论文检索

Michael Lan, Narmeen Fatimah Oozeer, Chaithanya Bandi, Philip Quirke, Austin Meek, Fazl Barez, Amir Abdullah

While mechanistic interpretability (MI) has produced important insights into neural network internals, the field has yet to establish a standardized system to audit experiments. As such, many of its findings remain underutilized in safety-critical applications such as medical AI and autonomous systems, as stakeholders cannot certify their validity. Recent work demonstrates this concretely: two papers found conflicting conclusions for the same behavior, and a third study revealed that both were partially correct but incomparable due to methodological inconsistencies. Without standardized auditing, such ambiguities hinder adoption in high-stakes contexts requiring strong correctness guarantees. We call for the MI community to work towards developing a novel reviewing system that complements peer review via: (1) Continuous reviewing supported by a Collaborative Reviewing Platform where meta-science results and discussions (such as critiques, negative results, post-hoc extensions, reproductions, replications, and partial results) that fit outside of papers are organized and discussed, allowing for comments and revisions to be made at any time (2) Generalizing good practices found on this platform into expert-verified guidelines and protocols to improve auditing efficiency, and (3) Source-based auditing systems that track arguments which claims depend on. This position paper encourages constructive debate over the necessity, design and implementation of such a framework, providing early concrete examples to help catalyze these dialogues. Overall, we propose that auditing MI itself is essential for its application in AI safety, industry, and governance.

Minjae Lee, Minhyuk Seo, Tingyu Qu, Tinne Tuytelaars, Jonghyun Choi

In continual instruction tuning (CIT) scenarios, where new instruction tuning data continuously arrive in an online streaming manner, training delays from large-scale data significantly hinder real-time adaptation. Data selection can mitigate this overhead, but existing strategies often rely on pre-trained reference models, which are impractical in CIT setups since future data are unknown. Recent reference model-free online sample selection methods address this, but typically select a fixed number of samples per batch (e.g., top-k), making them vulnerable to distribution shifts where informativeness varies across batches. To address these limitations, we propose OASIS, an adaptive online sample selection approach for CIT that (1) selects informative samples by estimating each sample’s informativeness relative to all previously seen data, beyond batch-level constraints, and (2) minimizes informative redundancy of selected samples through iterative selection score updates. Experiments on various large foundation models show that , using only 25% of the data, achieves comparable performance to full-data training and outperforms the state-of-the-art sampling methods.

Mingtian Tan, Mike A Merrill, Zachary Gottesman, Tim Althoff, David Evans, Thomas Hartvigsen

A common goal in analyzing time series data is to understand how events cause observed variations. We study whether Large Language Models (LLMs) can infer natural language events associated with time series data.We introduce an automated method for generating tasks that test a model’s ability to reason about events associated with time series data based on sports data, and develop a new benchmarking method. In experiments spanning 18 LLMs, we prompt LLMs to infer unobserved events given time series data and observe surprising successes, even when providing minimal context. We then show that combining distillation with Reinforcement Learning (RL) can improve the performance for small language models to approach that of large proprietary reasoning models. All resources needed to reproduce our work are available: https://github.com/hartvigsen-group/GAMETime.

Ido Andrew Atad, Itamar Zimerman, Shahar Katz, Lior Wolf

Attention matrices are fundamental to transformer research, supporting a broad range of applications including interpretability, visualization, manipulation, and distillation. Yet, most existing analyses focus on individual attention heads or layers, failing to account for the model’s global behavior. While prior efforts have extended attention formulations across multiple heads via averaging and matrix multiplications or incorporated components such as normalization and FFNs, a unified and complete representation that encapsulates all transformer blocks is still lacking. We address this gap by introducing TensorLens, a novel formulation that captures the entire transformer as a single, input-dependent linear operator expressed through a high-order attention-interaction tensor. This tensor jointly encodes attention, FFNs, activations, normalizations, and residual connections, offering a theoretically coherent and expressive linear representation of the model’s computation. TensorLens is theoretically grounded and our empirical validation shows that it yields richer representations than previous attention-aggregation methods. Our experiments demonstrate that the attention tensor can serve as a powerful foundation for developing tools aimed at interpretability and model understanding.

Ruixuan Deng, Xiaoyang Hu, Miles Gilberti, Shane Storks, Aman Taxali, Mike Angstadt, Chandra Sripada, Joyce Chai

We identify semantically coherent, context-consistent network components in large language models (LLMs) using coactivation of sparse autoencoder (SAE) features collected from just a handful of prompts. Focusing on concept-relation prediction tasks, we show that ablating these components for concepts (e.g., countries and words) and relations (e.g., capital city and translation language) changes model outputs in predictable ways, while amplifying these components induces counterfactual responses. Notably, composing relation and concept components yields compound counterfactual outputs. Further analysis reveals that while most concept components emerge from the very first layer, more abstract relation components are concentrated in later layers. Lastly, we show that extracted components more comprehensively capture concepts and relations than individual features while maintaining specificity. Overall, our findings suggest a modular organization of knowledge and advance methods for efficient, targeted LLM manipulation.

Zirui Yan, Dennis Wei, Dmitriy A Katz, Prasanna Sattigeri, Ali Tajer

Causal tracing systematically intervenes on a large language model’s (LLM’s) internal representations to uncover and quantify the causal pathways linking specific inputs or computations to specific metrics of interest, quantifying the LLM’s behavior. Building on previous single-component or single-layer studies, this paper presents a unified framework for causally tracing multiple components simultaneously. This framework systematically identifies the subsets of components (e.g., attention heads and multi-layer perceptron neurons) most critical to a desired target performance metric (e.g., accuracy and fairness). This is achieved by incorporating flexible interventions applied to a wide range of desired metrics. To address the combinatorial complexity of the multi-component problem, an efficient algorithm is designed that leverages soft interventions and a carefully designed metric transformation, converting the combinatorial search problem into a continuous one that can be solved efficiently under proper constraints, thereby generating proper binary decisions for selecting components. Experimental results demonstrate that the proposed method efficiently identifies subsets of the model’s components that have a high impact on the target metric, outperforming existing baseline approaches.

Abdelrahman Abdallah, Mohammed Ali, Bhawna Piryani, Adam Jatowt

Although Large Language Models (LLMs) show strong potential for zero-shot document ranking, current listwise approaches face three critical limitations. First, context length constraints prevent processing many documents simultaneously. Second, sequential generation creates bottlenecks that cannot run in parallel. Third, ranking results depend heavily on initial document order, leading to inconsistent performance. We introduce BracketRank, a reasoning-driven competitive elimination framework that addresses these challenges through systematic group competition. Our method uses adaptive grouping to automatically optimise group sizes based on LLM context limits, reasoning-enhanced prompts that require explicit relevance explanations, and bracket-style elimination where documents compete through winner and loser brackets. This structure ensures every document has fair advancement opportunities regardless of initial positioning while allowing for parallel processing across competition stages. We evaluate BracketRank on TREC DL 19, TREC DL 20, and eight BEIR benchmark datasets. Results show that BracketRank achieves 77.90 NDCG@5 on TREC DL 19 and 75.85 NDCG@5 on TREC DL 20, outperforming RankGPT and other state-of-the-art methods. On BEIR datasets, BracketRank reaches 54.66 average NDCG@10, demonstrating robust performance across diverse domains.

Guangya Wan, Mingyang Ling, Xiaoqi Ren, Rujun Han, Sheng Li, Zizhao Zhang

Long-horizon tasks that require sustained reasoning and multiple tool interactions remain challenging for LLM agents: small errors compound across steps, and even state-of-the-art models often hallucinate or lose coherence. We identify context management as the central bottleneck—extended histories cause agents to overlook critical evidence or become distracted by irrelevant information, thus failing to replan or reflect from previous mistakes. To address this, we propose COMPASS (Context-Organized Multi-Agent Planning and Strategy System), a lightweight hierarchical framework that separates tactical execution, strategic oversight, and context organization into three specialized components: (1) a Main Agent that performs reasoning and tool use, (2) a Meta-Thinker that monitors progress and issues strategic interventions, and (3) a Context Manager that maintains concise, relevant progress briefs for different reasoning stages. Across three challenging benchmarks—GAIA, BrowseComp, and Humanity’s Last Exam—COMPASS improves accuracy by up to 20% relative to both single- and multi-agent baselines. We further introduce a test-time scaling extension that elevates performance to match established DeepResearch agents, and a post-training pipeline that delegates context management to smaller models for enhanced efficiency.

Nikita Tatarinov, Vidhyakshaya Kannan, Haricharana Srinivasa, Arnav Raj, Harpreet Singh Anand, Varun Singh, Aditya Luthra, Ravij Lade, Agam Shah, Sudheer Chava

We introduce KG-MuLQA (Knowledge-Graph-based Multi-Level Question-Answer Extraction): a framework that (1) extracts QA pairs at multiple complexity levels (2) along three key dimensions – multi-hop retrieval, set operations, and answer plurality, (3) by leveraging knowledge-graph-based document representations. This approach enables fine-grained assessment of model performance across controlled difficulty levels. Using this framework, we construct a dataset of 20,139 QA pairs based on financial credit agreements and evaluate 16 proprietary and open-weight Large Language Models, observing that even the best-performing models struggle with set-based comparisons and multi-hop reasoning over long contexts. Our analysis reveals systematic failure modes tied to semantic misinterpretation and inability to handle implicit relations.

Jinu Lee, Kyoung-Woon On, Sophia Simeng Han, Arman Cohan, Julia Hockenmaier

Evaluating the quality of LLM-generated reasoning traces in expert domains (e.g., law) is essential for ensuring credibility and explainability, yet remains challenging due to the inherent complexity of such reasoning tasks. We introduce LEGIT (LEGal Issue Trees), a novel large-scale (24K instances) expert-level legal reasoning dataset with an emphasis on reasoning trace evaluation. We convert court judgments into hierarchical trees of opposing parties’ arguments and the court’s conclusions, which serve as rubrics for evaluating the issue coverage and correctness of the reasoning traces. We verify the reliability of these rubrics via human expert annotations and comparison with coarse, less informative rubrics. Using the LEGIT dataset, we show that (1) LLMs’ legal reasoning ability is seriously affected by both legal issue coverage and correctness, and that (2) retrieval-augmented generation (RAG) and RL with rubrics bring complementary benefits for legal reasoning abilities, where RAG improves overall reasoning capability, whereas RL improves correctness albeit with reduced coverage.

Adarsh Singh, Kushal Raj Bhandari, Jianxi Gao, Soham Dan, Vivek Gupta

Open-Domain Table Question Answering (TQA) involves retrieving relevant tables from a large corpus to answer natural language queries. Traditional dense retrieval models, such as DTR and DPR, not only incur high computational costs for large-scale retrieval tasks but also require retraining or fine-tuning on new datasets, limiting their adaptability to evolving domains and knowledge. In this work, we propose **CRAFT**, a zero-shot, cascaded retrieval approach that first uses a sparse retrieval model to filter a subset of candidate tables before applying more computationally expensive dense models as re-rankers.To improve retrieval quality, we enrich table representations with descriptive titles and summaries generated by *Gemini Flash 1.5*, enabling richer semantic matching between queries and tabular structures. Our method outperforms state-of-the-art (SOTA) sparse, dense, and hybrid retrievers on the NQ-Tables dataset. It also demonstrates strong zero-shot performance on the more challenging OTT-QA benchmark, achieving competitive results at higher recall thresholds, where the task requires multi-hop reasoning across both textual passages and relational tables.This work establishes a scalable and adaptable paradigm for table retrieval, bridging the gap between fine-tuned architectures and lightweight, plug-and-play retrieval systems. Code and data are available at: [https://coral-lab-asu.github.io/CRAFT/](https://coral-lab-asu.github.io/CRAFT/)

Zhenhua Liu, Lijun Li, Ruizhe Chen, Yuxian Jiang, Tong Zhu, Zhaochen Su, Wenliang Chen, Jing Shao

While guided decoding, especially value-guided methods, has emerged as a cost-effective alternative for controlling language model outputs without re-training models, its effectiveness is limited by the accuracy of the value function. We identify that this inaccuracy stems from a core distributional gap: existing methods train static value functions on trajectories sampled exclusively from the base policy, which inherently confines their training to a narrow and suboptimal view of the potential output space. We propose Iterative Value Refinement, a novel framework designed to bridge this gap. It employs Value Exploration to provide a more comprehensive and robust training signal, complemented by Iterative Self-Refinement, which uses the improved value function from one iteration to guide the generation of higher-quality data for the next. Extensive experiments on text summarization, multi-turn dialogue, and instruction following demonstrate the effectiveness of our framework in aligning language models. Our approach not only achieves alignment but also significantly reduces computational costs by leveraging principled value function optimization for efficient and effective control.

Yupeng Hou, Jiacheng Li, Xiangjun Fu, Zhankui He, An Yan, Xiusi Chen, Julian McAuley

Feature engineering has long been central to recommender systems, yet effectively leveraging textual item features remains challenging. Recent advances in large language models (LLMs) have enabled their use as semantic encoders for recommendation, but their roles and behaviors in this setting are still not well understood. Prior studies often rely on general-purpose embedding benchmarks (e.g., MTEB) when selecting LLMs, overlooking the unique characteristics of recommendation tasks. To address this gap, we introduce BLaIR, a comprehensive benchmark for evaluating LLMs as semantic encoders in recommendation scenarios. We contribute (1) a new large-scale Amazon Reviews 2023 dataset with over 570 million reviews and 48 million items, (2) a unified benchmark covering sequential recommendation, collaborative filtering, and product search, and (3) a new complex-query product search task featuring both semi-synthetic and real-world evaluation datasets. Experiments with 11 leading LLMs show that their rankings on BLaIR show little correlation with MTEB, highlighting the unique challenges of semantic encoding in recommendation.

Zhengxiang Cheng, Dongping Chen, Mingyang Fu, Tianyi Zhou

Large Reasoning Models (LRMs) have achieved remarkable success, yet they often suffer from producing unnecessary and verbose reasoning chains. We identify a core aspect of this issue as ”invalid thinking”— models tend to repeatedly double-check their work after having derived the correct answer. To address this specific inefficiency, we move beyond the general principles of Efficacy and Efficiency to propose two new, fine-grained principles: Brevity, which advocates for eliminating redundancy, and Sufficiency, which ensures critical reasoning steps are preserved. Guided by these principles, we introduce LC-R1, a post-training method based on Group Relative Policy Optimization (GRPO). LC-R1 employs a novel combination of a Length Reward for overall conciseness and a Compress Reward that is specifically designed to remove the invalid portion of the thinking process. Extensive experiments on multiple reasoning benchmarks demonstrate that LC-R1 achieves a significant reduction in sequence length (5̃0%) with only a marginal (2̃%) drop in accuracy, achieving a favorable trade-off point on the Pareto frontier that prioritizes high compression. Our analysis further validates the robustness of LC-R1 and provides valuable insights for developing more powerful yet computationally efficient LRMs.

Dahyun Jung, Jaewook Lee, Heuiseok Lim

Large language models (LLMs) require frequent knowledge updates to reflect changing facts and mitigate hallucinations. To meet this demand, lifelong knowledge editing has emerged as a continual approach to modify specific pieces of knowledge without retraining the entire model. Existing parameter-editing methods struggle with stability during sequential edits due to catastrophic forgetting. While retrieval-based approaches are proposed to alleviate this issue, their applicability remains limited across various datasets because of high training costs. To address these limitations and enhance scalability in lifelong settings, we propose LightEdit. Our framework first selects relevant knowledge from retrieved information to modify the query effectively. It then incorporates a decoding strategy to suppress the model’s original knowledge probabilities, thereby enabling efficient edits based on the selected information. Extensive experiments on ZSRE, Counterfact, and RIPE benchmarks demonstrate that LightEdit outperforms existing lifelong knowledge editing methods. Furthermore, by minimizing training costs, LightEdit achieves cost-effective scalability, enabling easy adaptation to various datasets.

Jason S Lucas, Matt Murtagh White, Ali Al-Lawati, Uchendu Uchendu, Adaku Uchendu, Dongwon Lee

Harmful content detectors—particularly disinformation classifiers—are predominantly developed and evaluated on Standard American English (), leaving their robustness to dialectal variation unexplored. We present , the first benchmark for evaluating disinformation detection robustness across 50 English dialects spanning U.S., British, African, Caribbean, and Asia-Pacific varieties. Using Multi-VALUE’s linguistically-grounded transformations, we introduce D-CUBE (Dialectal Disinformation Detection Corpus), a core corpus component of comprising 195K samples derived from established disinformation benchmarks. Our evaluation of 16 detection models reveals systematic vulnerabilities: human-written dialectal content degrades detection by 1.4–3.6% F1, while AI-generated content remains stable. Fine-tuned transformers substantially outperform zero-shot LLMs (96.6% vs. 78.3% best-case F1), with some models exhibiting catastrophic failures exceeding 33% degradation on mixed content. Cross-dialectal transfer analysis across 2,450 dialect pairs shows that multilingual models (mDeBERTa: 97.2% average F1) generalize effectively, while monolingual models like RoBERTa and XLM-RoBERTa fail on dialectal inputs. These findings demonstrate that current disinformation detectors may systematically disadvantage hundreds of millions of non- speakers worldwide. We release the benchmark, including the , and evaluation tools.

Junyi Li, Yongqiang Chen, Ningning Ding

Machine unlearning has gained increasing attention in recent years, as a promising technique to selectively remove unwanted privacy or copyrighted information from Large Language Models that are trained on a massive scale of human data. However, the emergence of Large Reasoning Models (LRMs), which emphasize long chain-of-thought (CoT) reasoning to address complex questions, presents a dilemma to unlearning: existing methods either struggle to completely eliminate undesired knowledge from the CoT traces or degrade the reasoning performances due to the interference with the reasoning process. To this end, we introduce Counterfactual Unlearning through iterative Preference Optimization (CiPO), a novel framework that redefines unlearning as the targeted intervention of the CoT reasoning in LRMs. More specifically, given a desired unlearning target answer, CiPO instructs LRMs to generate a logically valid counterfactual reasoning trace for preference tuning. As the LRM adjusts to the counterfactual trace, CiPO iteratively updates the preference learning data to increase the discrepancy from the original model. This iterative loop ensures both desirable unlearning and smooth optimization, effectively mitigating the dilemma. Experiments on challenging benchmarks demonstrate that CiPO excels at unlearning, completely removing knowledge from both the intermediate CoT steps and the final answer, while preserving the reasoning abilities of LRMs.

Kevin Du, Clara Kümpel, Michelle Wastl, Alex Warstadt

Users frequently express their beliefs to large language models (LLMs). In some situations, the LLM should accept these contextual beliefs as true. In others, they should stick to their prior knowledge. Notably, users’ expressions of belief (EoBs) can take linguistically diverse forms—using presuppositions, evidential and certainty markers, or varied tones—each of which may have a different persuasiveness over the LLMs. We introduce a typology to systematically evaluate how different EoBs affect whether models follow context versus prior knowledge. The typology is grounded in four linguistically motivated dimensions: form, evidentiality, epistemic stance, and tone, spanning 17 fine-grained types. By pairing these EoBs with world knowledge facts, we generate controlled EoB–query pairs that isolate the effect of linguistic variation. Using this benchmark, we evaluate 16 LLMs that differ in architecture (Llama3, Qwen3, Gemma3), scale (1B-30B parameters), and training stages (base vs instruct). We identify meaningful variations in response behavior across these axes, e.g., that bigger models and instruction models tend to be less context–following than smaller models and base models. We further identify specific EoBs that statistically significantly persuade LMs more consistently than others. Our work reveals systematic patterns in how linguistic framing affects LLM context integration, with implications for prompt engineering and model robustness.

Diana Abagyan, Alejandro R. Salamanca, Andres Felipe Cruz-Salinas, Kris Cao, Hangyu Lin, Acyr Locatelli, Marzieh Fadaee, Ahmet Üstün, Sara Hooker

Pretraining massively multilingual Large Language Models (LLMs) for many languages at once is challenging due to limited model capacity, scarce high-quality data, and compute constraints. Moreover, the lack of language coverage in the tokenizer makes it harder to address the gap for new languages purely at the post-training stage. In this work, we study what relatively cheap interventions early on in training improve *language plasticity*, or adaptation capabilities of the model post-training to new languages. We focus on tokenizer design and propose using a *universal* tokenizer that is trained for more languages than the primary pretraining languages to enable efficient adaptation in expanding language coverage after pretraining. Our systematic experiments across diverse groups of languages and different training strategies show that a universal tokenizer enables significantly higher language adaptation, with up to 20.2% increase in win rates compared to tokenizers specific to pretraining languages. Furthermore, a universal tokenizer also leads to better plasticity towards languages that are completely unseen in the tokenizer and pretraining, by up to 5% win rate gain. We achieve this adaptation to an expanded set of languages with minimal compromise in performance on the majority of languages included in pretraining.

Zhihao Gong, Zeyu Sun, Dong Huang, Qingyuan Liang, Jie M. Zhang, Dan Hao

While Large Language Models (LLMs) have substantially improved the functional correctness of code translation, the critical dimension of execution efficiency remains overlooked. We present Trace, the first benchmark to explicitly assess efficiency in LLM-translated code. Trace includes 1,000 efficiency-critical tasks across C++, Java, and Python, each augmented with stress tests that reveal efficiency disparities often overlooked by small-scale tests. Using Trace, we conduct an extensive evaluation of 28 representative LLMs and highlight several key insights: 1) Correctness and efficiency are often misaligned: the correctness leader Claude-Sonnet-4-Think achieves only moderate time efficiency, outperformed by smaller open-source LLMs such as Qwen2.5-Coder-14B-Instruct. 2) Inefficiency is both prevalent and patterned: 23.5% of correct translations suffer from notable inefficiency, mainly arising from algorithm implementation discrepancy (11.9%), language construct mismatch (66.4%), and resource management inefficiency (21.7%).3) Inference-time prompt strategies bring only modest improvements, indicating that simple prompting alone is insufficient to improve translation efficiency. Together, our results establish execution efficiency as an essential dimension of code translation and position Trace as a principled foundation for efficiency-oriented evaluation. Our code and data are available at: https://github.com/Albert-Gong/TRACE.