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

You Only Pass Once: Answering and Abstaining Together in a Single Forward Pass of a Frozen Language Model

Ziyang Luo, Zhongyao Chu, Xinjie He, Youting Wang, Xukui Qin, Runxiong Wu, Yan-Syuan Chen

A frozen language model on reasoning tasks has two coupled weaknesses: it under-uses evidence its own residual stream already encodes, and it fails to detect when the input is insufficient to answer, so it confabulates. This paper consolidates two research lines that address these on the same residual stream: a conditional steering probe writes the stream at mid-stack layers and recovers reasoning accuracy from a frozen backbone, and a zero-shot sufficiency direction reads the stream and abstains when information is insufficient. Deployed in one forward pass they interfere: the steering write shifts the state the direction reads, costing up to 8 AUROC points of cross-domain transfer on small models; a separate clean pass doubles inference cost. We keep the direction fixed and train a small network to reconstruct the pre-steering residual from the steered one -- mean-squared error on (steered, clean) pairs, no sufficiency labels -- and read the direction on the reconstruction. The resulting system, YOPO (You Only Pass Once), answers, steers, and abstains in one forward pass of a frozen Qwen2.5 backbone (1.5B/3B/7B). End to end, three-way accuracy more than doubles the frozen baseline (0.375->0.798 on 1.5B alphaNLI) and one pass beats the two-pass reference at every scale (0.798/0.830/0.893 vs 0.753/0.790/0.863) and on ten backbones across six model families. We chart the capacity-transfer frontier quantifying the principle that abstention should not be trained in; a source-side audit catches our own alphaNLI construction leaking a surface artifact, so architectural claims are anchored on native-label replications (SQuAD2, RepLiQA, MuSiQue); and on the standard four-domain suite we contribute, to our knowledge, the first answer-or-abstain benchmark, where our gate tops every in-domain dataset and the label-free direction is the only gate family to survive domain transfer.

arXiv

Information Satisfaction: A Reader-Centered Axis for Summarization Evaluation

Isabel Cachola, William Walden, Reno Kriz, Mark Dredze

The majority of work on summarization evaluation focuses on general summary quality (e.g., ROUGE, BERTScore) or specific desired properties (e.g., readability, factuality). However, these metrics fail to measure the utility of a summary to an individual user. For example, a biomedical researcher learning about the latest vaccine research will have different informational needs from a family doctor. Query-focused summarization captures part of this need, but in practice, users rarely state everything relevant in a query: a single short query is likely inadequate to distinguish the needs of a researcher from those of a physician. By contrast, a reader's background or persona (their role and expertise) is comparatively stable across queries and recovers much of this missing context, which makes it a practical signal for assessing whether a summary satisfies that reader's needs. In this work, we assess how sensitive popular summarization metrics are to both informational and persona differences, and find that many popular metrics, including strong LLM-as-judge metrics, fail basic perturbation tests of informational content. We additionally conduct an expert human evaluation, measuring summary preferences based on information satisfaction given a specific person's background and use case. We find that both traditional and LLM-based metrics are insufficient measures of information satisfaction and agree poorly with human judgment.

arXiv

Whose doctor does the AI recommend? An algorithm audit of reputation and demographic signals in large language model-assisted physician choice

Syeda Anshrah Gillani, Mirza Samad Ahmed Baig

Patients increasingly ask large language model (LLM) assistants which doctor to see, making these systems AI infomediaries: algorithms that intermediate one person's choice among other people and thereby decide, silently and at scale, which physicians become visible. We report a prespecified randomized algorithm audit of what causally moves those recommendations. Seven models (six open-weight; gpt-4o-mini) each chose among five synthetic family-medicine physician cards whose attributes were independently randomized across 3,024 choice sets, three patient personas, nine prompt paraphrases and nine experimental arms, yielding 40,068 scored responses; gender and ethnicity were signaled through names following correspondence-audit methodology. Reputation signals dominate: raising a rating from 3.9 to 4.7 increases choice probability by 31.4 percentage points (pp), and raising the fee from $90 to $190 lowers it by 20.0 pp. Demographic parity is rejected, but not in the direction human audit studies predict: female-signaled names gain 2.5 pp, and Hispanic-, South-Asian- and Black-signaled names gain 1.3-2.9 pp over White-signaled names, tilts worth $7-$14 per visit in fee-equivalent terms, and a content-free first-listed position is worth $11. Yet models mentioned gender or ethnicity in at most 0.03% of their stated reasons and abstained in 0.39% of trials, so these effects are invisible in the models' own explanations, and transparency obligations relying on model self-report would not detect them. One reasoning model failed the prespecified auditability gate outright. The frozen design makes the audit repeatable: any new model can be assessed against identical stimuli, making recurring behavioural audit, rather than self-reported explanation, the monitoring technology fit for purpose.

arXiv

LLMs Don't Pay for the Jump

Paras Balani, Subhrakanta Panda

Zahavy [2026] argues that Large Language Models, despite their capabilities in induction and deduction, cannot perform the abductive "Jump" that produced Einstein's equivalence principle, and attributes this limitation to the absence of embodied simulation. Zheng-Xin [2026] and Farmer [2026] question whether embodiment is necessary for abduction, pointing to alternative routes to General Relativity and forms of abduction that require no sensorimotor grounding. Max Planck resolved the blackbody radiation problem in 1900. Planck's move to E = hν required no embodied simulation. It was motivated by a mathematical consequence of classical theory, an infinite predicted energy for a finite measured quantity, that could not be physically accepted. We show that neither induction nor deduction could have produced the postulate and argue that its adoption required a coupling between epistemic error and physical cost. We formalize this distinction through thermodynamic coupling and show that fixed-weight transformer inference lacks such coupling, regardless of model scale. This is consistent with empirical results showing that output entropy remains nearly unchanged across tasks with sharply increasing causal difficulty, even as accuracy falls from 100% to 17%. We therefore argue that the missing ingredient in machine abduction may lie deeper than embodiment: a system must have some physical mechanism through which epistemic error becomes costly enough to force revision.

arXiv

A Survey of Large Models in Sports

Yichen Xu, Jianzhe Ma, Chuhan Wang, Zhonghao Cao, Liangyu Chen, Wenxuan Wang, Qin Jin

Sports have witnessed growing global enthusiasm in recent years, serving as a vital force for physical health, cultural exchange, social connection, and economic growth. The rapid advancement of large models, particularly (multimodal) large language models (M)LLMs, has demonstrated transformative potential to reshape sports understanding, analysis, and interaction across diverse domains. This paper presents a comprehensive survey of large models in sports, including (i) an overview of tasks and applications across different participant groups; (ii) a detailed analysis of sports-related datasets and benchmarks; and (iii) a critical discussion of current challenges and future directions. Our goal is to establish a foundation for advancing research and practical development of large-model-driven sports intelligence. An open-source GitHub repository is maintained at: https://github.com/Road2Redemption/Awesome_Large_Models_In_Sports1.

arXiv

Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages

Chih-Hsuan Yang, Anjir Ahmed Chowdhury, Cheng-Hau Yang, Weijian Zheng, Fernando Llorente, Xiaolong Ma, Xinyang Li, Eliu A. Huerta, Ian T. Foster, Rajeev Thakur

Multi-agent reasoning systems often use agreement, confidence, or automated scores to decide which messages should shape a final answer. Such filtering assumes that a message likely to be correct is also worth keeping. Yet a wrong answer can contain a useful decomposition, constraint, or scientific principle. We test this distinction with Diverse Hypothesis Deliberation (DHD), a controlled measurement protocol that caches five independently generated messages and replays the same downstream solver, called the integrator, with each message available or hidden. The replay comparison measures a message's trajectory value: whether making the message available helps or harms subsequent reasoning. Across five mathematics and science benchmarks and two openly available model families, gpt-oss-120b and gemma-4-31B-it, wrong-helpful messages appear in every benchmark-model combination. Among wrong-answer messages that change final correctness, more than four in ten changes are helpful in each model. Controlled repeats show that the number of repeatable message effects is unlikely to arise from replay variation alone (p=0.0002). A focused intervention on repeatable wrong-helpful messages finds that the complete message works best, while retaining its reasoning preserves more success than retaining only its answer; the source of the complete-message advantage remains open. Within the same problem, repeated trajectory-value evidence also identifies a better keep-or-remove choice than answer correctness alone. Answer correctness is therefore informative but does not determine trajectory value. DHD measures this missing property and produces reusable labels for learning when agents should listen.

arXiv

Local and Global Regimes of Geometric Complexity in Language Model Representations

Arwa Osman, Marco Baroni, Iuri Macocco

Intrinsic dimensionality (ID) is widely used to probe the representational complexity of language models, but it remains unclear whether ID differences reflect properties of language itself or artefacts of how the underlying dataset was constructed. In this paper, we focus specifically on how lexical diversity, the number of unique last-token items present in a dataset, affects ID estimates of that dataset. We find a scale-dependent transition between two regimes: at low lexical diversity, conditions with fewer unique final words produce higher ID, while at high lexical diversity, this ordering reverses, and conditions with more unique words produce higher ID. We derive an exact, parameter-free formula for the point at which this reversal occurs, which matches the observed transition point at every scale tested. On the one hand, our results highlight how care must be taken when interpreting the intrinsic dimensionality of a set of representations as a straightforward cue of their complexity. On the other hand, our discovery of the two ID regimes reveals a general principle of organisation of linguistic data in LLMs that sheds new light on their inner manifold structures.

arXiv

A Four-Axis Trustworthiness Benchmark for LLM-as-Judge in Principle-Based Regulation

Dipankar Sarkar

Principle-based regulation, with evaluative standards such as "fair, clear, and not misleading" or "deliver good outcomes", cannot be reduced to binary predicates, and LLM-as-judge is increasingly used as the substitute. Our position is that any such judge must be evaluated on four axes: accuracy, paraphrase robustness, adversarial robustness, and calibration. We release Principle-Bench, 168 cryptoasset financial-promotion scenarios mapped to two UK FCA principles, with paraphrase, adversarial keyword-stuffing, and boundary perturbations authored under a pre-registered rubric; the first benchmark covering all four axes for principle-based regulation. We also introduce Ceca (Calibrated Exemplar-Cluster Assessment): a calibrated, auditable assessor that emits exact per-exemplar counterfactual attributions. Across keyword counting, three sentence-transformer embedders, an open-weight LLM-judge, and a calibrated cascade, no method dominates all four axes. A 120B LLM-judge, strongest on benign inputs, loses 47 accuracy points (0.74 to 0.27) on keyword-stuffed Consumer Duty inputs: "compliance theatre." A second judge from a different model family agrees only at Cohen's kappa = 0.16 on that split, localising the failure to the model rather than the corpus. Any deployment-grade LLM-judge for principle-based regulation must report per-principle adversarial deception and post-hoc calibration alongside aggregate accuracy.

arXiv

AnchorBench: A Multi-Pathway Benchmark for the Anchoring Effect in LLMs

Yiderigun Borjigin, Alexander Hermann, Christian Cyron, Roland Aydin

The anchoring effect is a cognitive bias in which an initial reference value shifts a later judgment toward itself. This effect is well established in human judgment and decision-making, and recent work suggests that large language models (LLMs) exhibit similar behavior. However, existing work on anchoring in LLMs typically evaluates only a narrow set of anchor pathways and rarely distinguishes irrelevant from plausible anchors. We introduce AnchorBench, a benchmark for the anchoring effect in LLMs that evaluates multiple anchor pathways under an explicit anchor relevance axis. Across fourteen models, including ten open-weight models and four frontier API models, and a large set of controlled prompts, we find that (1) anchoring is strongly pathway-dependent, (2) plausible anchors usually induce larger shifts than irrelevant ones when introduced through stronger pathways, (3) anchor influence generally weakens as the anchor moves farther from the evidence-supported answer, most clearly on External and RAG, and (4) high task accuracy on the anchor-free control condition (Acc$_{10}$: answers within 10 points of gold) does not guarantee robustness: even frontier API models above 95% control accuracy remain susceptible to plausible anchors.

arXiv

Envs-FORGE: Frontier-Optimized Reward-Grounded Environment Synthesis for Agent RL

Xiaojun Wu, Cehao Yang, Honghao Liu, Xueyuan Lin, Zhichao Shi, Hao Zhou, Xuhui Jiang, Chengjin Xu, Jia Li, Jian Guo

Reinforcement learning (RL) for terminal agents needs executable training environments with reliable rewards and useful difficulty. Fixed recipes such as few-shot, Self-Instruct, and Evol-Instruct apply the same prompting policy to every seed, even when the current policy would benefit from a harder, easier, or simply different task. We present Envs-FORGE, a prompting policy that converts verifier rewards into per-seed environment-synthesis actions. Envs-FORGE estimates seed pass rates, scores six projection--direction actions around a target learning frontier, and solves a per-seed mixed-integer linear program (MILP) to choose the action that conditions generation. The selected action drives synchronized rewriting of the instruction, fixtures, oracle solution, tests, and Docker environment; only gold-verified bundles enter RL training. The indexed MILP form also supports optional soft skill coverage for portfolio planning. On Qwen 3.5 35B, Envs-FORGE improves Pass@1 over Base by 9.2 percentage points on tb-core (40.0% to 49.2%) and 6.4 points on tb-2.0 (23.0% to 29.4%), exceeding the strongest fixed-recipe baseline by 2.4 and 2.1 points. It reaches 77.1% on SWE-bench Verified versus 73.4% for Base, and improves tb-core by 6.8--9.2 points across the evaluated 4B--35B models. All synthesis methods export 100 verified environments and use 2.27M--2.88M synthesis tokens, placing the comparison at the same downstream training-set size and the same operational scale. The source code is available at https://github.com/DataArcTech/DataArc-SynData-Toolkit/.

arXiv

Seeing Red, Thinking Bad: Color Bias in Vision Language Models

Kohsuke Ide, Ryousuke Yamada, Yoshihiro Fukuhara, Hirokatsu Kataoka, Yutaka Satoh

Vision language models (VLMs) are increasingly used in industrial decision-making systems, such as recruitment support and recommendation. This motivates careful analysis of how VLMs process visual and textual information. In this work, we study how VLMs interpret text rendered as an image, and investigate the influence of visual styling biases. To this end, we introduce Stealth Visual Prompts, which subtly change visual styling of text, such as color and contrast, while preserving semantic content. Using these prompts, we systematically control the visual styling of words in text and measure their impact on the analysis performed by VLMs. We further analyze how such visual perturbations affect the latent representations of the vision encoder. From our experiments, we observed that coloring positive words in green consistently shifts sentiment predictions toward a positive direction. As a result, VLMs often fail to properly account for negative words present in the text. Our analysis suggests that this behavior is correlated with changes in the latent representations of the vision encoder induced by color variations. In addition, we show that reducing text--background contrast increases reliance on visually salient cues and leads to more incorrect Visual Question Answering (VQA) outputs. These results suggest that the visual styling of rendered text can guide VLMs' interpretation in ways that diverge from human semantic understanding. Project page: https://github.com/KohsukeIde/color-bias-vlm

arXiv

SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoning

Haonan He, Haodi Lei, Yun Luo, Haoran Zhang, Shunkai Zhang, Yizhuo Li, Shengji Tang, Zhilin Wang, Runzhe Zhan, Lei Bai, Ganqu Cui, Fangchen Yu, Yafu Li, Peng Ye, Ning Ding, Yu Cheng

On-policy distillation (OPD) offers a promising way to transfer reasoning capabilities from stronger teacher models, but applying it to long-context reasoning teachers and short-context students introduces practical challenges, including tokenizer mismatch, teacher-student distribution mismatch, response length explosion, and training instability. In this work, we study this setting by transferring proof-reasoning capabilities from the long-context reasoning model SU-01 to short-context student models. To handle tokenizer differences, we perform OPD in a shared text space and align only tokens that occupy identical text spans under the student and teacher tokenizers. To mitigate the problem of excessive generation length and frequent truncation, we introduce a student reference KL loss and mask the advantages of special termination tokens such as </think> and <|im_end|>. This strategy constrains the student from drifting excessively from its initial policy, thereby mitigating the teacher-student distribution mismatch problem and fostering steady length growth. Experiments on both same-family and different-family student models, including Qwen3, Qwen3.5, Intern-S2, GLM-4.7, Gemma-4, show consistent gains in mathematical reasoning, especially natural-language math proving. Notably, Intern-S2-Preview improves by 21.2 points on ProofBench, reaching 55.2 and surpassing Gemini-2.5-Pro. It also improves on science benchmarks such as HLE and HiPhO, suggesting that OPD transfers reasoning capabilities that generalize beyond the mathematical training domain.

arXiv

Grounding Without Corrective Control: Truth-Tracking Profiles for Large Language Models

Brett Reynolds

Recent work suggests that some large language model representations have content or reference. Grounding can secure either without supplying live routes for correction. This paper asks what follows from that gap. An output is answerable when discrepancies can affect what a target- and task-specific arrangement produces, accepts, or withdraws. The arrangement has corrective control only when live, sufficiently independent routes can detect and repair fresh discrepancies. A route profile records which routes constrain the arrangement and how they are related. Those profiles support analysis of truth-tracking: patterned support for representational success. Language models are the pressure case; text-only arrangements provide a task-relative limiting case. Text-trained models inherit patterns of testimony, coherence, and prior correction. Where target-sensitive correction survives training, these can supply derivative answerability (inherited constraint); live answerability is the relation supplied by a current route for fresh discrepancies. Fluent failures should follow when a task requires independently informative access to the facts. Self-consistency, retrieval, tools, code execution, multimodal input, and feedback should help selectively. Route-by-task interactions test the distinctions. The decomposition's empirical burden is to predict held-out route--task combinations or improve intervention choice without conceptual refitting. Surface improvement and truth-tracking improvement can come apart.

arXiv

The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning

Anna Borisiuk, Andrey Savchenko, Alexander Panchenko, Elena Tutubalina

Popular facts are memorised more deeply during pretraining and resist removal longer than rare ones, yet existing LLM unlearning methods apply uniform gradient pressure regardless of training-data frequency. We propose the AdaPop (Adaptive Popularity) method, which combines local token confidence with a per-fact popularity-dependent exponent derived from an external proxy (e.g., Wikidata sitelinks, LLM-as-Judge), and automates the forget-retain balance via a dual-ascent controller that adjusts the retain penalty each epoch. Across three model families and two benchmarks, AdaPop leaks ~5x less forgotten content than competing methods under paraphrased queries and ~1.6x less under adversarial reformulations. We support our analysis with internal metrics: under our method, forget-set hidden states move further from the pre-unlearning model's states than under other methods, while retain-set representations remain close.

arXiv

MathForm: Scaling Mathematical Autoformalization with Knowledge Retrieval and Verification-Guided Refinement

Lushi Pu, Weiming Zhang, Xinheng Xie, Zixuan Fu, Bingxiang He, Hengyu Zhao, Hongya Lyu, Xin Li, Jie Zhou, Yudong Wang

Autoformalization is commonly framed as translating natural-language mathematical statements into machine-verifiable formal languages such as Lean 4. However, faithful formalization requires more than translation. Models must map mathematical concepts to the complex hierarchy of types and definitions in formal libraries such as Mathlib, while ensuring that generated statements preserve the meaning of the source propositions. Existing approaches struggle because they rely heavily on the model's parametric memory for library-specific knowledge, while common data construction pipelines often resort to filtering single-pass outputs and lack mechanisms for feedback-driven revision. To address these challenges, we introduce MathForm, an autoformalization framework for constructing verified training data through Mathlib knowledge retrieval and verification-guided iterative refinement. Before generation, a retrieval planner gathers relevant definitions and existing formalizations from Mathlib to guide the formalization generator. Generated statements are then revised using compiler diagnostics and semantic-consistency feedback. Using this framework, we construct FormalVerse, a Lean 4 dataset containing approximately 367K verified examples across diverse mathematical domains and sources. We then train MathForm-8B through supervised fine-tuning followed by reinforcement learning. Across six benchmarks, MathForm-8B achieves average Pass@8 rates of 88.06% under Syntax Check (SC) and 72.37% under Consistency Check (CC), outperforming multiple specialized 32B autoformalizers. On the challenging FATE-H and FATE-X subsets, it attains CC pass rates of 63% and 37%, exceeding the strongest specialized baselines in both cases.

arXiv

How Much Do Legal RAG Systems Still Hallucinate?

Souvick Das, Sallam Abualhaija, Domenico Bianculli

Hallucination is a major challenge for retrieval-augmented generation (RAG) systems in the legal domain, where ungrounded answers can lead to serious consequences. To better understand this problem, we conduct a fine-grained analysis of hallucination behavior in eight legal RAG systems across two legal corpora, the GDPR (in English) and a national civil law (in French). Using claim-level and answer-level evaluation, we report on hallucination density and severity, analyze performance across question categories and user personas, and validate our findings on an independent set of 142 legal-expert-authored questions. Our results show that hallucinations remain pervasive, ranging from less than 10% of responses for the best-performing systems to nearly half in the worst case. We further find that false-premise questions, containing incorrect assumptions that must be rejected, produce high hallucination rates on the manually-drafted questions.

arXiv

MINT: A Universal Zero-Shot Predictor for Transaction Data

Parameswaran Kamalaruban, Viktor Drobnyi, Maeve Madigan, Julia Rozanova, David Sutton, Stuart Burrell

Banks analyse sequential financial transaction data to perform many tasks, including fraud prevention, credit risk assessment and offer personalization. To improve the predictive accuracy of these tasks, Payments Foundation Models encode transaction sequence data as rich contextual embeddings, which can then be provided to task-specific models as features. However, these Foundation Models are not designed for flexible zero-shot reasoning across novel downstream prediction tasks, limiting their adaptability and utility. Existing LLM-based approaches to zero-shot prediction often fail to fully exploit the predictive signal within transaction data, while relying on costly text serialization or task-specific architectures that scale poorly. To address these limitations, we present the Multimodal Instruction Network for Transactions (MINT), a framework that connects a pretrained transaction sequence encoder to a decoder-only LLM through lightweight embedding injection, transaction-language alignment, and instruction tuning. We find that MINT achieves state-of-the-art predictive question-answering performance in both in-distribution and out-of-distribution questions, while substantially reducing input tokens, latency, and memory consumption compared to text-serialization baselines. Through comprehensive analyses of representations, alignment strategies, training data, and history length, we establish that compact transaction embeddings are a superior approach to transaction representation than text serialization for multimodal reasoning and zero-shot prediction tasks.

arXiv

KV Cache Compression Through the Lens of Transform Coding

Hannah Laus, Claudio Mayrink Verdun, Hao Wang, Flavio du Pin Calmon, Felix Krahmer

The key-value (KV) cache stores information from past tokens and is a major memory bottleneck in long-context inference. Existing quantization methods address this bottleneck by representing the KV cache uniformly with lower-precision data types and designing quantization schemes to minimize reconstruction error in the cache itself, without accounting for how that error propagates through attention mechanisms. We prove that, under a white-noise quantization model, the expected attention-aware distortion decomposes into additive key and value contributions that factor across tokens and channels. Building on transform coding and reverse water-filling, which are classical tools from signal processing and rate-distortion theory, we introduce Attention-Aware Transform Coding (AATC), which allocates bits over a calibration set to minimize attention-aware distortion. On Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct, evaluated across LongBench, RULER, GSM8K, MMLU-Pro, and MATH-500, our method achieves near-lossless accuracy at approximately $5.8\times$ compression, whereas each baseline degrades in at least some settings.

arXiv

Leading-Silence Augmentation and Multi-Stage Synthetic Supervision for the Second MLC-SLM Challenge

Kexin Shi, Renhe Sun, Yuge Huang, Ximeng Wang, Jiayi Zhou, Jian Liu, Malu Zhang

The second Multilingual Conversational Speech Language Model (MLC-SLM) Challenge evaluates two tasks over complete, unsegmented multilingual conversations: speaker diarization and recognition (Task 1) and conversational speech understanding (Task 2). Neither task provides oracle utterance boundaries or speaker labels at evaluation, and Task 2 provides no question-answer training set. For Task 1, we fine-tune VibeVoice-ASR-7B with random leading-silence cropping, consistent timestamp correction, and an exponential moving average (EMA) training strategy. For Task 2, we construct synthetic question-answer pairs through multimodal candidate generation, silent-audio filtering, and distribution-matched augmentation, and fine-tune Qwen3-Omni-30B-A3B-Instruct for tagged direct answering. On the Task 1 evaluation set, cropping reduces tcpMER from 18.30% to 17.27%, and EMA further reduces it to 16.73%. On the Task 2 evaluation set, jointly applying distribution-matched augmentation and tagged direct answering raises accuracy from 83.0% to 86.0%.

arXiv

Regime-Conditional Verification: Correctness Estimation for Adapting and Monitoring Safety Classifiers

Thiago Sandoval, Ufuk Topcu

Safety classifiers deployed with large language models often fail for two reasons: their decisions reflect the policy learned during training rather than the deployer's desired policy, and their performance degrades as deployment traffic evolves. We present Regime-Conditional Verification (RCV), a lightweight wrapper that adapts an off-the-shelf safety classifier without retraining it. RCV estimates, from the classifier's internal representations, the probability that each prediction disagrees with the deployer's policy, and selectively corrects predictions likely to be wrong. The same correctness estimates also provide a label-free signal for detecting distribution shift, enabling a maintenance loop that updates the correctness estimation layer and resorts to classifier fine-tuning only when necessary. Across three off-the-shelf safety classifiers and two benchmark datasets, RCV improves adherence to the deployer's policy in every classifier-dataset combination, catching up to 0.81 of previously missed unsafe content without modifying the underlying classifier. In a deployment study with ten attack campaigns, each a harm category held out of RCV's training, RCV detects every campaign in a dedicated injection panel; in the maintenance census most drift episodes are repaired without updating the classifier, and the fine-tune is reserved for the residual episodes that repair does not restore.

Keep browsing

More papers from top AI conferences

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