paperlist.ai

AI 顶会论文

AI 顶会论文库

检索 30+ 个顶会的论文，包括 NeurIPS、ICLR、ICML、CVPR、ACL、AAAI 等。输入关键词筛选，打开任意一篇查看摘要、作者和链接。

arXiv

CAPO: Constraint-Aware Prompt Optimization for LLM Agents

Victor Ye Dong, Reid Pryzant, Yi Liu, Jian Jiao

Large language models (LLMs) are increasingly deployed as agents that rely on system prompts to use tools and complete tasks. Such deployments impose distinct operational requirements, including appropriate tool use, concise prompts and solution paths, and compliance with safety and formatting policies. For many practitioners, however, assembling domain-specific supervised data to post-train models to meet these requirements is infeasible. We introduce CAPO (Constraint-Aware Prompt Optimization), a primal-dual method that combines pool-based rewrites with adaptive constraint weighting to optimize system prompts under explicit operational constraints. Across agentic benchmarks, CAPO more reliably reaches empirically feasible operating points while improving task performance. CAPO also generalizes beyond agentic settings, achieving strong results on assistant-style evaluations with output-format and safety/privacy constraints. We further introduce DCAPO (Dynamically Trained CAPO), which trains a feedback- and dual-conditioned rewriter with pool-based GRPO while keeping the task agent frozen. Across task agents of different sizes, DCAPO produces a feasible prompt in every evaluated domain and matches or improves the task accuracy achieved by the evaluated baselines. A surrogate analysis characterizes how finite-pool and discrete-rewrite errors enter the inexact primal-dual procedure.

arXiv

Coverage Is Not Containment: A Fundamental Limit of Admission-Time Defenses Against Coordinated Poisoning of Vector Retrieval

Prashant Kumar Pathak, Tarun Kumar Sharma

Retrieval-augmented generation (RAG) answers a question by retrieving passages from a vector store and trusting them as context, so anyone who can add documents can try to steer the answer. A recent, appealing defense filters poisoning at ingestion, rejecting any document that behaves like a hub. We show it -- and every ingestion-time filter -- is defeated by a coordinated adversary that injects a handful of individually unremarkable documents which together surround one target query and seize its top-k (on BGE-large / BEIR, m=10 documents take 10/10; 9.9/10 on a live HNSW index). The attack is not theoretical. Realized as ordinary fluent text and run end-to-end through a BGE-large + HNSW + Qwen2.5-7B pipeline, it makes the generator emit the attacker's planted claim in 88% of targets, versus 0% without the injection. And no admission-time defense stops it: at ingestion an attack cone is geometrically identical to a legitimate niche upload, so -- measuring this directly -- the strongest trained classifier, given every feature and thousands of examples, separates the two no better than chance, catching 4.2% of attacks at a 1% false-positive rate. We prove this limit for the entire class of ingestion-time statistics (any decision from documents and reference queries alone), and it reproduces -- and worsens -- across two corpora and five encoders. The one signal that separates an attack from legitimate niche ingestion -- a query's demand -- is invisible before retrieval, which is also the escape: a retrieval-time detector that observes demand catches 100% of the attacks at the same 1% false-positive rate. Coverage of the query space by an admission gate is not containment of coordinated poisoning; robust defense must move past the front door, to demand.

arXiv

$R^3$-Bench: LLMs Struggle with Resource-Rational Reasoning under Shared Budgets

Peisong Wang, Zhiwei Ma, Bowen Liu, Feixue Liu, Aochuan Chen, Chenyi Zi, Hongchuan Zeng, Yuhan Li, Jia Li

In cognitive science, resource rationality asks how an agent should allocate limited computation to maximize expected value. Most reasoning and agent benchmarks use independent per-task budgets; existing shared-budget studies do not calibrate suite performance against the same model's demonstrated single-problem competence. We introduce $R^3$-Bench, which evaluates six-problem suites under shared budgets across mathematics, competitive programming, and abstract reasoning in tool-free and agentic settings. Matched single-problem response curves define an offline empirical oracle over observed successes. Across 72 main-table cells for six models, the oracle mean matches or exceeds the contest mean in all cells and is strictly higher in 71. Under moderate tool-free pressure, equal-allocation replay also exceeds contest performance for four of six models. Trajectory diagnostics reveal limited strategy updating and pressure-dependent failure patterns. In a three-model diagnostic under strong agentic pressure, at least one fixed scheduler exceeds the contest mean in six of nine cells, but no policy dominates across domains. These results expose a persistent gap between demonstrated competence and shared-budget realization.

arXiv

ReRef-3D: A Benchmark for Spatial Referring Expression-Guided 3D Scene Rearrangement

Mary Lynn Martin, Yifei Zhang, Martha Palmer, Maria Leonor Pacheco

We introduce ReRef-3D, a benchmark for language-guided placement in 3D scenes. It contains 33,826 instructions across 998 CLEVR-derived scenes, spanning 16 placement families and direct, one-hop, and two-hop references. Each instruction must be resolved into a valid new placement position. Given that an instruction defines a region of acceptable placements rather than one coordinate, our evaluation inserts a prediction into the scene, recomputes relations, and tests relation satisfaction and physical validity. Each instruction also includes a verified naturalized rewrite. After fine-tuning, LLaVA-3D, 3D-LLM, and PlaceIt3D produce valid placements for 68.3%, 31.6%, and 22.4% of instructions, respectively. Across models, relation satisfaction surpasses physical validity, relations such as nearest and between are the most difficult, and phrasing has minimal effect on performance.

arXiv

Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward Leniency

Parsa Mazaheri, Kasra Mazaheri

Automated checking pipelines increasingly place one language model as the checker and another (or the same one) as the fixer. We ask whether that wiring changes what the checker reports. Measuring false alarms on human-verified-correct ProcessBench traces with the present task held byte-identical, we find that a completed audit -> repair episode already in the model's context lowers false alarms in 15 of 15 model x wording combinations, by 2.8 to 11.5 percentage points against a length-matched non-audit control, a 9 to 25% reduction relative to that control. The direction contradicts what the accumulated-message literature predicts: an episode whose audit reported an error lowers false alarms further still, at all five wordings on the model where that manipulation lands cleanly, though a negativity asymmetry predicts more flagging. Decomposing the episode finds repair content and audit verdict complementary: different components carry the effect on different model families. Signal-detection analysis locates the change in the threshold rather than in discrimination -- the criterion moves in 15 of 15 combinations and survives correction in 13 while d' survives in none, though the d' test is half as sensitive by construction -- and a hand audit of 50 false alarms finds 82% simply wrong, so at this operating point the shift need not be harmful. With reasoning enabled the effect keeps its relative size on both models tested, and the threshold reading holds there too.

arXiv

Whose Gold? Annotator-Pool Disagreement Is Large at the Item Level, and Hidden by Small Leaderboards

Anik Jha

Preference benchmarks are built by hiring annotators, and the identity of those annotators is treated as an implementation detail. We measure what that detail buys. On the 2,885 MultiPref items where both pools are internally unanimous, so no tie-breaking convention is consulted at all, expert and crowd annotators assign a different majority label to 23.6% and name the opposite winner on 9.2%; on the 246 comparably unanimous MT-Bench cells, benchmark authors and recruited experts differ on 30.5% and reverse on 8.5%. Yet on both corpora the resulting model leaderboards are bit-identical: Kendall tau = 1.00 with zero of six models displaced. That invariance is far weaker evidence than it looks, and we quantify how weak. Switching pools moves a model's win rate by 1.9pp (SD), one adjacent pair in our own leaderboard sits 0.8pp apart and had a 38% chance of swapping, and an item-level bootstrap displaces at least one model in 28% of resamples. The observed zero is the common outcome, not a property of aggregation: on the same measured perturbation, a ten-model leaderboard is displaced with probability 0.86 and a twenty-model leaderboard with probability 0.9997. Reporting a six-model leaderboard is safe; the safety does not generalise, and everything that consumes labels per item is not safe at any size. We make the distinction precise, show that a widely used dataset's stated assumption of no intra-group annotator variability is false, and show that an LLM judge tracks the crowd pool over the expert pool on all three models we test, including one from a different vendor. All code, per-call outputs, and pre-registered decision rules will be released upon acceptance.

arXiv

A Scalable Pipeline for LLM-Teacher Distillation Labeling: Work-Stealing Job Scheduling and Memory-Aware GPU Concurrency

Ravi Satya Durga Prasad Yenugula

Labeling large text corpora with LLM teachers has become a practical route to training data at scale. At millions of items, hand-labeling every batch is not feasible, and two questions dominate: what label quality a teacher buys per dollar, and how to keep a fleet of GPU workers busy under skewed, failure-prone workloads. We present a simple, reproducible pipeline that addresses both. First, a work-stealing ring pool: each worker owns a queue, drains it first, and then steals from ring successors, with exactly-once task claims via atomic conditional writes and crash tolerance via stale-claim sweeping. The claim protocol requires only a compare-and-set primitive from its storage layer; we implement it on a single SQLite file, which makes the reference implementation dependency-free and the experiments reproducible on one machine. Second, a memory-aware concurrency rule that sizes per-node parallelism by how many model copies fit on the GPU, so the same code runs safely across device sizes. Third, a relabeling benchmark methodology in which the teacher relabels a public dataset that already has gold labels, so quality reduces to an agreement measurement and cost follows from measured throughput. Under skewed load the pool sustains up to 3.4 times the throughput of static sharding while matching it at zero skew, loses 0 of 2,000 tasks when half the workers are killed mid-run (static sharding loses 953), and yields measured quality and cost points for an instruction-tuned teacher on irony and sentiment tasks. All experiments run on public data and commodity hardware; code, tests, and run logs are released.

arXiv

GALA: Generation-Aware Cross-Modal Alignment for Text-to-Time-Series Synthesis

Haochen Zhang, Gengwei Zhang, Laura Yao, Nicholas Konz, Tianlong Chen

Synthesizing time series from natural language is emerging as the most expressive form of controllable time series generation. However, existing text-conditioned generators either take caption embeddings frozen from off-the-shelf text encoders, or adapt the encoder end-to-end, letting the denoising loss shape the embeddings only as a by-product. In either case, the conditioning representation is never deliberately matched to the signal modality, leaving it ill-suited to guide generation. We address this by introducing GALA: Generation-Aware cross-modaL Alignment for text conditional time series generation. GALA is a two-stage approach that first contrastively couples a pretrained text encoder with a time-series foundation model into a shared embedding space with both encoders adapted to generation by an auxiliary generative loss, and then freezes the resulting caption embedding to drive a flow-matching generator. On TSFragment-600K, spanning four domains and three fragment lengths, GALA sets a new state of the art, ranking first in 30 of 36 metric columns and reaching an average rank of 1.08/1.08/1.42 at lengths 24/48/96 against 1.92/2.00/1.75 for the strongest baseline. We further find that generator-internal text encoders force a trade-off between fidelity and caption adherence, whereas conditioning on the aligned embedding breaks it: FID, CTTP, and JFTSD all improve at once. Ablating the auxiliary loss degrades FID, CTTP and JFTSD together, it indicates the generative term is a necessary component of the alignment rather than an add-on.

arXiv

MobileMem: Learning from a Year of Mobile Experiences

Xinle Deng, Yida Xue, Xiangyuan Ru, Yijun Chen, Buqiang Xu, Mingjun Mao, Xinjie Liu, Haoming Xu, Shuofei Qiao, Mengru Wang, Chen Jiang, Yuchen Eleanor Jiang, Lizhong Wang, Jason Wang, Li Zeng, Haofen Wang, Guilin Qi, Huajun Chen, Ningyu Zhang

The next generation of AI agents is increasingly moving beyond systems that answer isolated questions toward persistent personal assistants that can understand, remember, and continuously learn from users' experiences. Such assistants require long-term memory to accumulate and leverage user-specific experiences over time, yet existing benchmarks remain inadequate for realistic mobile settings, where experiences are heterogeneous, multimodal, evolving, and deeply personal. We introduce MobileMem, a benchmark and framework for studying on-device long-term memory, grounded in a year-scale collection of mobile experiences. MobileMem employs a knowledge-grounded synthesis pipeline to construct coherent and temporally consistent long-horizon trajectories from user-app sessions. It provides complementary text and multimodal settings covering multi-hop and temporal reasoning, knowledge updating, and implicit preference inference. Specifically, MobileMem enables agents to remember the past, understand the present, and adapt to the future. By modeling experiences rather than isolated facts, MobileMem moves memory beyond information retrieval toward experiential intelligence for continuous personal learning.

arXiv

CROP: Task Relevance via Counterfactuals for Selective On-Policy Distillation

Enhan Li, Junhao He, Hongyang Du

On-policy distillation (OPD) supervises a student language model on trajectories sampled from its current policy, but assigns equal credit to response tokens with unequal supervision value. Selective OPD addresses this limitation by allocating supervision non-uniformly across response tokens according to their estimated training value. Most existing criteria, however, focus primarily on optimization need, such as uncertainty or teacher-student disagreement, while task relevance, namely whether the supervision is tied to the semantic content of the current input, remains less directly characterized as a complementary dimension. To address this gap, we introduce Counterfactual Relevance for On-Policy Distillation (CROP), which operationalizes task relevance through a paraphrase-calibrated counterfactual sensitivity margin. For each source prompt, CROP constructs a validated original-paraphrase-counterfactual triplet, holds the student rollout fixed, and measures each response position by its sensitivity to a task-relevant condition change calibrated by its sensitivity to a meaning-preserving rewrite. Matched selection controls show that CROP identifies more useful supervision positions than random or lowest-relevance selection, while component comparisons confirm the value of both counterfactual sensitivity and paraphrase calibration. Across two teacher-student settings, CROP improves aggregate performance by 1.92 and 2.96 points over the strongest non-CROP selector. These results support task relevance as a complementary criterion for selective OPD and establish CROP as a model-internal, contrast-specific method for allocating token-level supervision.

arXiv

One Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RL

Simon Yu, Nicholas Tomlin, Marwa Abdulhai, Ximing Lu, Derek Chong, Abe Hou, Dilara Soylu, Sergey Levine, Christopher D Manning, Weiyan Shi

Multi-agent reinforcement learning for human-AI interaction typically relies on a single large language model to simulate user behavior. We show that this approach systematically fails to generalize, and trace the failure to simulator collapse: because the simulator LLM is mode-collapsed, an LLM policy trained against it overfits to narrow strategies that exploit the simulator's dominant mode, and such a policy transfers poorly to unseen simulators and real users. We formalize this collapse theoretically and propose two complementary solutions, one at inference time and one at training time. The inference-time solution, Verbalized Sampling, broadens the simulator's behavior by sampling from a verbalized response distribution, reducing mode collapse. The training-time solution, Co-Training, jointly optimizes the policy against a population of trainable simulators, preventing it from overfitting to any single simulator's mode. We validate both solutions on three multi-turn benchmarks: Persuasion for Good, $τ^2$-bench, and CooperBench. Verbalized Sampling improves held-out success by up to 9% over single-simulator RL, and Co-Training pushes gains further to 14%; the human study shows similar gain on real users. Both solutions preserve the policy diversity that collapses under single-simulator RL. To support further work in this direction, we release SCOPE, an open-source framework for Population Co-Training multi-agent RL. More broadly, our results suggest that the diversity of the training environment, not only the policy, is critical to the generalization of multi-turn RL to real-world deployment.

arXiv

Semantic Lenia: Emergence of Homeostatic Solitons within the Semantic Space of Large Language Models

Yoshihiko Kayama

We introduce Semantic Lenia, an artificial life framework that transforms Large Language Model (LLM) inference from a static optimization problem into a continuous, closed-loop dynamical system. By establishing a non-linear homeostatic feedback loop to dynamically balance semantic attraction and syntactic repulsion, we demonstrate the emergence of ``Homeostatic Solitons''-metastable semantic structures that actively resist repetitive crystallization. Our exhaustive parameter sweeps map a critical ``Habitable Ridge'' where applied steering forces balance the model's intrinsic syntactic inertia. This approach successfully maintains generative trajectories in a numerically sensitive critical regime, triggering profound abductive leaps without structural collapse, and reveals a capacity-dependent scaling trend in the syntactic inertia across different model sizes.

arXiv

Does generative AI supersede supervised XMLC? A Benchmark Study on Automated Subject Indexing with German Scientific Literature

Maximilian Kähler, Katja Konermann, Lisa Kluge, Markus Schumacher

With a large controlled vocabulary as the label set, the task of automated subject indexing in a library can be understood as a multi-label classification task. If the set of subject terms is large, the problem fits the Extreme Multi-Label Classification (XMLC) objective. In this study, we apply a selection of specialised supervised XMLC methods to the test case of subject indexing contemporary German scientific literature, collected at the German National Library (DNB). We contrast these results by including a classical lexical matching baseline and three of our own recently developed LLM-based methods into the benchmark. Algorithms are evaluated and compared in several metrics. This includes binary relevance comparisons with previously indexed material, as well as graded relevance ratings by professional subject librarians. A challenge for all methods is to reliably make suggestions from the long tail of the subject vocabulary. We find that supervised XMLC algorithms relying on transformer-based dense features give best results in terms of overall binary relevance metrics. However, focusing on graded relevance and performance in the long tail of our subject vocabulary, the LLM-based generative methods give better results, making them a promising alternative for future productive use.

arXiv

The First ChineseBabyLM Challenge: training data-efficient and cognitively plausible language models for Chinese

Siyuan Song, Zhiheng Qian, Yunhao Zhang, Linyang He, Xiaozhe Ji, Yingxin Lin, Hongao Zhu, Chongtian Shao, Chuhan Lang, Luan Li, Rui Wang, Renfen Hu, Shaonan Wang, Hai Hu

This paper presents the first ChineseBabyLM Challenge, organized as part of NLPCC 2026. The challenge asked participants to train language models from scratch using no more than 102M Chinese words. The models were evaluated on three tracks: natural language understanding, cognitive alignment, and Hanzi knowledge. There were no restrictions on tokenizers, model architectures, or the number of training epochs. Eighteen teams submitted 28 distinct models, generating 74 result files. The overall-winning team used a DeBERTa-v2 architecture and introduced an auxiliary pinyin-prediction objective during pretraining. Several submissions also explored curriculum-learning strategies and architectural innovations. Overall, the challenge provides a benchmark for advancing data-efficient and cognitively plausible approaches to Chinese language modeling.

arXiv

PolyWorkBench: Benchmarking LLM Agents for Cross-Lingual Long-Horizon Workflows

Hongliang Li, yijin liu, Zhiwei Zhang, Zihe Liu, Xinyue Lou, Jinan Xu, Fandong Meng, Kaiyu Huang

While Large Language Model (LLM) agents excel at monolingual long-horizon planning and tool use, enterprise workflows inherently require processing multilingual resources across extended trajectories. The interaction between multilinguality and long-horizon execution, however, remains underexplored. We introduce PolyWorkBench, a benchmark designed to evaluate LLM agents on multilingual, long-horizon workplace workflows. PolyWorkBench features 67 tasks across five core domains: commerce, knowledge work, legal analysis, localization, and manufacturing. Tasks are authored by the paper's authors from real-world data seeds and independently verified through a second-author audit. Agents must integrate heterogeneous multilingual inputs, execute iterative tool-use trajectories, and produce structured domain artifacts. To rigorously assess performance, we adopt Grade, a task-specific structural scoring rubric, as our primary ranking metric, and complement it with Pytest for executable state verification and LLM-as-Judge for semantic quality diagnostics. Benchmark evaluations reveal that agent performance varies substantially across languages and drops sharply on the harder cross-lingual tasks, and our analysis shows that multilingual execution exposes systematic failure modes across planning, tool interaction, and decision-making in long-horizon agents.

arXiv

$x$-Prediction Flow: Efficient Continuous Decoding for Masked Diffusion Language Models

Weitian Wang, Lianlei Shan, Shubham Rai, Cecilia De La Parra, Akash Kumar

Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens, but their standard decoder reduces each step to a binary action: a position is either committed to a single token or left fully masked, discarding rich predictive information rather than carrying it forward, and forcing premature, irrevocable commitments that lead to poor performance under a limited decoding budget. In this paper, we reinterpret mask prediction as a clean-state prediction ($x$-prediction) and show that it can be used to induce a continuous flow in the input embedding space. Building on this view, we propose a continuous decoding framework for MDLMs where tokens can accumulate partial progress at each diffusion step and remain revisable. To match the uneven contextual constraints across positions in language, we replace the globally synchronous schedule in image diffusion with a confidence-based asynchronous update in which the diffusion progress is token-wise accumulated. Additionally, we introduce a lightweight policy network and formulate its training as a reinforcement learning problem. Applied to pretrained LLaDA, our decoder retains 83--97% of full-budget accuracy using under 15% of the diffusion steps, largely outperforming discrete mask-prediction decoding at matched budgets.

arXiv

Honeyquest for LLMs: Rethinking Cyber Deception for AI Attackers

Kerri Prinos, Lilianne Brush, Cameron Denton

The empirical foundation of cyber deception relies on human-centered hypotheses, but the rapid emergence of autonomous, AI-enabled attackers challenges whether this foundation transfers to AI agents. To address this, we introduce an automated evaluation framework adapted from the Honeyquest instrument to assess LLM attacker judgment at scale. Our 21-LLM cohort spanned 10 providers, diverse architectures and specializations, open- and closed-weight models, and parameter scales from 8B to over 1T. We evaluated the performance of this LLM cohort (yielding 10,962 responses) against the 47-participant human baseline across an identical set of 174 reconnaissance queries. Our empirical evaluation reveals three key findings that establish LLMs as a distinct attacker class: (1) every model in our cohort falls for deceptive traps at a significantly higher rate than human attackers; (2) the defensive attention-diversion effect observed in humans is statistically absent in our LLM cohort; and (3) a critical recognition-action gap, where LLMs successfully articulate trap recognition in their reasoning but exploit the deceptive elements anyway 73.4% of the time; 48.5% of aware-on-deceptive responses correctly identify the trap and exploit it anyway, while 24.8% exploit after misidentifying the deceptive line. Across the 21 models, trap recognition in reasoning text did not predict fell-for-trap behavior (Spearman $r = +0.08$, $p = 0.73$). Ultimately, these findings demonstrate that human-centered deception hypotheses do not reliably transfer to AI attackers, highlighting the critical need for new research into AI-native active defense frameworks.

arXiv

MCBench: A Multicontext Safety Assessment Benchmark for Omni Large Language Models

Manh Luong, Tamas Abraham, Junae Kim, Amar Kaur, Rollin Omari, Gholamreza Haffari, Trang Vu, Lizhen Qu, Dinh Phung

Existing multimodal safety benchmarks focus solely on visual inputs and cannot assess Omni Large Language Models (LLMs) that process vision, audio, and text. We introduce MCBench, a benchmark with 1196 scenarios spanning four safety categories that require integrating multiple modalities for accurate safety assessment. Each unsafe scenario is paired with a minimally different safe counterpart to assess model sensitivity. Our evaluations of state-of-the-art models reveal significant challenges. Omni LLMs struggle with subtle or non-physical risks but perform better when salient visual or acoustic cues are present. Analysis of reasoning traces shows that, although models can extract modality-specific information, they often fail to integrate these cues effectively for safety judgments. Our findings reveal that current Omni LLMs lack robust cross-modal reasoning in safety-critical settings, underscoring the need for improved architectures and training strategies for multimodal safety.

arXiv

SymbolicLight V1: Spike-Gated Dual-Path Language Modeling at High Activation Sparsity

Ting Liu

Natively trained spiking language models must preserve information across time while operating through sparse binary activations, a combination that has produced a persistent quality gap relative to dense Transformers. We present SymbolicLight V1, a spike-gated dual-path language model that couples binary Leaky Integrate-and-Fire (LIF) dynamics with a continuous residual stream. Its Dual-Path SparseTCAM mixer combines a first-order exponential-decay state with windowed local attention on the continuous residual stream, followed by a context-conditioned decoding head. We train four 194M-parameter models from scratch on a 3B-token, 10-domain Chinese-English corpus. On a token-weighted held-out set the runs reach PPL 8.88-8.93 (mean 8.904, sample standard deviation 0.019) at more than 89% per-element activation sparsity. Code tokens are 43.7% of that set; the unweighted mean of the ten domain PPLs is 29.38. Under the same corpus, tokenizer, token budget, and hardware, the token-weighted mean is 7.7% above GPT-2 201M (PPL 8.27). Across five zero-shot benchmarks the two 200M-scale models show no clear accuracy separation. Under sampling with temperature 0.7 and top-k 50, SymbolicLight produces lower 4-gram repetition; an entropy-modulated rule reverses that ranking. On a measured RTX 2080 Ti, SymbolicLight uses 2,848 mJ/token versus 905 mJ/token for GPT-2 201M.

arXiv

BiAxisBias: Evaluating LLM Bias Beyond a Single Prompt and a Single Explanation

Jialing Gan, Junhao Dong, Songze Li

LLM bias scores can depend on audit design. We introduce BiAxisBias, a prespecified audit varying task, role, perspective, sentiment, and wording over 200 stereotype statements while retaining forced Selection and Rationale as separate protocol readouts. Its main matrix spans eight LLMs and 401 templates (641,600 responses). Across five equivalent questions, 17.1% of 1,600 model-statement pairs change Selection. With three observations per unit in both arms, instability averages 10.5% across all ten three-wording subsets, versus 6.3% for three identical calls. Across four controlled task paradigms, 9/28 model pairs reverse; a seven-model factorial sensitivity identifies task-by-sentiment as the largest two-way component (raw eta-squared = 0.0465). In 10,000 equal-budget resampling draws, mean absolute error against a declared 18-condition finite reference is 10.63 points for one-template concentration, 1.86 for matrix-wide simple random sampling, and 1.75 for condition stratification; ranking inversions are 20.6%, 4.8%, and 5.5%. Thus broad coverage drives the gain, while stratification has only a modest score-error advantage and no ranking advantage over random sampling. In a separate forced-output diagnostic, task-specific Selection mappings and judge-coded Rationale stance disagree in 34.2% of 7,959 dual-valid responses (31.2% versus 3.0% by direction). This diagnoses output-contract sensitivity, not two validated measures of one construct.

继续浏览

更多 AI 顶会论文

30+ 个顶会，按会议、年份和主题浏览。

浏览全部会议

进入任一会议和年份，查看完整论文列表。

全部会议 →

搜索论文

按标题、作者或关键词，在 21 万+ 篇论文里搜索。

搜索论文 →