safety Alignment: Synthesizing and Internalizing Guidelines in Reasoning Models

Yuhang Wang, Yanxu Zhu, Jiaming Zhang, Dongyuan Lu, Jitao Sang

Explicit safety policies can improve reasoning-model safety, but their effective coverage may lag behind evolving jailbreak strategies. We study whether a reasoning model can synthesize and internalize a task-specific safety guideline from a small set of harmful and benign examples. We introduce Self-Guided Adaptive Safety Alignment (SGASA). The model generates a guideline, refines it on its own errors, and selects a version by self-evaluation, which can then be applied in context or distilled into the model for guideline-free inference. Across two adversarial prompt datasets and three Qwen3 scales, in-context guidelines improve a combined safety and non-over-refusal score by 20.0-45.5 points on the 8B and 14B models. Self-evaluation selects the externally best refinement round in five of six settings, while guideline generation and utilization show distinct scaling patterns. Using alignment supervision derived only from WildJailbreak, internalized models retain 13.3-14.1 point gains on WildJailbreak without an inference-time guideline. These results support self-generated guidelines as a useful intermediate representation for low-resource safety adaptation.

arXiv

$π$-Attention: Online Efficient Sparse Transformers for Long-Context Modeling

Pike D. Liu, Chang Liu, Yanxuan Yu

Sparse attention is crucial in long-context Transformers, which restricts each token to a limited neighborhood and thereby reduces the quadratic cost of full self-attention. Local windows capture nearby context effectively, yet they induce a receptive-field bottleneck for dependencies beyond the window, limiting long-range modeling under moderate depth. In this paper, we propose $π$-Attention, an \emph{online efficient} sparse attention operator: as tokens arrive, each step maintains a streaming working set of local neighbors plus a $π$-indexed long-range fetch, fused by an adaptive prior under a shared softmax. Rather than materializing a global sparse mask in advance, $π$-Attention computes attention on the live working set with hierarchy-aware IO. We analyze causal reachability and minimum depth under this online rule, and show per-step cost remains $\mathcal{O}(k)$. Experiments on language modeling, Long Range Arena, and efficiency profiling---across 4K--32K context lengths---show consistent gains over local-window and other sparse baselines, approaching dense attention quality at linear cost.

arXiv

MIDI-LLM: Improving Text-to-MIDI Music Generation via Adapting Large Language Models

Shih-Lun Wu, Dave Carlton, Ryan Miyakawa, Yoon Kim, Chris Donahue, Cheng-Zhi Anna Huang

We present MIDI-LLM, a recipe that improves multitrack text-to-MIDI generation via adapting Large Language Models (LLMs). MIDI-LLM expands an LLM's text vocabulary to include MIDI tokens and employs a two-stage training pipeline: (i) unimodal continued pretraining on music-adjacent text and standalone MIDIs, and (ii) multimodal supervised finetuning on text-MIDI pairs. Our instantiation of MIDI-LLM based on Llama 3.2 (1B) outperforms the recent Text2midi model in both text control and musical quality, and readily integrates with optimized inference ecosystems like vLLM. To align with real-world songwriting workflows, we further finetune our MIDI-LLM on the TheoryTab dataset for text-conditioned lead sheet (i.e., melody + chords) generation and infilling. A comprehensive ablation study validates the synergy between LLM text pretraining, standalone MIDI pretraining, and supervised text-to-MIDI finetuning. Finally, an in-the-wild blind user study conducted in a real-world creative workflow at scale with 58 Hookpad Aria users and 4,002 generated outputs demonstrates that our MIDI-LLM achieves the highest acceptance rate in zero-to-one lead sheet generation over baselines without text control or LLM pretraining, confirming its efficacy in human-AI music co-creation.

arXiv

Toward Understanding the Transferability of Adversarial Suffixes in Large Language Models

Sarah Ball, Niki Hasrati, Alexander Robey, Avi Schwarzschild, Frauke Kreuter, Zico Kolter, Andrej Risteski

Discrete optimization-based jailbreaking attacks on large language models aim to generate short, nonsensical suffixes that, when appended onto input prompts, elicit disallowed content. Notably, these suffixes are often transferable -- succeeding on prompts and models for which they were never optimized. And yet, despite the fact that transferability is surprising and empirically well-established, the field lacks a rigorous analysis of when and why transfer occurs. To fill this gap, we identify three statistical properties that strongly correlate with transfer success across numerous experimental settings: (1) how much a prompt without a suffix activates a model's internal refusal direction, (2) how strongly a suffix induces a push away from this direction, and (3) how large these shifts are in directions orthogonal to refusal. On the other hand, we find that prompt semantic similarity only weakly correlates with transfer success. These findings lead to a more fine-grained understanding of transferability, which we use in interventional experiments to showcase how our statistical analysis can translate into practical improvements in attack success.

arXiv

Obfuscation Rules for Detecting and Detoxifying Korean Toxicity

Yejin Lee, Su-Hyeon Kim, Hyundong Jin, Dayoung Kim, Yeonsoo Kim, Yo-Sub Han

As language models become increasingly deployed in online environments, toxicity detection and detoxification have received growing attention. Existing studies primarily focus on non-obfuscated text, which limits robustness when users intentionally disguise toxic expressions. In particular, Korean toxic expressions can be easily disguised through agglutinative morphology and Hangeul-specific orthographic variation. However, obfuscation in Korean remains largely unexplored, which motivates us to introduce a KOTOX: Korean toxic dataset for deobfuscation and detoxification. We categorize Korean obfuscation patterns into linguistically grounded classes, define transformation rules derived from real-world examples, and provide the resulting obfuscation framework as an open transformation package. Using these rules, we provide paired neutral and toxic sentences alongside their obfuscated counterparts. Models trained on our dataset better handle obfuscated text without sacrificing performance on non-obfuscated text. This is the first dataset that simultaneously supports deobfuscation and detoxification for the Korean language. We expect the dataset to facilitate better understanding and mitigation of obfuscated toxic content in LLM for Korean. Our code and data are available at https://github.com/leeyejin1231/KOTOX.

arXiv

Token Buncher: Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning

Weitao Feng, Lixu Wang, Peizhuo Lv, Tianyi Wei, Jie Zhang, Chongyang Gao, Sinong Zhan, Wei Dong

As large language models (LLMs) continue to grow in capability, so do the risks of harmful misuse through fine-tuning. While most prior studies assume that attackers rely on supervised fine-tuning (SFT) for such misuse, we systematically demonstrate that reinforcement learning (RL) enables adversaries to more effectively break safety alignment and facilitate more advanced harmful task assistance, under matched computational budgets. To counter this emerging threat, we propose TokenBuncher, the first effective defense specifically targeting RL-based harmful fine-tuning. TokenBuncher suppresses the foundation on which RL relies: model response entropy. By constraining entropy, RL-based fine-tuning can no longer exploit distinct reward signals to drive the model toward harmful behaviors. We realize this defense through entropy-as-reward RL and a Token Noiser mechanism designed to prevent the escalation of harmful capabilities. Extensive experiments across multiple models and RL algorithms show that TokenBuncher robustly mitigates harmful RL fine-tuning while preserving benign task performance and finetunability. Our results highlight that RL-based harmful fine-tuning poses a greater systemic risk than SFT, and that TokenBuncher provides an effective and general defense.

arXiv

Uncovering Spontaneous Physics Representations in In-Context Learning

Yeongwoo Song, Jaeyong Bae, Dong-kyum Kim, Hawoong Jeong

In-context learning (ICL) lets large language models (LLMs) solve new tasks from prompts alone, across an ever-widening range of domains, yet the mechanisms underlying this ability remain poorly understood. Physical systems offer a controlled testbed for this question as they provide experimentally controllable data with structured dynamics grounded in fundamental principles. Here we study the ICL ability of LLMs, focusing on physical reasoning. Using dynamics forecasting as a proxy task, we first show that LLMs forecast physical dynamics in context, with accuracy improving as more history is provided. Analyzing the model's residual stream reveals internal activations that correlate with key physical quantities such as energy. These correlations strengthen gradually with context length, indicating that LLMs spontaneously form representations aligned with physical concepts without any physics-specific supervision. To assess whether these representations contribute to the model's predictions, we introduce a layer-wise gradient-based attribution analysis. We find that, residual directions more strongly correlated with energy also receive greater attribution to numerical predictions. This pattern is not observed for features correlated with directly observed quantities such as displacement, suggesting that the energy-related signal is not merely numerical information copied from the input. Our results broaden ICL analysis to structured physical dynamics and give a mechanistic account of how LLMs organize physical structure in context.

arXiv

Speech LLMs in Low-Resource Scenarios: Data Volume Requirements and the Impact of Pretraining on High-Resource Languages

Seraphina Fong, Marco Matassoni, Alessio Brutti

Large language models (LLMs) have demonstrated potential in handling spoken inputs for high-resource languages, reaching state-of-the-art performance in various tasks. However, their applicability is still less explored in low-resource settings. This work investigates the use of Speech LLMs for low-resource Automatic Speech Recognition using the SLAM-ASR framework, where a trainable lightweight projector connects a speech encoder and a LLM. Firstly, we assess training data volume requirements to match Whisper-only performance, re-emphasizing the challenges of limited data. Secondly, we show that leveraging mono- or multilingual projectors pretrained on high-resource languages reduces the impact of data scarcity, especially with small training sets. Using multilingual LLMs (EuroLLM, Salamandra) with whisper-large-v3-turbo, we evaluate performance on several public benchmarks, providing insights for future research on optimizing Speech LLMs for low-resource languages and multilinguality.

arXiv

The Eloquence team submission for task 1 of MLC-SLM challenge

Lorenzo Concina, Jordi Luque, Alessio Brutti, Marco Matassoni, Yuchen Zhang

In this paper, we present our studies and experiments carried out for the task 1 of the Challenge and Workshop on Multilingual Conversational Speech Language Model (MLC-SLM), which focuses on advancing multilingual conversational speech recognition through the development of speech language models architectures. Given the increasing relevance of real-world conversational data for building robust Spoken Dialogue Systems, we explore three approaches to multilingual ASR. First, we conduct an evaluation of the official baseline to better understand its strengths and limitations, by training two projectors (linear and qformer) with different foundation models. Second we leverage the SLAM-ASR framework to train a custom multilingual linear projector. Finally we investigate the role of contrastive learning and the extended conversational context in enhancing the robustness of recognition.

arXiv

BOW: Training Language Models to Reason Over Plausible Next Words

Ming Shen, Zhikun Xu, Jacob Dineen, Xiao Ye, Ben Zhou

Next-word prediction (NWP) trains language models against a single observed continuation, even though many contexts admit multiple plausible next words. Recent RL-based next-word reasoning methods make this tension explicit: they reward a model for producing a rationale that supports one context-conditioned continuation, which can turn a pre-existing preference into a confident, self-justifying trajectory. We introduce BOW, an RL framework that instead trains models to produce self-contained, neutral, and comprehensive descriptions of the plausible next-word space. The policy generates a next-word reasoning trajectory from the full context, but a frozen scorer computes the core reward from that trajectory alone, without separately receiving the context. BOW-Reg adds a lightweight breadth regularizer to this core reward to discourage premature collapse. On two model backbones, BOW remains competitive with the original models and often outperforms trained baselines across ten general reasoning benchmarks. On benchmarks testing ambiguous references and word meanings, BOW-Reg achieves the highest SharedRef correctness and the lowest HoWN-Simple single-sense collapse on both backbones. Human evaluation further shows that BOW-Reg produces broader next-word reasoning trajectories, while direct next-word-prediction evaluation shows that these trajectories remain predictive.

Keep browsing

More papers from top AI conferences

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