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

Semantic Watermarking with Order-Robust Detection over Sub-sentence Units

Abdulrahman Diaa, Jonathan Petit, Florian Kerschbaum

Semantic watermarks tie the mark to sentence meaning rather than token choices, promising robustness to content-preserving edits. However, the detector only observes attacker-supplied text, which can be reworded, reordered, or resegmented to evade detection without content loss. Rewording, reordering, and resegmentation all cause embedding displacement: detection tests embeddings different from those selected during watermarking and can therefore lose the mark. Our adaptive embedding displacement attack (EDA) admits all three edits under a single objective that maximizes this displacement. It uses a public paraphraser and surrogate encoder without access to the provider's generator or secret key. At a 5% false-positive rate (FPR) and content-preservation threshold $\bar{q}=90\%$, EDA successfully removes the mark on between 32.6% and 47.9% of documents across four schemes, the highest among the tested attacks. Therefore, EDA evaluates the schemes' robustness more thoroughly than passive paraphrasing. To address these vulnerabilities, we design (k)-SwordStamp: semantic watermarks with order-robust detection over sub-sentence units, reducing sensitivity to attacker-chosen structure at a small quality cost. Against k-SwordStamp, the strongest no-box attack we test is an EDA variant adapted to its design, with a 10.8% attack-success rate. A stronger EDA with access to the provider's detector and secret key reaches a 39.7% attack-success rate, compared with 65.5% on k-SemStamp. Our code is available at https://github.com/D-Diaa/SwordStamp.

arXiv

Knowing Before Answering: Decoding Language Models for Reliable RAG

Syed Mahbubul Huq, Christopher Child, Tillman Weyde, Pranava Madhyastha

In Retrieval-Augmented Generation (RAG), retrieval may provide insufficient or conflicting information needed to answer a question. The system should not only know when to answer but also be able to identify cases in which the documents provided in RAG are insufficient or contain conflicting information. This can be framed as a three-way classification problem, where we use the model's internal signals to determine whether the provided information in the input can be classified as sufficient, insufficient, or conflicting. We create a controlled benchmark dataset that replicates a RAG setup with fictitious information and labels each instance as answerable, insufficient, or conflicting. We use hidden activations and attention-derived features as inputs to train a lightweight linear model to distinguish among the three classes. Across 16 language models spanning different architectures and a range of model sizes, our feature-based router consistently outperforms prompting-based baselines and the performance of specialised RAG-models. We further conduct analyses into the information dynamics of the models. We show that the most informative signals for the classification are available in the middle layers, with hidden activation states being more effective than attention values or the MLP-feature outputs in most of the tested models. Overall, our results suggest that language models internally encode whether retrieved evidence is sufficient to support answering, and that this signal can be decoded reliably for RAG triage.

arXiv

When Tokenizers Fail: Byte-Level Chunking for Zero-Shot Transfer to Low-Resource Languages

Sanjeev Kumar, Atsuki Yamaguchi, Nikolaos Aletras

Subword tokenization hinders low-resource language processing by imposing frequency patterns from dominant languages onto script-sharing variants. Byte-level models bypass this issue by processing raw UTF-8 characters, yet they create a granularity mismatch for word-level tasks in non-Latin scripts. Hierarchical byte-level architectures address this mismatch by grouping bytes into word-aligned chunks. However, these architectures require massive training data and suffer from representational misalignment when paired with frozen subword-based language models. In this paper, we propose an adapted hierarchical network framework that bridges this modality gap without extensive training. Our method initializes byte embeddings directly from the subword representations of a frozen base model. We apply a chunk alignment loss to project dynamically grouped byte chunks toward precomputed subword targets, and interleave lightweight part-of-speech (POS) supervision to guide boundary detection. Experiments across six languages demonstrate that our tokenizer-free approach improves performance for word-level morphological tasks, yielding up to a 13.3% improvement on POS tagging.

arXiv

Quantization-Triggered Backdoors in Language Models: Cross-Quantizer Transferability and the Validation--Deployment Gap

Jacopo Dardini, Claudio Stanzione, Giordano Colò, Giuseppe Fenza

Post-training quantization is often treated as a semantically neutral optimization for edge deployment of Large Language Models. When a full-precision source checkpoint is evaluated and quantization is applied downstream without equivalent re-evaluation, this workflow creates a structural validation--deployment gap: because quantization is a many-to-one mapping over parameter space, source-precision certification does not guarantee behavioral equivalence in the deployed configuration. We formalize this gap through Quantization Behavioral Equivalence Classes (QBECs) and prove that QBEC membership does not imply behavioral equivalence, providing a theoretical basis for quantization-triggered backdoor attacks. Building on a three-stage adversarial fine-tuning framework, we embed latent malicious payloads into models that satisfy the source-precision checks used in our evaluation, yet activate targeted adversarial behavior upon INT8 or 4-bit compression. We evaluate this threat in two operationally motivated scenarios, tactical machine translation and political content analysis, extending prior work from decoder-only causal LMs to multilingual encoder-decoder sequence-to-sequence models. Results show that backdoored translation models move from zero measured friend--foe corruption at repaired FP16 to up to 85.02% inversion after quantization, and that a paired stance classifier measures an ideological shift of up to $Δ\mathrm{Bias}=0.33$ upon compression. A cross-quantizer transferability analysis further shows that attack persistence varies across quantization schemes and model architectures, rather than being determined by nominal bit-width alone. These findings demonstrate that source-precision auditing alone does not rule out quantization-triggered behavior and that the final deployed configuration must be included in behavioral certification for trustworthy edge AI.

arXiv

How Do Linear Probes Emerge? A Circuit-Tracing Framework with Concept-Targeted Attribution

Vedant Palit, Florent Draye, Terry Jingchen Zhang, Bernhard Schölkopf, Zhijing Jin

Transcoder attribution graphs are usually trained to explain why a model assigns high probability to a particular next token. We introduce Concept-Targeted Attribution (CTA), which instead trains attribution graphs with respect to a linear probe direction. CTA therefore yields probe-specific circuits that explain why an internal concept representation arises in a prompt, independently of whether it is expressed in the generated token. Using Cross-Layer Transcoders, we show that these probe-targeted graphs contain predictive structure: graph-level features predict probe accuracy across four widely studied concept categories ($ρ= 0.91$, $R^2 = 0.84$), while local features identify the sparse components driving per-prompt classification. This connects probe performance to interpretable circuit structure, allowing us to ask not only whether a probe works, but which internal computations make it work. Causal ablations further show that probe-targeted and logit-targeted graphs capture functionally distinct mechanisms. Removing probe-relevant features reduces internal concept scores while largely preserving generated tokens, whereas removing logit-relevant features changes the generated token in 92% to 100% of cases with near-zero effect on probe scores. CTA provides a framework for moving from behavioral probe accuracy to mechanistic explanations of probe performance, enabling more detailed audits of internal concept representations, including safety-critical ones. Our code is available at https://github.com/vedantpalit/concept-targeted-attribution

arXiv

INSPIRE: An Internalize-Then-Improve Approach for Example-Driven Mathematical Reasoning

Shuai Wang, Jiayi Kuang, Yinghui Li, Haojing Huang, Xinnian Liang, Ying Shen, Liang Lin

Mathematical reasoning has seen rapid progress in large language models (LLMs), yet existing methods optimize predominantly for final-answer correctness, raising the question whether models truly internalize mathematical concepts or merely memorize solution patterns. In human mathematics education, example-based reasoning such as constructing counterexamples to test theorem boundaries reflects deep conceptual understanding, but remains underdeveloped in current LLMs. Enhancing this capability through preference optimization presents two key challenges: (1) the model's limited example-based reasoning ability makes constructing effective preference pairs inherently difficult; and (2) capability acquisition is progressive, as the model must first learn to adopt this strategy before learning to apply it correctly. Therefore we propose INSPIRE, an Internalize-Then-Improve approach combining Reference-Guided Student Internalization (RGSI), which produces high-quality preference candidates under the policy model's own distribution, with a stage-wise rubric preference training strategy that decomposes learning into method-oriented and correctness-oriented stages. Experiments across multiple model scales and families demonstrate consistent improvements, even surpassing larger open-source models, while evaluations on out-of-distribution benchmarks confirm no degradation in general mathematical reasoning ability.

arXiv

An LLM-Based Framework for Intent-Driven Network Topology Design

Kholoud El-Habbouli, Fen Zhou, Stephane Huet

Designing deployable and resilient network topologies from natural language requirements remains a challenging problem in network automation. This work investigates the ability of Large Language Models (LLMs) to generate structurally valid and constraint-compliant network topologies through a constraint-driven pipeline combining hierarchical modeling and systematic validation. The framework is evaluated via a multimodel comparison of proprietary and open-weight LLMs across four realistic network scenarios released as a public dataset. We assess structural correctness using node and edge F1-scores against reference topologies, and evaluate resilience through server and content connectivity metrics. In addition, we analyze common failure modes, including interface mismatches and directional inconsistencies in generated topologies. Overall, this work provides a systematic benchmark for understanding how LLMs handle structural and resilience constraints in topology synthesis, and supports informed model selection for AI-driven network design.

arXiv

CASPER in the Machine: Insights into Character Variety in LLM-Generated Stories

Anneliese Brei, Abhisheik Sharma, Nicholas Sanaie, Lu Wang, Snigdha Chaturvedi

As LLM-generated text is increasingly used, especially in fictional domains, we explore how much LLM-generated stories differ from human-written stories. In this work, we focus on characters. We borrow definitions from narratology to analyze eight intricate dimensions of character, such as stylization and wholeness. These dimensions consider more than just basic characteristics. They assess how characters are portrayed within their stories. After automatically inferring categories of characters within both LLM and human-written stories, we compare and contrast these two sets of stories. We consider the following overarching questions: (1) Do LLMs and human-written stories have similar characters? and (2) Do LLMs generate stories with a variety of characters? Our analysis includes research questions that focus on stories generated by popular LLMs and recently published human-written stories. We describe a number of interesting similarities, differences and key takeaways.

arXiv

Persuasion Index: A Theory-Guided Framework for Persuasion Analysis

Liancheng Gong, Zhiyang Wang, Yiwei Xu, Julia Mendelsohn

Identifying persuasive rhetorical cues is critical across domains, from detecting information manipulation and improving AI safety to advancing public health communication. We propose the Persuasion Index (PI), a taxonomy of 15 dimensions grounded in persuasion theories from psychology and communication, and one transparent implementation using 55 sub-features built from lexicons and rule-based detectors. The taxonomy is modular: individual detectors can be replaced while preserving the theoretical structure. We evaluate PI on four public datasets for English argumentative text that vary in domain, style, and outcome measures, and show that PI provides a shared feature space for interpreting rhetorical patterns associated with persuasion-related outcomes. Linear models show that PI features carry meaningful predictive signal while remaining computationally lightweight. Dimension-level analyses reveal recurring associations between PI dimensions and persuasion outcomes across datasets, while also highlighting topic- and stance-specific variation. We release PI as an open-source package and web interface for principled and auditable analysis of human and AI-mediated communication.

arXiv

Do LLM Agents Mirror Socio-Cognitive Effects in Power-Asymmetric Conversations?

Anvesh Rao Vijjini, Sagar Manjunath, Snigdha Chaturvedi

Power differences shape human communication through well documented socio cognitive effects, including language coordination, pronoun usage, authority bias, and harmful compliance. We examine whether large language models (LLMs) exhibit similar behaviors when assigned high or low status personas. Using personas from diverse professions, we simulate multi turn, power asymmetric dialogues (e.g., principal teacher, justice lawyer) and measure (i) language coordination, (ii) pronoun usage, (iii) persuasion success, and (iv) compliance with unsafe requests. Our results show that LLMs show key socio-cognitive effects of power, albeit with nuances and variability, linking simulated interactions to both desirable and unsafe behaviors.

arXiv

G-Loss: Graph-Guided Fine-Tuning of Language Models

Aditya Sharma, Vinti Agarwal, Rajesh Kumar

Traditional loss functions, including cross-entropy, contrastive, triplet, and su pervised contrastive losses, used for fine-tuning pre-trained language models such as BERT, operate only within local neighborhoods and fail to account for the global semantic structure. We present G-Loss, a graph-guided loss function that incorporates semi-supervised label propagation to use structural relationships within the embedding manifold. G-Loss builds a document-similarity graph that captures global semantic relationships, thereby guiding the model to learn more discriminative and robust embeddings. We evaluate G-Loss on five benchmark datasets covering key downstream classification tasks: MR (sentiment analysis), R8 and R52 (topic categorization), Ohsumed (medical document classification), and 20NG (news categorization). In the majority of experimental setups, G-Loss converges faster and produces semantically coherent embedding spaces, resulting in higher classification accuracy than models fine-tuned with traditional loss functions.

arXiv

The Instability of Safety: How Random Seeds and Temperature Expose Inconsistent LLM Refusal Behavior

Erik Larsen

Current safety evaluations of large language models rely on single-shot testing, implicitly assuming that model responses are deterministic and representative of the model's safety alignment. We challenge this assumption by investigating the stability of safety refusal decisions across random seeds and temperature settings. Testing four instruction-tuned models from three families (Llama 3.1 8B, Qwen 2.5 7B, Qwen 3 8B, Gemma 3 12B) on 876 harmful prompts across 20 sampling configurations (4 temperatures x 5 seeds), we find that 18-28% of prompts exhibit decision flips--the model refuses in some configurations but complies in others--depending on the model. Our Safety Stability Index (SSI) reveals that higher temperatures significantly reduce decision stability (Friedman chi-squared = 396.81, p < 0.001), with mean within-temperature SSI dropping from 0.977 at temperature 0.0 to 0.942 at temperature 1.0. We validate findings across all model families using Claude 3.5 Haiku as a unified external judge, achieving 89.1% inter-judge agreement with the Llama 70B judge on the two models both judges labeled (Cohen's kappa = 0.62). Within each model, prompts with higher compliance rates exhibit lower stability (Spearman rho = -0.47 to -0.70, all p < 0.001), indicating that models "waver" more on borderline requests. These findings demonstrate that single-shot safety evaluations are insufficient for reliable safety assessment and that evaluation protocols must account for stochastic variation in model behavior. For Llama 3.1 8B, single-shot evaluation agrees with multi-sample ground truth only 92.5% of the time when pooling across temperatures (98.7% at greedy to 90.3% at temperature 1.0), and we recommend scaling samples with temperature--one at greedy, three at low temperature, more at higher temperatures (where three reach only ~95%), and ten when pooling--rather than a single flat threshold.

arXiv

SMRC: Aligning Large Language Models with Student Reasoning for Mathematical Error Correction

Biaojie Zeng, Min Zhang, Juan Zhou, Fengrui Liu, Ruiyang Huang, Yu Song, Xin Lin

Large language models (LLMs) often make reasoning errors when solving mathematical problems, and how to automatically detect and correct these errors has become an important research direction. However, existing approaches \textit{mainly focus on self-correction within the model}, which falls short of the "teacher-style" correction required in educational settings, \textit{i.e.}, systematically guiding and revising a student' s problem-solving process. To address this gap, we propose \texttt{SMRC} (\textit{\underline{S}tudent \underline{M}athematical \underline{R}easoning \underline{C}orrection}), a novel method that aligns LLMs with student reasoning. Specifically, \texttt{SMRC} formulates student reasoning as a multi-step sequential decision problem and introduces Monte Carlo Tree Search (MCTS) to explore optimal correction paths. To reduce the cost of the annotating process-level rewards, we leverage breadth-first search (BFS) guided by LLMs and final-answer evaluation to generate reward signals, which are then distributed across intermediate reasoning steps via a back-propagation mechanism, enabling fine-grained process supervision. Additionally, we construct a benchmark for high school mathematics, MSEB (Multi-Solution Error Benchmark), consisting of 158 instances that include problem statements, student solutions, and correct reasoning steps. We further propose a dual evaluation protocol centered on \textbf{solution accuracy} and \textbf{correct-step retention}, offering a comprehensive measure of educational applicability. Experiments demonstrate that \texttt{SMRC} significantly outperforms existing methods on two public datasets (ProcessBench and MR-GSM8K) and our MSEB in terms of effectiveness and overall performance. The code are available at https://github.com/ECNU-RAIL/SMRC-EMNLP2026.

arXiv

PRISM: Agentic Retrieval with LLMs for Multi-Hop Question Answering

Md Mahadi Hasan Nahid, Davood Rafiei

Retrieval plays a central role in multi-hop question answering (QA), where answering complex questions requires gathering multiple pieces of evidence. We propose PRISM, an agentic retrieval framework that leverages large language models (LLMs) in a structured loop to retrieve relevant evidence with high precision and recall. PRISM decomposes retrieval into three specialized agents: a Question Analyzer that breaks complex queries into sub-questions, a Selector that identifies the most relevant context for each sub-question (focusing on precision), and an Adder that brings in any missing evidence (focusing on recall). The iterative interaction between the Selector and Adder produces a compact yet comprehensive evidence set, avoiding both brittle error propagation and noisy context accumulation. It achieves higher retrieval accuracy while filtering out distracting content, enabling downstream QA models to surpass full-context answer accuracy while relying on significantly less irrelevant information. Experiments on four challenging multi-hop QA benchmarks, including HotpotQA, 2WikiMultiHopQA, MuSiQue, and MultiHopRAG, demonstrate that our approach consistently outperforms strong baselines.

arXiv

The Telephone Game: Evaluating Semantic Drift in Unified Models

Sabbir Mollah, Rohit Gupta, Sirnam Swetha, Qingyang Liu, Ahnaf Munir, Mubarak Shah

Unified models (UMs) combine visual understanding (I2T) and generation (T2I) in a single framework. We focus on T2I and I2T, where cross-consistency---what a model understands, it should be able to generate---is a promise of unification and a necessity when composing both capabilities. Yet, existing benchmarks evaluate them in isolation: FID/GenEval for T2I; MME/MMBench for I2T. We show this gap is consequential: models scoring competitively on these benchmarks can fail severely when understanding and generation are composed, losing entities, attributes, spatial relations, and counts, resulting in semantic drift. To quantify drift, we introduce the Semantic Drift Protocol (SDP), inspired by the Telephone Game: starting from a caption or image, we alternate I2T and T2I over multiple generations and measure semantic preservation. We propose Mean Cumulative Drift (MCD), an embedding-based measure of content retention across three representation spaces, and Multi-Generation GenEval (MGG), extending GenEval's object-level compliance scoring across generations. To stress-test models beyond COCO-style data, we create a benchmark of 400 image-text pairs sampled from NoCaps and DOCCI, emphasizing novel objects and fine-grained descriptions. Applying SDP to seven models reveals that drift varies dramatically and is not predicted by single-pass scores: BAGEL retains high semantic fidelity over multiple generations, while VILA-U and Janus variants collapse within five generations, despite comparable isolated metrics. We identify six recurring failure modes and find degradation is typically catastrophic rather than gradual: once a critical error occurs, subsequent generations compound it. SDP exposes failure modes that single-pass benchmarks miss, enabling a more faithful assessment of unified model reliability. Code and benchmark: https://github.com/mollahsabbir/telephone-game-semantic-drift

arXiv

Automatic Pronunciation Error Detection and Correction of the Holy Quran's Learners Using Deep Learning

Abdullah Abdelfattah, Mahmoud I. Khalil, Hazem Abbas

Assessing spoken language is challenging, and quantifying pronunciation metrics for machine learning models is even harder. However, for the Holy Quran, this task is enabled by the rigorous recitation rules (Tajweed) established through the efforts of Muslim scholars, making highly effective assessment possible. Despite this advantage, the scarcity of high-quality annotated data remains a significant barrier. In this work, we bridge these gaps by introducing: (1) A 98% automated pipeline to produce high-quality Quranic datasets -- encompassing collection of recitations from expert reciters, segmentation at pause points (waqf) using our fine-tuned wav2vec2-BERT model, transcription of segments, and transcript verification via our novel Tasmeea algorithm; (2) 848 hours of audio (286K annotated utterances); (3) qdat_bench, a benchmark covering phonemes, diacritization, and Tajweed rules (Ghunnah, Qalqalah, Madd) on real recitation errors containing 159 samples; (4) A novel ASR-based approach for pronunciation error detection utilizing our custom Quran Phonetic Script (QPS) to encode Tajweed rules (unlike the IPA standard for Modern Standard Arabic). QPS uses an 11-level script: phoneme level (encoding Arabic letters with short/long vowels) and sifat level (encoding articulation characteristics of every phoneme). We further present comprehensive modeling with our novel multi-level CTC model, which achieved 0.21% and 1.94% average Phoneme Error Rate (PER) on the test set and qdat_bench respectively, with a 75.8% Tajweed F1 score. We release our work as open-source: https://obadx.github.io/quran-muaalem/en/

arXiv

Pruning Laws for Large Language Models

Ayan Sengupta, Siddhant Chaudhary, Tanmoy Chakraborty

Scaling up model parameters and training data consistently improves the performance of large language models (LLMs), but at the cost of rapidly growing memory and compute requirements, which makes deployment on resource-limited hardware infeasible. Model pruning, a widely used compression technique, reduces inference costs by removing redundant parameters. However, its impact on downstream performance remains unpredictable and is typically assessed only through costly empirical sweeps. To address this gap, we introduce pruning laws, simple and interpretable scaling relations that connect a pruned LLM's post-pruning performance to its unpruned performance and pruning ratio. Across ten LLMs (1.3B-30B parameters), a 20B mixture-of-experts model, three pruning strategies (unstructured, width, and depth), and eight diverse tasks, we show that pruning laws achieve strong predictive accuracy (average extrapolation error less than 7%), reliably quantify performance degradation, and identify critical pruning thresholds beyond which recovery is infeasible. Moreover, we demonstrate that the functional form transfers across dense and mixture-of-experts architectures, pruning methods, and unseen models in zero-shot and one-shot setups, with task- and method-specific coefficients that vary in interpretable ways. These results provide both researchers and practitioners with a principled framework to select pruning strategies, estimate safe pruning ratios without exhaustive tuning, and deploy LLMs efficiently under real-world compute and latency constraints.

arXiv

CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes

Yufan Wu, Yinghui He, Zhengyi Hu, Lang Wei, Ruichen Li, Qifan Yang, Ting Zhu

Recent advances in inference-time scaling have significantly improved the reasoning performance of large language models (LLMs). However, these methods typically rely on repeated generation or external verification. To address this limitation, we introduce CritICL, a novel inference-time framework that improves reasoning while maintaining high efficiency. Our key insight is that LLM failure modes exhibit structured patterns across model scales within the same family. Instead of treating failures as undesirable outputs, CritICL leverages them as a source of guidance. Specifically, we utilize failure modes derived from weaker models and incorporate them into inference through critique-based in-context examples. We propose two variants: CritICL-dynamic, which adaptively predicts input-specific failure modes and retrieves critiques, and CritICL-static, which uses a global failure mode profile to provide stable guidance. Experimental results show that CritICL consistently outperforms standard in-context learning and achieves performance competitive with or superior to test-time scaling methods, while requiring significantly fewer generations and lower token cost. Code available at: https://github.com/umwyf/CRITICL

arXiv

WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution

Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, Tu Vu

Agent skills package specialized knowledge and workflows into reusable resources that extend AI agent capabilities. Recent work automatically discovers such skills from agent experience, which enables agents to progressively adapt through interaction. However, the insights that guide skill development typically remain scattered across optimization histories, limiting their systematic reuse across iterations. We introduce WikiSkill, a framework that co-evolves agent skills with a persistent knowledge base (wiki). At a high level, WikiSkill separates raw execution experience, accumulated knowledge, and executable skills, while continuously consolidating experience into the wiki, which subsequent skill updates can build on. Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods and improves over no-skill baselines in most model-benchmark settings. We find that skill evolution complements model scaling: larger models generally benefit more from evolved skills, while smaller models with skills can outperform substantially larger models without them. We also find that evolved skills transfer effectively across models and model families, and skills evolved by other models can outperform self-evolved skills. Finally, our ablation studies confirm that persistent knowledge accumulation in the wiki is critical for effective skill evolution. These results demonstrate the benefits of systematically accumulating and refining agent experience for developing reusable and transferable skills.

arXiv

SWE-Prime: Fewer Trajectories, Better Performance

Dewu Zheng, Ruizhe Ye, Yanlin Wang, Yang Ye, Hongyu Zhang, Ensheng Shi, Xilin Liu, Yuchi Ma, Jianxing Yu, Zibin Zheng

To improve large language models' ability to resolve real-world software issues, prior work has focused on constructing large-scale agent trajectory datasets and performing supervised fine-tuning (SFT) on successful trajectories. However, task success does not guarantee high-quality supervision: successful trajectories may still contain ineffective, redundant, or risky steps. Directly using such trajectories for SFT can introduce noisy supervision and encourage models to imitate undesirable problem-solving behaviors. Therefore, we propose SWE-Prime, a multi-granularity, two-stage SFT data selection method that progressively filters training data at the trajectory and segment levels. Specifically, the first stage performs trajectory-level screening based on process quality, result quality, and data representativeness, selecting a high-quality and representative subset of successful trajectories. The second stage performs segment-level selection by grouping consecutive steps into semantic segments and assessing each segment based on its contribution to the final solution, learnability, and potential risks. During SFT, all segments remain in the sequence to preserve context, while only selected segments contribute to the loss computation. Experiments on SWE-Bench Pro and SWE-Bench Verified show that training on the 10% trajectory subset selected by SWE-Prime outperforms training on the full resolved dataset, yielding relative performance gains of up to 12.2% and 24.2%, respectively.

Keep browsing

More papers from top AI conferences

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