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

HANDBOOK.md: A Benchmark for Long-Context Agentic Instruction Following

Liudas Panavas, Sebastian Minus, Bradley Monton, Derek Ray, Suhaas Garre, Sushant Mehta, Edwin Chen

Language-model agents are increasingly deployed under standing instructions: a system prompt, a policy file, or a skills document is placed in context, and the agent is trusted to let that document govern every action that follows. Existing benchmarks rarely test this deployment pattern directly; they measure whether an agent can complete a task, not whether a long, binding policy document constrains its behavior over an extended tool-use horizon. We present HANDBOOK_md, a benchmark of 65 agentic tasks modeled on how employees follow company handbooks. Each task places an agent in a self-contained company environment (a file workspace with mock email, chat, calendar, issue-tracking, and commerce services exposed over the Model Context Protocol) and instructs it to carry out routine professional work governed by an expert-written standard operating procedure of 20-124 pages. Tasks span five domains (finance, medical billing, insurance, logistics, and HR) and 10 fictional companies. To resist memorization, every task modifies one of 10 base handbooks, altering the specific rules and thresholds on which grading depends, so no two tasks share the same set of policies. Grading is fully deterministic: each task carries a rubric of programmatic criteria (824 in total) that check both that required actions occurred and that prohibited actions did not. Under strict grading, where a trial passes only if every criterion is satisfied, the strongest evaluated model passes 36.2% of trials, and most frontier models remain below 25%. Failures follow consistent patterns: agents let a plausible but unauthorized in-environment request override the standing policy, perform a required check and then act against its result, lose rule details over long horizons, and report compliance they did not achieve. We release the tasks, environments, and evaluation harness.

arXiv

Do We Really Need Multimodal Emotion Language Models Larger Than 1B Parameters?

Kaiwen Zheng, Junchen Fu, Wenhao Deng, Hu Han, Joemon M. Jose, Xuri Ge

Recent advances in multimodal large language models (MLLMs) have significantly improved the performance of multimodal emotion recognition (MER) and enabled interpretable description generation by jointly modeling video, audio, and language, etc. However, these performance improvements are often accompanied by an increase in model parameter size (e.g, at least 7B), which simultaneously incurs high computational costs and reduces inference efficiency, thereby hindering real-time deployment on resource-constrained platforms such as robots and mobile devices. This raises a fundamental question: do we really need the multimodal MER model larger than 1B parameters for high-quality MER? In this paper, we challenge the assumption that larger models are inherently necessary and proposes a lightweight MER framework (called Light-MER), which achieves better and faster multimodal sentiment understanding and recognition through knowledge distillation. It can transfer knowledge from a strong, large-scale teacher model to a lightweight sub-billion-parameter student model, aiming to preserve rich multimodal emotion reasoning and recognition while substantially improving deployment efficiency. Specifically, we introduce two new optimization strategies to enhance knowledge transfer: (1) a new optimal transport loss that combines Sliced Wasserstein Distance with hidden-state alignment, and (2) a new multi-reward optimization strategy based on GRPO that balances MER performance and efficiency, aimed at further enhancing the learning capabilities of student models. Extensive experiments on nine benchmark datasets demonstrate that Light-MER achieves state-of-the-art performance while significantly improving inference efficiency. This highlights the strong potential of small multimodal emotion language models for future research. Code is available at https://github.com/GAIR-Lab/Light-MER.

arXiv

Transplanting, inverting, and preventing a misalignment persona: method-conditional emergent misalignment in Qwen2.5

Lyndon Drake, Zandi Eberstadt

Emergent misalignment (EM) --- the broad misbehaviour a language model acquires after fine-tuning on narrow harmful data --- is mediated in Qwen2.5 models by a latent persona direction, and that direction is causal in open weights. Transplanting it into a model that shares only pretraining with its source induces broad EM (2.83 $\pm$ 0.26\% misaligned against a random-direction floor of $\sim$1.1\%), and ablating a model's own direction roughly halves an overt inducer's broadcast (21\% to 10\%). The transplant doubles as a measurement method, causally assaying directions that a source model represents but cannot itself express. Whether a fine-tune recruits this persona depends on method and capacity, and since low-rank PEFT is the cheaper regime at scale, the recruiting method is also the economical one. On Qwen2.5-32B, LoRA at low ranks on insecure code recruits it (3.4\% misaligned) while full SFT on identical data does not (0.3\%) and moves against the persona axis (drift--persona cosine $+0.17$ at rank 1 to $-0.10$), the far-inducer, high-capacity exception consistent with a representational-distance $\times$ capacity account. The persona's causal role is itself conditional. Steering a bad-medical SFT run away from the direction during training raises the broadcast from ${\sim}24\%$ to ${\sim}50\%$ while matched random controls stay at or below baseline, replicated across three training seeds, so removing the direction is no blanket recipe. Because recruitment is a loss-reducing shortcut that capacity renders redundant, it can be screened for and prevented in the tested instances. Persona loss-relevance at the SFT solution orders four inducers' broadcasts rank-perfectly within Qwen2.5, inoculation removes recruitment selectively (4.75\% to 0.0\%, code coherence 65\% to 87\%), and fine-tuning orthogonal to the single behaviour-derived axis reduces it persona-specifically.

arXiv

The Illusion of Superposition? A Principled Analysis of Latent Thinking in Language Models

Michael Rizvi-Martel, Guillaume Rabusseau, Marius Mosbach

Latent reasoning via continuous chain-of-thoughts (Latent CoT) has emerged as a promising alternative to discrete CoT reasoning. Operating in continuous space increases expressivity and has been hypothesized to enable superposition: the ability to maintain multiple candidate solutions simultaneously within a single representation. Despite theoretical arguments, it remains unclear whether language models actually leverage superposition when reasoning using latent CoTs. We investigate this question across three regimes: a training-free regime that constructs latent thoughts as convex combinations of token embeddings, a fine-tuned regime where a base model is adapted to produce latent thoughts, and a from-scratch regime where a model is trained entirely with latent thoughts to solve a given task. Using Logit Lens and entity-level probing to analyze internal representations, we find that only models trained from scratch exhibit signs of using superposition. In the training-free and fine-tuned regimes, we find that the superposition either collapses or is not used at all, with models discovering shortcut solutions instead. We argue that this is due to two complementary phenomena: i) pretraining on natural language data biases models to commit to a token in the last layers ii) capacity has a huge effect on which solutions a model favors. Together, our results offer a unified explanation for when and why superposition arises in continuous chain-of-thought reasoning, and identify the conditions under which it collapses.

arXiv

The production of meaning in the processing of natural language

Christopher J. Agostino, Quan Le Thien, Nayan D'Souza, Louis van der Elst

Understanding the fundamental mechanisms governing the production of meaning in the processing of natural language is critical for designing safe, thoughtful, engaging, and empowering human-agent interactions. If meaning is constituted rather than retrieved, then the search for context-independent features or circuits in the pursuit of mechanistic interpretability may be fundamentally limited. Experiments in cognitive science and social psychology have demonstrated that human semantic processing exhibits contextuality more consistent with quantum logical mechanisms than classical Boolean theories, and recent works have found similar results in large language models---in particular, clear violations of the Bell inequality in experiments of contextuality during interpretation of ambiguous expressions. In this work, we explore the CHSH $|S|$ parameter---the metric associated with the inequality---across the inference parameter space of models spanning four orders of magnitude in scale and cross-reference our findings with MMLU, hallucination rate, and nonsense detection benchmarks. We find that the interquartile range of the $|S|$ distribution is completely orthogonal to all external benchmarks, while overall violation rate shows weak anticorrelation with all three benchmarks. We investigate how $|S|$ varies with sampling parameters and word order, and discuss the information-theoretic constraints that genuine contextuality imposes on prompt injection defenses and its human analogue, whereby careful construction and maintenance of social contextuality can be carried out at scale, shaping the space of possible interpretations before any particular one is reached. We consider the implications for mechanistic interpretability and how genuine contextuality sets an information-theoretic bound on the decomposability of semantic processing.

arXiv

Assessing the Effect of Cross-Domain Mapping on Creativity in Humans and Large Language Models

Qiawen Ella Liu, Marina Dubova, Henry Conklin, Takumi Harada, Thomas L. Griffiths

Creativity is the ability to come up with novel ideas, a capacity crucial for human development and flourishing. Are large language models (LLMs) creative in the same way humans are, and can the same interventions increase creativity in both? We study a promising but largely untested intervention for creativity: forcing creators to draw an analogy from a random, remote source domain (''cross-domain mapping''). Human participants and LLMs generated novel designs for ten daily products (e.g., backpack, TV) under two prompts: (i) cross-domain mapping, which required drawing inspiration from a randomly assigned source (e.g., octopus, cactus, GPS), and (ii) user need, which required proposing innovations targeting unmet user needs. We show that humans reliably benefit from randomly assigned cross-domain mappings, while LLMs, on average, generate more original ideas than humans and do not show a statistically significant effect of cross-domain mappings. However, in both systems, the impact of cross-domain mapping increases when the inspiration source becomes more semantically distant from the target, and above a distance threshold, cross-domain mapping benefits the most capable LLMs too. Humans and LLMs differed in how they used the source concept: Humans tended to transfer surface features of the source, whereas LLMs transferred structural and functional properties. Our results highlight the role of remote association in creative ideation and systematic differences in how humans and LLMs respond to the same intervention for creativity.

arXiv

Large Language Models provide support for the parallelogram theory of analogy

Qiawen Ella Liu, Raja Marjieh, Jian-Qiao Zhu, Adele E. Goldberg, Thomas L. Griffiths

Four-term word analogies (A:B::C:D) are classically modeled geometrically as parallelograms: adding the vector B-A+C produces D. Recent work suggests that this model poorly captures how humans produce analogies, with simple local-similarity heuristics often providing a better account (Peterson et al., 2020). But does the parallelogram model fail because it is a bad model of analogical relations, or because people are not very good at generating relation-preserving analogies? We compared human and large language model (LLM) analogy completions on the set of problems from Peterson et al. (2020). We find that LLM-generated analogies are reliably judged as better than human-generated ones, and are also more consistent with parallelograms in a distributional embedding space. Crucially, we show that the improvement over human analogies is driven by greater parallelogram alignment and reduced reliance on accessible words rather than enhanced sensitivity to local similarity. Finally, fine-tuning GloVe to better satisfy the parallelogram constraint makes the model's top-ranked candidates more likely to be the completions humans and LLMs actually produced, and improves its prediction of human ratings. Overall, these results provide support for the parallelogram model of word analogies.

arXiv

Beyond Simulations: What 20,000 Real Conversations Reveal About Mental Health AI Safety

Caitlin A. Stamatis, Jonah Meyerhoff, Richard Zhang, Olivier Tieleman, Matteo Malgaroli, Thomas D. Hull

Mental-health AI safety is typically evaluated with small, simulation-based benchmarks that may not reflect the linguistic and contextual diversity of deployment. We pair four benchmark replications with an ecological audit of real-world conversations to evaluate a purpose-built mental-health AI alongside six frontier general-purpose models spanning four families (OpenAI GPT-5, GPT-5.1, GPT-5.2; DeepSeek V3; Google Gemini 3 Flash; Moonshot Kimi K2). The purpose-built system produced significantly lower overall potentially harmful content rates than every frontier comparator on suicide/self-harm, eating-disorder, and substance-use prompts (CCDH Benchmark: Ash 6.2% vs frontier models 18.0-52.0%, all p < .001). In an audit of 20,000 deployment conversations, clinician review within the audit pipeline confirmed no suicide-risk conversations lacking crisis resources and three NSSI-related conversations without crisis intervention, a within-pipeline conditional rate of 3/800 (0.38%). To extend population-level inference, we adjudicated 600 conversations randomly sampled from the full deployment distribution via blinded clinician review. Against these labels, the LLM judge showed 100% sensitivity (6/6; 95% CI 54.1-100%) and 99.2% specificity (589/594) at its operating threshold, and the conversational model delivered crisis resources in all 6 clinician-confirmed cases (zero end-to-end false negatives; rule-of-three upper 95% CI 0.50%). Findings argue for ecological auditing as a complement to pre-deployment evaluation in mental-health AI safety

arXiv

Meta Prompting for AI Systems

Yifan Zhang, Yang Yuan, Andrew Chi-Chih Yao

We introduce Meta Prompting (MP), a framework that emphasizes the formal structure of a task rather than content-specific worked examples. We give a categorical formalization in which a functor maps typed task transformations to typed prompt transformations. Functoriality encodes preservation of identities and composition, but it does not by itself guarantee semantic correctness. We extend MP to Recursive Meta Prompting (RMP), in which a proposer model generates candidate prompt edits, a validator enforces the edit schema, and an executor model uses the resulting prompt. We model the accumulation of valid edit scripts with the Writer monad and their application by a monoid action on prompts. This construction makes accumulated edits independent of parenthesization; it does not guarantee convergence or improved task accuracy. Empirically, a Qwen-72B base model guided by example-free meta-prompts attains 46.3% pass@1 on MATH and 83.5% accuracy on GSM8K. Separately, an MP-CR agent synthesizes a tool-executed Game of 24 solver; the original experiment reports 100% success on 1,362 instances, subject to exact re-verification from the corresponding input and per-instance outputs. The batch amortizes one LLM response across all puzzles, so its token accounting is not protocol-matched to per-instance few-shot or search-based methods.

arXiv

AURORA-LM: Autoencoding Unified Representation for Continuous-Latent Diffusion Language Modeling

Jiajun Liang, Yucheng Liao, Yukang Cao, Jiazhe Wei, Ken Li, Wende Tan, Jiankun Zhang, ZY Cui, Jingkang Yang, Liucheng Guo, Shiqi Yang, B. Yang, Caifeng Shan, Ziwei Liu, Chenyang Si

Language remains an outlier in generative modeling: while images, video, and audio are increasingly modeled in continuous latent spaces, text generation still relies predominantly on discrete tokens. Existing continuous language models either inherit embedding spaces not designed for joint generation and decoding, or compress autoencoded latents to ease diffusion, sacrificing token-level fidelity. Instead of simplifying the representation to suit the generative model, we preserve a high-capacity, decodable text latent and design the diffusion model to learn its distribution directly. We introduce AURORA-LM, a continuous-latent diffusion language model that separates the construction of a decodable text representation from the modeling of its distribution. A Query-based Encoder-Decoder organizes text into a high-capacity, prefix-aligned latent sequence, and a Block-causal Diffusion Transformer learns its distribution through flow matching, generating blocks left to right while denoising positions within each block in parallel. Because such a latent is harder for diffusion to model, AURORA-LM restricts only the noisy-input pathway while retaining the full clean-latent prediction target, accommodating full-width latents without reducing decoder-facing capacity. We further calibrate the noise-level distribution to the latent width, and introduce self-trajectory consistency to bridge independently sampled training noise and iterative denoising at inference. AURORA-LM achieves the strongest performance among evaluated continuous and diffusion-based language models on OpenWebText free generation and XSum summarization. Scaling to 1B parameters with about 1500 EFLOPs of total compute yields further gains, surpassing a larger publicly released latent-diffusion language model under a matched evaluation protocol. All experiments are conducted on Ascend NPUs.

arXiv

UEmbed: Unified Sparse and Dense Multimodal Embeddings

Tingyu Song, Mingxin Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Zhijie Nie, Yilun Zhao, Shu Wu

Sparse retrieval underpins modern search systems, from web search to retrieval-augmented generation. Existing work has introduced Learned Sparse Retrieval (LSR) to push beyond exact lexical matching toward richer semantics. Yet LSR has so far remained tied to encoder-style bidirectional architectures, and its extension to multimodal settings still relies heavily on auxiliary cross-modal modules. To address these limitations, we introduce UEmbed (Unified Embedding), a decoder-only multimodal embedding model that produces both sparse lexical and dense representations in one causal forward pass. UEmbed appends N learnable special tokens to the input and partitions the vocabulary into N disjoint subsets. Each token's causal hidden state predicts sparse weights over its assigned subset, and the N subsets are concatenated into the full sparse vector. Trained on public data, we release UEmbed at 2B, 4B, and 9B scales. UEmbed-9B reaches 71.8 (dense) and 71.0 (sparse) on MMEB-v2, outperforming multimodal embedding models trained on publicly available data (e.g., RzenEmbed). On BEIR, UEmbed also remains competitive with strong dense and sparse baselines. Furthermore, we demonstrate the practical utility of UEmbed across three dimensions: effectiveness, efficiency, and agentic applications. Overall, UEmbed offers a new paradigm: it unifies dense and sparse embeddings in one model, while further extending sparse retrieval to unify text and multimodal inputs.

arXiv

Romanized Arabic Across Dialects: Views, Usage Patterns, and Linguistic Variation

Amr Keleg, Ahmed Amine Ben Abdallah, Taha Yassine, Chadi Helwe, Imane Guellil, Nedjma Ousidhoum

Arabizi refers to Arabic written in Latin script. Although previous studies have shown that the prevalence and usage of Arabizi vary by factors such as region and age group, most NLP research on Arabic texts treats it as a temporary phenomenon resulting from limited technological support for the Arabic script. In this work, we engage with Arabic speakers to collect insights on their perceptions and usage of Arabizi. We further examine writing norms among speakers of different dialects, focusing on Algerian, Egyptian, Lebanese, Moroccan, and Tunisian Arabic. To this end, we release two resources. First, a character-level alignment of Arabic words to study inter- and intra-dialectal variation across these five dialects, based on words transliterated by survey participants, finding systematic intra-dialectal regularity and inter-dialectal variation. Second, to study Arabic speakers' ability to identify this stylistic variation at the sentence-level, we build a manually curated parallel corpus of sentences written in Arabic script alongside multiple Arabizi transliterations, collected from speakers of the same five dialects. Our study presents the largest human-centered, cross-dialectal study of Arabizi's perceptions and practices to date.

arXiv

Who Should Be Generated? Justifying Demographic Targets in Open-Ended Generation

Zeshen Zheng, Yujia He, Qianmian Lin, Xiangyue Huang, Wenqing Chen

Fairness evaluation concerns not only what a model produces, but also what its outputs ought to be compared against. When a model generates "a CEO in the United States," the prompt leaves demographic realization to the model. Existing group fairness definitions assume that sensitive attributes are given on the input side. Generative audits instead examine output-side demographic composition, yet the targets they compare it against are typically supplied rather than justified. The upstream question is what the target distribution should be. We formalize this missing-target problem for demographic-value-unspecified generation and decompose target construction into four commitments: the evaluative object, prior admissibility, allocation, and operationalization. In this framework, we admit the geographic prior under a geographic-membership interpretation for the declared public-world use. The occupational prior, under an incumbency interpretation, requires an independently defended objective such as workforce-composition fidelity. Instantiating this construction in AP-Bench, we find substantial distribution divergence from geography-derived targets, ranging from 0.508 to 0.606 on a 0-to-1 scale. Replacing each geography-derived target with an equal-category comparator, while holding generations and measurement fixed, produces model-specific mean absolute cell-level $\mathrm{JSD}_2$ changes ranging from 0.279 to 0.355. Target construction is therefore not a preliminary to fairness evaluation but a component of it. What we supply is not a universal target, but a framework that makes explicit the justification required before a distribution can serve as a fairness standard.

arXiv

MedPRESS: A Multi-turn Benchmark for Patient-Pressure-Induced Medical Sycophancy in LLMs

Saman Sarker Joy, Niloy Farhan

Large language models (LLMs) are increasingly used for health-related advice. Existing research measures their safety with static questions rather than pressured patient-facing conversations. We introduce MedPRESS, a multi-turn benchmark for measuring patient-pressure-induced sycophancy in LLMs. MedPRESS contains 600 medically grounded five-turn dialogues across three scenario families: medication and treatment demand, personal health self-care, and symptom triage and care resistance. Each dialogue begins with a health query and escalates through personal experience, social proof, external evidence claims, and direct adversarial challenge. We evaluate 20 LLMs across general, medical-domain, lightweight, large, open-weight, and proprietary families using structured judging and safety-focused metrics. Results show that models frequently shift toward unsafe agreement under repeated patient pressure, with substantial variation across model families, model scale, and prompt type. Anti-sycophancy prompting improves robustness for several models, but does not eliminate unsafe agreement. MedPRESS highlights a critical gap in medical LLM evaluation: safe medical knowledge is not enough unless models can maintain it under conversational pressure.

arXiv

SWE-Touch: Benchmarking Coding Agents When Users Touch the Code

Yuqiao Tan, Jinxiang Meng, Fangyu Lei, Minzheng Wang, Shizhu He, Jun Zhao, Kang Liu

Real-world software development requires coding agents to operate in shared workspaces where users may inspect and modify code during an ongoing task, yet existing repository-level benchmarks typically evaluate agents working alone or restrict user participation to messages. This leads us to ask: how do coding agents understand and respond to code changes in a shared workspace? We introduce SWE-Touch, a framework that stress-tests this setting through validated Counter-Edits: plausible edits to task-relevant code that conflict with task completion. SWE-Touch mines task-critical regions from multiple repair trajectories, uses a separate User Patch Generator to construct the edits, and injects them with contextual user messages when agents reach the relevant code. We evaluate nine coding models on SWE-bench Verified, with additional experiments on longer-horizon tasks from SWE-Bench Pro and DeepSWE. Counter-Edit lowers average resolve rate by 7.7 percentage points on SWE-bench Verified, with degradation also persisting on both longer-horizon benchmarks. Trajectory analysis links these failures to limited awareness of the evolving workspace: agents may retain conflicting code or replace it without sufficiently re-inspecting the repository and validating the revised code with targeted tests. These findings show that strong autonomous performance does not yet ensure the state awareness and adaptive behavior needed for shared-workspace collaboration, and point to detecting workspace changes, reconciling conflicting edits with the task, and verifying the affected behavior as key capabilities for future optimization.

arXiv

Cultural Awareness is Represented but Not Decoded: Tracing Mythological Knowledge across 18 Open-Source LLMs

Iaroslav Chelombitko, Ekaterina Chelombitko, Mika Hämäläinen

Open-source LLMs reliably name Zeus, Jupiter, and Thor, but recover their counterparts in less-represented traditions like Finnish, Slavic, Egyptian, or Chinese mythology far less consistently. We ask where inside the model this cultural default is produced. On a parallel cross-cultural substrate of Thompson-motif entities, we instrument 18 open-source LLMs from 8 architecture families with linear probing, logit lens, activation patching, and output extraction. The residual stream cleanly distinguishes cultures, well above a name-string baseline, yet the decoder collapses culturally-specific tokens onto dominant-tradition ones. The failure is at readout, not at representation. Asking the same question in the target culture's native language versus English produces failures that cluster within language but decouple across language: the decoder is gated on prompt language. We release a per-entity (probe, output) decomposition framework, a citation-anchored cross-cultural ground truth, a within- versus cross-mode correlation test for language-conditioned readout, and per-entity predictions for all 18 models.

arXiv

CTRAG: An In-Context Retrieval-based Framework for Automated Compliance Checking using LLMs

Muhammad Roman, Karen Rafferty, Barry Devereux

Trust is fundamental in modern regulatory ecosystems, and compliance checking plays a critical role in fostering that trust. Regulatory compliance verification is essential for businesses operating in highly controlled environments, as it ensures alignment with sector-specific guidelines across domains such as financial reporting, data privacy, and cybersecurity. Manual compliance testing, however, is often time-intensive and prone to inconsistencies, particularly when compliance depends indirectly on third-party services such as cloud providers, where vendors rely on external providers to meet regulatory standards. In this paper, we present CTRAG, a novel Retrieval-Augmented Generation (RAG) pipeline designed for automated compliance checking. CTRAG employs advanced strategies, including adaptive chunking, dynamic retrieval configurations, and in-context learning, to improve the precision and relevance of compliance assessments. By extracting control questions from regulatory texts and cross-referencing them with unstructured company documentation, CTRAG achieves highly accurate, document-informed compliance verification, even in cases of indirect compliance through third-party services. Empirical evaluations demonstrate significant improvements, with CTRAG achieving an F1-score of 78% and a recall of 85% in the final deployed configuration, ensuring minimal missed non-compliance cases while reducing manual reviewer effort in a real-world deployment. To validate CTRAG value, we developed and deployed a POC within a Big Four professional services firm, applying it to real-world cases and cross-checking results against manual compliance reports. These findings highlight CTRAG potential to streamline compliance workflows, mitigate risks, and enhance regulatory trust in complex, high-stakes environments.

arXiv

Right Answer, Wrong Method: Shortcut Hacking Misleads the Evaluation of LLM Reasoning on Frontier Science Benchmarks

Xuan Ren, Weiqi Zhai, Tianle Pu, Yihua Zhu, Hu Wei, Bing Zhao

Scientific reasoning benchmarks typically evaluate large language models (LLMs) using final-answer accuracy. However, a correct answer does not necessarily demonstrate the reasoning capability targeted by the problem. We identify Solution Hacking, a failure mode in which an LLM reaches the correct answer through invalid shortcuts, such as numerical search, enumeration, guessing, or answer-first verification, without providing a valid task-targeted derivation. We systematically analyze this phenomenon across difficulty levels, scientific domains, and frontier models. Solution hacking increases sharply with benchmark difficulty, from 2.2\% on common problems to 28.3\% on Olympiad-level problems and 37.4\% on HLE. Moreover, 8.2\%-44.1\% of answers credited as correct across frontier models are identified as hacked solutions. We further develop expert-inspired anti-hacking strategies, including an automatic judge and a test-time instruction. The results show that suppressing shortcut behavior substantially reduces reported accuracy while having a smaller effect on correct and non-hacked accuracy. These findings reveal that answer-only evaluation can overestimate the scientific reasoning capabilities of frontier LLMs.

arXiv

Training-Free versus Training-Based Intent Classification in LLMs: Accuracy, Robustness, and Failure Modes

Nan Chen, Zhouhao Yang, Soufiane Hayou

Intent classification in Large Language Models (LLMs) involves categorizing user prompts into predefined classes. For instance, given a user prompt, the system must determine whether it primarily concerns mathematics, coding, or general text processing. Such classification enables routing prompts to specialized models optimized for specific domains, improving both accuracy and computational efficiency. In this work, we conduct a systematic study comparing training-free vs training-based approaches for intent classification. For this purpose, we consider two lightweight, training-free methods based on statistics of internal representations and compare them against MLP classifiers and linear probes. Our comprehensive empirical evaluation reveals that 1) Both training-free and training-based methods saturate easy benchmarks (mathematics vs. coding vs. natural language), 2) Training-based classifiers have an advantage on harder classification tasks (e.g. Java vs Python), and 3) Training-free methods are generally more robust to mixed-intent and adversarial prompts.

arXiv

PredAct-Bench: Benchmarking Tool-Augmented Dialogue under Controlled Tool Noise

Abdulrahman AlRabah, Xiaocheng Yang, Dilek Hakkani-Tür, Abdussalam Alawini

Large Language Models (LLMs) are increasingly deployed in task-oriented dialogue systems that support multi-step decision-making in high-stakes domains such as education, healthcare, and finance. However, existing benchmarks typically assume perfectly accurate tool outputs, overlooking the reality that deployed systems must operate with noisy tools and human decision-makers whose trust in the agent is itself uncertain. Such conditions are common in practice, for example, a clinician using a diagnostic prediction tool or an advisor relying on a model that forecasts student outcomes from historical records. We introduce PREDACTBENCH, a benchmark for evaluating dialogue agents paired with statistically imperfect tools, using education as a measurable testbed where ground truth outcomes and clear intervention decisions are available. First, we build a benchmark for AI-assisted human decision-making, where the AI uses noisy predictors to help guide a user. Second, we introduce episode-level Relative AI-Reliance (RAIR) and Relative self-reliance (RSR) metrics, extending prior trust calibration framework to multi-turn dialogue. Third, we evaluate 13 state-of-the-art closed and open source LLMs on two educational datasets, OULAD (real assessment trajectories from the UK Open University) and PREDACT-CS (60 courses with real final grade outcomes and synthetically generated weekly score trajectories), alongside a human study with instructors and teaching assistants. We find that when tools are noisy, SOTA models are supposed to provide visibility to teachers so that they do not over-rely on wrong suggestions or hallucinations, but current models fail to do that. We offer PREDACTBENCH to help build better LLMs as AI decision support systems to help teachers.

Keep browsing

More papers from top AI conferences

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