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

Toward Robust LLM-Based Judges: Taxonomic Bias Evaluation and Debiasing Optimization

Hongli Zhou, Hui Huang, Rui Zhang, Kehai Chen, Bing Xu, Conghui Zhu, Tiejun Zhao, Muyun Yang

Large language model (LLM)-based judges are widely adopted for automated evaluation and reward modeling, yet their judgments are often affected by judgment biases. Accurately evaluating these biases is essential for ensuring the reliability of LLM-based judges. However, existing studies typically investigate limited biases under a single judge formulation, either generative or discriminative, lacking a comprehensive evaluation. To bridge this gap, we propose JudgeBiasBench, a benchmark for systematically quantifying biases in LLM-based judges. JudgeBiasBench defines a taxonomy of judgment biases across 4 dimensions, and constructs bias-augmented evaluation instances through a controlled bias injection pipeline, covering 12 representative bias types. We conduct extensive experiments across both generative and discriminative judges, revealing that current judges exhibit significant and diverse bias patterns that often compromise the reliability of automated evaluation. To mitigate judgment bias, we propose bias-aware training that explicitly incorporates bias-related attributes into the training process, encouraging judges to disentangle task-relevant quality from bias-correlated cues. By adopting reinforcement learning for generative judges and contrastive learning for discriminative judges, our methods effectively reduce judgment biases while largely preserving general evaluation capability.

arXiv

RAG Strategies for Natural Language-Based SQL Query and REST API Call Generation

Tim Schlippe, Simon Martin, Michael Marketsmüller

Enterprise software systems commonly expose business functionality through both relational databases and REST APIs. Accessing these interfaces requires specialized technical knowledge, as users must determine whether a request requires a database query or an API operation and understand the corresponding schemas, endpoints, and parameters. This creates demand for natural language interfaces that translate user requests into SQL queries and REST API calls. While large language models (LLMs) show promise for structured code generation, they typically lack reliable knowledge of enterprise-specific schemas, endpoints, and documentation. Retrieval-augmented generation (RAG) addresses this limitation by grounding generation in external documentation. However, prior work largely studies SQL query generation and REST API call generation separately, despite enterprise documentation environments often containing both database schemas and API specifications. We systematically evaluate standard RAG, Self-RAG, and CoRAG across SQL query generation, REST API call generation, and a combined task requiring routing between both operation types. Using SAP Transactional Banking as a realistic enterprise use case, we constructed an execution-validated dataset and compared retrieval strategies under database-only, API-only, and mixed-documentation settings. Retrieval augmentation proved essential for reliable enterprise structured generation, substantially improving performance over a no-retrieval baseline. CoRAG achieved the best results in the combined SQL query and REST API call setting, with statistically significant improvements in exact-match accuracy over standard RAG, primarily driven by stronger SQL query generation under mixed-documentation retrieval conditions. Overall, findings show that retrieval strategy substantially affects structured generation performance under mixed-documentation settings.

arXiv

Bagpiper: Solving Open-Ended Audio Tasks via Rich Captions

Jinchuan Tian, Haoran Wang, Bo-Hao Su, Chien-yu Huang, Qingzheng Wang, Jiatong Shi, William Chen, Xun Gong, Siddhant Arora, Chin-Jou Li, Masao Someki, Takashi Maekaku, Keita Goto, Yusuke Shinohara, Jin Sakuma, Chao-Han Huck Yang, Shinji Watanabe

Current audio foundation models typically rely on rigid, task-specific supervision (e.g., speech recognition), addressing isolated factors of audio rather than the whole. In contrast, human processes audio holistically, seamlessly bridging raw audio waveform with abstract cognitive concepts (e.g., all perception details of audio events) to execute complex tasks. Grounded in this philosophy, we introduce Bagpiper, an 8B audio foundation model that interprets physical audio via rich captions, i.e., comprehensive natural language descriptions that encapsulate the critical cognitive concepts inherent in the audio. By pre-training on a massive corpus of 600B tokens, the model establishes a robust bidirectional mapping between raw audio and this high-level conceptual space. During fine-tuning, Bagpiper adopts a caption-then-process workflow, simulating an intermediate cognitive reasoning step to solve diverse tasks without knowing prior task-specific practice. Experimentally, Bagpiper achieves universal generation that can uniformly generate speech, sound effects, music, and their arbitrary combinations. It also maintains comparable performance with the 7B Qwen-2.5-Omni for audio understanding. To the best of our knowledge, Bagpiper is among the first works that achieve open-ended audio understanding and generation on speech, sound, and music. Model, data, and code will be released at Bagpiper Home Page.

arXiv

Transformers perform adaptive partial pooling

Vsevolod Kapatsinski

Any language model must decide what to say in novel contexts based on information from similar contexts. But what about contexts that are not novel but merely infrequent? In hierarchical regression, the model's predictions for behavior in a context are affected by observations from similar contexts to the extent that 1) the current context is infrequent and 2) different contexts behave similarly. This is called adaptive partial pooling. This paper shows that next-word predictions of a transformer (GPT2) are affected by observations from outside the current context, but this pooling reduces with more training. Pooling is affected by context frequency, context number (type frequency) and context variability in a qualitatively similar way to hierarchical regression. However, there is a "sweet spot" in training at which the transformer best matches the behavior of hierarchical regression. This is the point at which the effect of context frequency on pooling is at maximum.

arXiv

When LLM Essays Outscore Student Essays: What a Korean Writing Rubric Rewards and Where Readers Disagree

Shinwoo Park, Yo-Sub Han

LLMs now help students plan, draft, and revise essays. Educational assessment therefore faces a basic question: how should student and LLM writing be compared? Rubrics assign points to content, organization, and expression. Their total can still hide which criteria drive the comparison, where ratings approach the maximum, and where readers disagree. We therefore conducted a secondary, post hoc audit of a Korean writing study with source-informed scoring. Three Korean language and literature majors used a 16-criterion, 100-point rubric to score six student essays and 24 essays from four LLMs prompted for three school levels. After source disclosure, they helped develop the rubric and, according to the protocol, scored the Phase 2 essays without gold labels while recording source judgments. The pooled LLM mean exceeds the student mean by 18.35 points. Orthographic norms and genre-appropriate register contribute 8.19 points, or 44.7% of the gap. All 72 LLM ratings reach the orthographic maximum, and 70 reach the register maximum. The student source mean is highest or tied highest on both creativity criteria and natural Korean phrasing, where reader agreement is weak. Criterion-level auditing therefore offers a deeper account of human and LLM writing than the total alone.

arXiv

Hylog: A Hybrid Approach to Logging Text Production in Non-alphabetic Scripts

Roberto Crotti, Giovanni Denaro, Zhiqiang Du, Ricardo Muñoz Martín

Research keyloggers are essential for cognitive studies of text production, yet most fail to capture the on-screen transformations performed by Input Method Editors (IMEs) for non-alphabetic scripts. To address this methodological gap, we present Hylog, a novel hybrid logging system that combines analytical keylogging with ecological text logging for a more complete and finer-grained analysis. Our modular, open-source system uses plug-ins for standard applications (Microsoft Word, Google Chrome) to capture both keyboard output and rendered text, which a hybridizer module then synchronizes into a dual trace. To validate the system's technical feasibility and demonstrate its analytical capabilities, we conducted a proof-of-concept study where two volunteers translated a text into simplified Chinese. Hylog successfully captured keypresses and temporal intervals between Latin letters, Chinese characters, and IME confirmations -- some measurements invisible to traditional keyloggers. The resulting data enable the formulation of new, testable hypotheses about the cognitive restrictions and affordances at different linguistic layers in IME-mediated typing. Our plug-in architecture enables extension to other IME systems and fosters more inclusive multilingual text-production research.

arXiv

DeepSurvey-Bench: Evaluating Academic Value of Automatically Generated Scientific Surveys

Guo-Biao Zhang, Xian-Ling Mao, Ding-Yuan Liu, Da-Yi Wu, Tian Lan, Huihui Li, Heyan Huang

The rapid development of automated survey generation technology has made it increasingly important to establish a comprehensive benchmark to evaluate the quality of generated surveys. Most existing benchmarks first construct ground-truth datasets by selecting human-written surveys based on limited selection criteria, such as citation counts and structural coherence, and evaluate generated surveys primarily based on conventional quality dimensions, including structural quality and reference relevance. However, these benchmarks have two key issues: (1) the datasets are insufficiently reliable because the selection criteria only identify highly cited or structurally coherent surveys without verifying their academic value; (2) the evaluation metrics mainly reflect the surface-level quality of generated surveys and are insufficient to assess their academic value. Together, these issues prevent existing benchmarks from effectively assessing the academic value of generated surveys. To address the above problems, we propose DeepSurvey-Bench, a comprehensive benchmark for evaluating the academic value of automatically generated surveys. Specifically, our proposed benchmark introduces a set of academic value evaluation criteria covering three dimensions: informational value, scholarly communication value, and research guidance value. We first construct a reliable dataset with academic value annotations based on these criteria, and then evaluate the academic value of generated surveys according to these criteria through a multi-LLM-as-a-judge approach. Extensive experiments demonstrate that DeepSurvey-Bench not only aligns closely with human assessments in evaluating the academic value of surveys, but also reveals underlying academic value beyond the reach of surface-level quality metrics, providing a foundation for fine-grained diagnosis and iterative improvement of generated surveys.

arXiv

Interpretable Recognition of Cognitive Distortions in Natural Language Texts

Anton Kolonin, Anna Arinicheva

We propose a new approach to multi-factor classification of natural language texts based on weighted structured patterns such as N-grams, taking into account the heterarchical relationships between them, applied to solve such a socially impactful problem as the automation of detection of specific cognitive distortions in psychological care, relying on an interpretable, robust and transparent artificial intelligence model. The proposed recognition and learning algorithms improve the current state of the art in this field. The improvement is tested on two publicly available datasets, with significant improvements over literature-known F1 scores for the task, with optimal hyper-parameters determined, having code and models available for future use by the community.

arXiv

LM-mixup: Text Data Augmentation via Language Model based Mixup

Zhijie Deng, Zhouan Shen, Ling Li, Yao Zhou, Zhaowei Zhu, Yanji He, Wei Wang, Jiaheng Wei

Instruction tuning is crucial for aligning Large Language Models (LLMs), yet the quality of instruction-following data varies significantly. While high-quality data is paramount, it is often scarce; conversely, abundant low-quality data is frequently discarded, leading to substantial information loss. Existing data augmentation methods struggle to augment this low-quality data effectively, and the evaluation of such techniques remains poorly defined. To address this, we formally define the task of Instruction Distillation: distilling multiple low-quality and redundant inputs into high-quality and coherent instruction-output pairs. Specifically, we introduce a comprehensive data construction pipeline to create MIXTURE, a 144K-sample dataset pairing low-quality or semantically redundant imperfect instruction clusters with their high-quality distillations. We then introduce LM-Mixup, by first performing supervised fine-tuning on MIXTURE and then optimizing it with reinforcement learning. This process uses three complementary reward signals: quality, semantic alignment, and format compliance, via Group Relative Policy Optimization (GRPO). We demonstrate that LM-Mixup effectively augments imperfect datasets: fine-tuning LLMs on its distilled data, which accounts for only about 3% of the entire dataset, not only surpasses full-dataset training but also competes with state-of-the-art high-quality data selection methods across multiple benchmarks. Our work establishes that low-quality data is a valuable resource when properly distilled and augmented with LM-Mixup, significantly enhancing the efficiency and performance of instruction-tuned LLMs.

arXiv

Unpacking Hateful Memes: Presupposed Context and False Claims

Weibin Cai, Jiayu Li, Reza Zafarani

While memes are often humorous, they are frequently used to disseminate hate, causing serious harm to individuals and society. Current approaches to hateful meme detection mainly rely on pre-trained language models. However, less focus has been dedicated to \textit{what make a meme hateful}. Drawing on insights from philosophy and psychology, we argue that hateful memes are characterized by two essential features: a \textbf{presupposed context} and the expression of \textbf{false claims}. To capture presupposed context, we develop \textbf{PCM} for modeling contextual information across modalities. To detect false claims, we introduce the \textbf{FACT} module, which integrates external knowledge and harnesses cross-modal reference graphs. By combining PCM and FACT, we introduce \textbf{\textsf{SHIELD}}, a hateful meme detection framework designed to capture the fundamental nature of hate. Extensive experiments show that SHIELD outperforms state-of-the-art methods across datasets and metrics, while demonstrating versatility on other tasks, such as fake news detection.

arXiv

TS-Reasoner: Aligning Time Series Foundation Models with LLM Reasoning

Fangxu Yu, Hongyu Zhao, Tianyi Zhou

Time series reasoning is crucial to decision-making in diverse domains, including finance, energy, and scientific discovery. While existing time series foundation models (TSFMs) can capture low-level dynamic patterns and provide accurate forecasting, further analysis usually requires additional background knowledge and sophisticated reasoning, which are lacking in most TSFMs but can be achieved through Large Language Models (LLMs). On the other hand, without expensive post-training, LLMs often struggle with the numerical understanding of time series data. Although it is intuitive to integrate the two types of models, developing effective training recipes that align the two modalities for reasoning tasks is still an open challenge. To this end, we propose TS-Reasoner that aligns the latent representations of TSFMs with the textual inputs of LLMs for downstream understanding/reasoning tasks. Specifically, we propose a simple yet effective method to curate diverse, synthetic pairs of time series and textual captions for alignment training. We then develop a two-stage training recipe that applies instruction fine-tuning after the alignment pretraining. Unlike existing works that train an LLM to take time series as inputs, we leverage a pretrained TSFM and freeze it during training. Experiments on several benchmarks demonstrate that TS-Reasoner not only outperforms a wide range of open-source LLMs, Vision-Language Models (VLMs), and Time Series LLMs of comparable scale, but also does so with remarkable data efficiency, e.g., using less than half the training data.

arXiv

SoM-1K: A Thousand-Problem Benchmark Dataset for Strength of Materials

Qixin Wan, Zilong Wang, Jingwen Zhou, Wanting Wang, Ziheng Geng, Jiachen Liu, Ran Cao, Lu Cheng

Foundation models have shown remarkable capabilities in various domains, but their performance on complex, multimodal engineering problems remains largely unexplored. We introduce SoM-1K, the first large-scale multimodal benchmark dataset dedicated to evaluating foundation models on problems in the strength of materials (SoM). The dataset, which contains 1,065 annotated SoM problems, mirrors real-world engineering tasks by including both textual problem statements and schematic diagrams. Due to the limited capabilities of current foundation models in understanding complicated visual information, we propose a novel prompting strategy called Descriptions of Images (DoI), which provides rigorous expert-generated text descriptions of the visual diagrams as the context. We evaluate eight representative foundation models, including both large language models (LLMs) and vision language models (VLMs). Our results show that current foundation models struggle significantly with these engineering problems, with the best-performing model achieving only 56.6% accuracy. Interestingly, we found that LLMs, when provided with DoI, often outperform VLMs provided with visual diagrams. A detailed error analysis reveals that DoI plays a crucial role in mitigating visual misinterpretation errors, suggesting that accurate text-based descriptions can be more effective than direct image input for current foundation models. This work establishes a rigorous benchmark for engineering AI and highlights a critical need for developing more robust multimodal reasoning capabilities in foundation models, particularly in scientific and engineering contexts.

arXiv

Quick on the Uptake: Eliciting Implicit Intents from Human Demonstrations for Personalized Mobile-Use Agents

Zheng Wu, Heyuan Huang, Yanjia Yang, Yuanyi Song, Xingyu Lou, Weiwen Liu, Weinan Zhang, Jun Wang, Zhuosheng Zhang

As multimodal large language models advance rapidly, the automation of mobile tasks has become increasingly feasible through the use of mobile-use agents that mimic human interactions from graphical user interfaces. To further enhance mobile-use agents, previous studies employ demonstration learning to improve mobile-use agents from human demonstrations. However, these methods focus solely on the explicit intention flows of humans (e.g., step sequences) while neglecting implicit intention flows (e.g., personal preferences), which makes it difficult to construct personalized mobile-use agents. In this work, to evaluate the Intention Alignment Rate between mobile-use agents and humans, we first collect MobileIAR, a dataset covering 40 English- and Chinese-speaking users and containing human-intent-aligned actions and ground-truth actions. This enables a comprehensive assessment of the agents' understanding of human intent. Then we propose IFRAgent, a framework built upon Intention Flow Recognition from human demonstrations. IFRAgent analyzes explicit intention flows from human demonstrations to construct a query-level vector library of standard operating procedures (SOP), and analyzes implicit intention flows to build a user-level habit repository. IFRAgent then leverages a SOP extractor combined with retrieval-augmented generation and a query rewriter trained through knowledge distillation to generate personalized queries and SOPs from a raw ambiguous query, enhancing the alignment between mobile-use agents and human intent. Experimental results demonstrate that IFRAgent consistently outperforms baselines on every model, achieving an average absolute improvement of 9.88\% (53.33\% relative) in human intention alignment rate and an average absolute improvement of 8.26\% (32.40\% relative) in step completion rate.

arXiv

An automatic-differentiation framework for time-lapse electrical resistivity tomography inversion of hydrologic dynamics

Pu Yang, Zhengyang Fang, Yuxin Liu, Xuan Su, Deshan Feng, Hang Chen

Time-lapse electrical resistivity tomography (TL-ERT) provides spatially distributed information on subsurface hydrologic changes. However, inversion of long monitoring sequences is computationally demanding. Modifying the data misfit, regularization, model parameterization, or petrophysical transformation may also require new gradient derivations and separate implementations. Here, we present AD-TLERT, a unified, GPU-accelerated framework for time-lapse ERT inversion based on automatic differentiation. The framework integrates model parameterization, differentiable petrophysical transformations, forward modeling, data misfit, regularization and auxiliary constraints into a single computational chain. Alternative inversion formulations can therefore reuse the same PDE derivative implementation without re-deriving the complete ERT sensitivity for each case. Comparisons with pyGIMLi showed close agreement in the forward responses, gradients, and recovered resistivity models. Under the tested configuration, AD-TLERT achieved an approximately 51-fold speedup. Synthetic experiments showed that inversion choices affect the amplitude, geometry, and temporal behavior of recovered anomalies. By propagating gradients through the embedded petrophysical relationship, AD-TLERT enabled direct water-content inversion and yielded more accurate estimates than post-inversion conversion for the tested model. A field application further demonstrated how ERT, temperature, and soil-moisture observations can be combined to image snowmelt-driven hillslope wetting. AD-TLERT provides an efficient and flexible framework for time-lapse ERT inversion and hydrologic interpretation.

arXiv

Ring-based Spatial Transformer: Learning Non-linear Spatial Interactions between Building Distribution and Pedestrian Flow

Shun Nakayama, Takahiro Kanamori, Wanglin Yan

This study proposes a ring-based SpatialTransformer to learn how building uses at different distances from a railway station interact to generate pedestrian flow. Concentric ring buffers at 100-meter intervals up to 800 meters were defined around 100 randomly selected stations in Tokyo, treating each ring as a spatial token. Self-Attention was applied to learn inter-zone interactions directly from data, without prior structural assumptions. GPS-derived walking trip counts served as the target variable and Geographically Weighted Regression as the baseline. Across 30 independent trials, the SpatialTransformer consistently outperformed GWR in predictive accuracy. SHAP analysis revealed that mid-to-outer distance zone features dominate pedestrian flow prediction, while features from the 0-100m zone contributed little. The attention matrix showed that each distance zone attends most strongly to spatially distant zones, demonstrating that pedestrian flow is regulated by structural interactions across the entire catchment area rather than by any single zone in isolation. These findings challenge the compact city assumption that station-proximate development maximizes pedestrian flow, and suggest that land use distribution across the full walkable catchment area deserves greater consideration in urban planning practice.

arXiv

When Uncertainty Isn't Enough: An Empirical Study of Self-Correction in Code Generation

Pranav Rakasi, Maanas Lalwani, Arnav Srivastava, Arya Palanivel, Tinuade Adeleke, Ruizhe Li, Sean Wu

Large language models for code generation often produce incorrect solutions without reliable indicators of failure. We study whether uncertainty estimation methods developed for natural language transfer to code generation, and whether such signals can improve code generation via selective self-correction. We evaluate five uncertainty methods: mean token entropy, verbalized confidence, $P(\text{True})$, entropy ensembles, and semantic entropy probes, across three small code LLMs on HumanEval and BigCodeBench. We find that multi-sample $P(\text{True})$ achieves the strongest correlation with correctness, while all the other methods, including semantic entropy probes, yield only weak correlation. We then use these uncertainty signals to drive three self-correction policies: adaptive decoding, uncertainty-based regeneration, and verification-based regeneration. Our results reveal a stronger negative finding than anticipated: uncertainty-based self-correction fails to reliably improve Pass@1, degrading accuracy in 5 of 6 configurations across both benchmarks ($-3$pp to $-10$pp), and adaptive decoding degrades accuracy in 4 of 6 configurations. Only verification-based self-correction reliably improves Pass@1, with gains of $+6$ to $+26$ percentage points on HumanEval and $+8$ to $+20$ percentage points on BigCodeBench, scaling inversely with baseline strength. These findings replicate consistently across both benchmarks and suggest that cheap uncertainty estimators are insufficient on their own to improve code correctness, and that their practical value lies in serving as gating signals for costlier execution-based correction loops rather than as standalone substitutes for verification.

arXiv

P2E-VQ: ECG-linked representation augmentation for PPG via discrete patch retrieval

Zhongli Wu, Zhuangzhi Gao, He Zhao, Feixiang Zhou, Fu Wang, Jinru Ding, Yuankai Wang, Hongyi Qin, Gregory Y. H. Lip, Bil Kirmani, Yalin Zheng

Photoplethysmography (PPG) is widely used in consumer wearables because of its low cost and ease of acquisition. However, unlike electrocardiography (ECG), PPG measures peripheral pulse dynamics rather than cardiac electrical activity, limiting its ability to predict cardiac conditions that rely on ECG-specific morphological cues. Existing methods attempt to bridge this gap by reconstructing ECG signals from PPG signals. However, this inverse mapping is inherently ill-posed, and faithful waveform reconstruction does not necessarily translate into improved downstream performance. To address this challenge, we propose P2E-VQ, a retrieval-augmented framework that replaces ECG waveform reconstruction with ECG-linked representation retrieval. Specifically, P2E-VQ converts PPG patches into discrete tokens and retrieves ECG-linked information from a memory bank constructed exclusively from the training data. This process augments PPG representations while requiring only PPG signals during inference. Extensive experiments on five public datasets covering six downstream tasks, including clinical endpoint prediction and affective state recognition, demonstrate that P2E-VQ consistently outperforms pretrained baselines under a unified frozen-feature linear-probing protocol.

arXiv

FedImp: Enhancing Federated Learning Convergence with Impurity-Based Weighting

Hai Anh Tran, Cuong Ta, Truong X. Tran

Federated Learning (FL) is a collaborative paradigm that enables multiple devices to train a global model while preserving local data privacy. A major challenge in FL is the non-Independent and Identically Distributed (non-IID) nature of data across devices, which hinders training efficiency and slows convergence. To tackle this, we propose Federated Impurity Weighting (FedImp), a novel algorithm that quantifies each device contribution based on the informational content of its local data. These contributions are normalized to compute distinct aggregation weights for the global model update. Extensive experiments on EMNIST and CIFAR-10 datasets show that FedImp significantly improves convergence speed, reducing communication rounds by up to 64.4%, 27.8%, and 66.7% on EMNIST, and 44.2%, 44%, and 25.6% on CIFAR-10 compared to FedAvg, FedProx, and FedAdp, respectively. Under highly imbalanced data distributions, FedImp outperforms all baselines and achieves the highest accuracy. Overall, FedImp offers an effective solution to enhance FL efficiency in non-IID settings.

arXiv

Do Uncertainty Signals Help? A Systematic Study of Uncertainty-Aware Decoding with Rollback Mechanisms

Xianzong Wu, Xiaohong Li, Yuejun Guo, Xinyang Liu, Tianlin Li, Junjie Wang, Qiang Hu

Prediction uncertainty is a widely adopted metric for quantifying model confidence, with downstream applications spanning model explanation, data selection, and prediction rollback. Despite its demonstrated utility, the potential of uncertainty quantification to enhance code generation in large language models (LLMs) remains largely underexplored, raising a critical question: to what extent can uncertainty serve as an effective signal for improving LLM-based code generation? To answer this question, we study uncertainty-aware rollback decoding, an inference-time strategy that uses uncertainty signals to identify unreliable generation regions and roll back to earlier valid prefixes without retraining the model. We evaluate this framework on seven code LLMs, five code generation benchmarks, and eight token-level uncertainty signals under a unified decoding setup. Our results show that the complete rollback framework improves over equal-budget restart across the evaluated benchmarks and model settings, with gains of up to 0.26 in pass@1 and 0.35 in AvgTestPassRate on functional code generation benchmarks, and an absolute improvement of up to 6.4\% in Patch-Aligned Safe Rate on Dsec-Python. Among the evaluated signals, information-theoretic measures such as token entropy and negative log-likelihood show the most favorable overall trend, frequently achieving the best or near-best results on standard benchmarks. A component-controlled ablation further shows that feedback-guided rollback provides the main improvement, while uncertainty localization provides an additional gain when checking, budget, rollback, and branch decay are held fixed.

arXiv

Measuring Cross-Task Behavioral Consistency in Language Model Agents

Amritesh Banerjee, Pranil Raichura

Agent evaluation relies almost entirely on outcome metrics such as success rate, which capture whether an agent succeeds but not how consistently it behaves. We argue that behavioral consistency across tasks is a distinct and measurable property, and we introduce the Behavioral Consistency Metric (BCM) to quantify it. BCM trains a model to predict task success from behavioral features of agent execution traces, derives a per-trajectory feature-attribution vector, and measures the mean pairwise similarity of these vectors within an agent system. Across roughly 9,000 trajectories from six language model agents on software engineering tasks, our central finding is that cross-task and within-task consistency are distinct axes that can diverge: some systems are locally reproducible, behaving similarly on repeated attempts at one task, yet globally fragmented, with no stable strategy across different tasks, while others are consistent at both scales. Prior work measures only same-task reproducibility and so cannot observe this separation. We further find that consistency is not reducible to success rate, since systems with comparable success can differ sharply in consistency, and that the frontier-versus-open-source consistency gap persists under a within-task control that holds task difficulty constant. We position BCM as a process-level reliability signal that complements outcome metrics, and we are explicit about the conditions under which it is meaningful.

Keep browsing

More papers from top AI conferences

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