Knowledge editing methods such as ROME and MEMIT update factual associations in transformer models by modifying MLP weights. While evaluated mainly by output behavior, their internal mechanism remains underexplored. We investigate whether edits rely on a common mechanism, regardless of which fact is modified. Despite fact-specific weight changes, we argue that ROME and MEMIT target the same subset of weights critical for maintaining edits. To isolate this subset, we train a compact binary mask over the edited weights. The mask reverses 80% of edits on the training set and over 70% on the test set, confirming that diverse edits share a common functional structure. Our analysis reveals that the mask reverses edits by eliminating overattention in later layers. Additionally, we show that injecting the mask during editing drops editing success from 98% to 38%, demonstrating that this mechanism is necessary for edits to succeed. Our finding that edits suppress rather than overwrite knowledge explains why ROME and MEMIT fail to propagate changes to related facts. The identified common functional subspace informs detection and defense against unwanted edits.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
From Shijing to English and German: Resources and Evaluation for LLM Translation of Early Chinese Poetry
PDF ↗While large language models (LLMs) show promise in literary translation, Shijing (The Book of Songs) serves as a rigorous yet under-explored testbed for testing their limits, given its linguistic antiquity and complex poetic constraints. Automated evaluation in this domain is currently hindered by a scarcity of multilingual resources and the inadequacy of existing metrics in capturing both semantic fidelity and aesthetic quality. In this paper, we bridge these gaps by curating a Shijing parallel corpus with line-by-line Chinese-English-German alignments, together with a fine-grained lexical knowledge base (KB) for archaic expressions. Based on these resources, we propose a hybrid evaluation framework that integrates knowledge-driven, rule-based, and LLM-as-judge metrics. Experimental results show that our framework achieves significantly higher human correlation than traditional metrics and demonstrates high statistical stability. By applying this framework to evaluate representative LLMs, we reveal that while top-tier models like Gemini-2.5-Pro and DeepSeek-3.1 show potential, achieving semantic precision and aesthetic sophistication—particularly in lower-resource directions like German—remains a persistent challenge. Our code, lexical KB, and corpus reconstruction protocols are available at https://github.com/ML-KULeuven/ShijingLLMTrans.
Hierarchical text classification (HTC) assigns documents to multiple levels of a pre-defined taxonomy. Automated patent subject classification represents one of the hardest HTC scenarios because of professional difficulties and extensive labels. Prior approaches only output a flat label set, which offers little insight into the reason behind predictions. Therefore, we propose Reasoning for Hierarchical Classification (RHC), a novel framework that reformulates HTC as a step-by-step reasoning task to sequentially deduce hierarchical labels. RHC trains large language models (LLMs) in two stages: a cold-start stage that aligns outputs with chain-of-thought (CoT) reasoning format and a reinforcement learning (RL) stage to enhance multi-step reasoning ability. RHC demonstrates four advantages in our experiments. (1) Effectiveness: RHC surpasses previous baselines and outperforms the supervised fine-tuning counterparts by approximately 3% in accuracy and macro F1. (2) Explainability: RHC produces natural-language justifications before prediction to facilitate human inspection. (3) Scalability: RHC scales favorably with model size with larger gains compared to standard fine-tuning. (4) Applicability: Beyond patents, we further demonstrate that RHC achieves state-of-the-art performance on other widely used HTC benchmarks, which highlights its broad applicability.
Recent advances in large language models for test case generation have improved branch coverage via prompt-engineered mutations. However, they still lack principled mechanisms for steering models toward specific high-risk execution branches, limiting their effectiveness for discovering subtle bugs and security vulnerabilities. We propose GLMTest, the first program structure-aware LLM framework for targeted test case generation that seamlessly integrates code property graphs and code semantics using a graph neural network and a language model to condition test case generation on execution branches. This structured conditioning enables controllable and branch-targeted test case generation, thereby potentially enhancing bug and security risk discovery. Experiments on real-world projects show that GLMTest built on a Qwen2.5-Coder-7B-Instruct model improves branch accuracy from 27.4% to 50.2% on TestGenEval benchmark compared with state-of-the-art LLMs, i.e., Claude-Sonnet-4.5 and GPT-4o-mini.
Spatial-RAG: Spatial Retrieval Augmented Generation for Real-World Geospatial Reasoning Questions
PDF ↗Answering real-world geospatial questions—such as finding restaurants along a travel route or amenities near a landmark—requires reasoning over both geographic relationships and semantic user intent. However, existing large language models (LLMs) lack spatial computing capabilities and access to up-to-date, ubiquitous real-world geospatial data, while traditional geospatial systems fall short in interpreting natural language. To bridge this gap, we introduce Spatial-RAG, a Retrieval-Augmented Generation (RAG) framework designed for geospatial question answering. Spatial-RAG integrates structured spatial databases with LLMs via a hybrid spatial retriever that combines sparse spatial filtering and dense semantic matching. It formulates the answering process as a multi-objective optimization over spatial and semantic relevance, identifying Pareto-optimal candidates and dynamically selecting the best response based on user intent. Experiments across multiple tourism and map-based QA datasets show that Spatial-RAG significantly improves performance over strong baselines.
Waking Up Blind: Cold-Start Optimization of Supervision-Free Agentic Trajectories for Grounded Visual Perception
PDF ↗Small Vision-Language Models (SVLMs) are efficient task controllers but often suffer from visual brittleness and poor tool orchestration. They typically require expensive supervised trajectory tuning to mitigate these deficits. In this work, we propose Self-supervised Perception Enabled by Cascaded Tool Rollout Alignment (SPECTRA), a supervision-free framework that bootstraps agentic capabilities via Coldstart Reinforcement Learning for SVLMs. SPECTRA enforces Soft Structured Multi-turn Rollouts, a topological constraint that directs agents to explicitly sequence tool derived evidence before synthesis, effectively grounding reasoning in visual observations. We employ a multi-objective reward signal that simultaneously maximizes task correctness, rollout structure, and tool utility, enabling agent to self-discover robust behaviors without human preference labels. We further introduce Tool Instrumental Utility (TIU), a novel metric to quantify tool efficacy in the absence of ground truth. Extensive evaluations across composite and out-of-distribution (MMMU-Pro) benchmarks demonstrate that SPECTRA boosts agentic trajectories, improving task accuracy by up to 5% and tool efficiency by 9%, enabling more efficient multimodal agents that learn effectively from environmental interaction alone.
Self-supervised speech models (S3Ms) are known to encode rich phonetic information, yet how this information is structured remains underexplored. We conduct a comprehensive study across 96 languages to analyze the underlying structure of S3M representations, with particular attention to phonological vectors.We first show that there exist linear directions within the model’s representation space that correspond to phonological features. We further demonstrate that the scale of these phonological vectors correlate to the degree of acoustic realization of their corresponding phonological features in a continuous manner. For example, the difference between [d] and [t] yields a voicing vector: adding this vector to [p] produces [b], while scaling it results in a continuum of voicing. Together, these findings indicate that S3Ms encode speech using phonologically interpretable and compositional vectors, demonstrating phonological vector arithmetic.All code and interactive demos are available at https://github.com/juice500ml/phonetic-arithmetic.
Most existing dialogue systems are user-driven, primarily designed to fulfill user requests. However, in many critical real-world scenarios, a conversational agent must proactively extract information to achieve its own objectives rather than merely respond. To address this gap, we introduce Inquisitive Conversational Agents (ICAs) and develop an ICA specifically tailored to U.S. Supreme Court oral arguments. We propose a Dual Hierarchical Reinforcement Learning framework featuring two cooperating RL agents, each with its own policy, to coordinate strategic dialogue management and fine-grained utterance generation. By learning when and how to ask probing questions, the agent emulates judicial questioning patterns and systematically uncovers crucial information to fulfill its legal objectives. Evaluations on a U.S. Supreme Court dataset show our method outperforms single-agent RL baselines in multiple metrics. Although specialized to a single legal domain, it represents an important first step toward broader high-stakes, domain-specific applications. We attached a part of the code as supplementary material. All code will be released upon publication for reproducibility.
Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse domains. However, the reliability of responses from LLMs remains a question. Uncertainty quantification (UQ) of LLMs is crucial for ensuring their reliability, especially in areas such as healthcare. Existing UQ methods, often designed around a single resource such as Natural Language Inference (NLI) or graph-based metrics, fail to capture the multifaceted nature of uncertainty in natural language generation. In this work, we propose MS-UQ, a novel Multi-Resource Uncertainty Quantification framework that integrates heterogeneous uncertainty signals into a unified measure. Our approach concatenates matrices from diverse resources and employs tensor decomposition to orthogonally disentangle unique and shared information. To ensure scalability, we construct an adaptive ensemble of outputs from different decomposition methods, enabling the incorporation of new uncertainty sources. Experiments on CoQA, NQ_Open, and HotpotQA demonstrate that MS-UQ consistently outperforms existing methods, offering a comprehensive and scalable solution for uncertainty estimation in black-box LLMs and a more robust framework for enhancing LLM reliability in high-stakes applications. Our code can be accessed at https://anonymous.4open.science/r/MDUQ-First-202E/README.md.
When a student fails an exam, do we tend to blame their effort or the test’s difficulty? Attribution, defined as how reasons are assigned to event outcomes, shapes perceptions, reinforces stereotypes, and influences decisions. Attribution Theory explains how people attribute causes to internal factors (effort, ability) or external ones (task difficulty, luck). LLMs’ attribution of event outcomes based on demographics carries important fairness implications. Most works exploring social biases in LLMs focus on surface-level associations or isolated stereotypes. This work proposes a cognitively grounded bias evaluation framework to identify how models’ reasoning disparities shape demographic bias across three contexts: single-actor, actor–actor, and actor–observer, capturing comparative and perspective-driven biases overlooked in prior work. Introducing a 140k-prompt benchmark covering ten scenarios and four social dimensions, our analyses reveal attribution asymmetries across identities that vary in multi-actor and observer settings, suggesting that other identities influence bias. This work underscores the need for cognitively grounded bias evaluation and informs future debiasing efforts through the proposed framework.
In-context learning (ICL) performance depends critically on which demonstrations are placed in the prompt, yet most existing selectors prioritize heuristic notions of relevance or diversity and provide limited insight into the \textbf{coverage} of a demonstration set. We propose \textbf{Unseen Coverage Selection (UCS)}, a training-free, subset-level coverage prior motivated by the principle that a good demonstration set should \textbf{expose the model to latent cluster unrevealed by the currently selected subset}. UCS operationalizes this idea by \textbf{(1)} inducing discrete latent \textbf{clusters} from model-consistent embeddings and \textbf{(2)} estimating the number of unrevealed clusters within a candidate subset via a Smoothed Good-Turing estimator from its empirical frequency spectrum. Unlike previous selection methods, UCS is coverage-based and training-free, and can be seamlessly combined with both query-dependent and query-independent selection baselines via a simple regularized objective. Experiments on multiple intent-classification and reasoning benchmarks with frontier Large Language Models show that augmenting strong baselines with UCS consistently improves ICL accuracy by up to \textbf{2-6}% under the same selection budget, while also yielding insights into task- and model-level latent cluster distributions. Code is available at https://github.com/Raina-Xin/UCS.
The systematicity of natural language interpretation—our ability to understand novel expressions by compositionally combining familiar elements—has been central to debates about symbolic versus neural approaches to cognition since Fodor and Pylyshyn (1988). We investigate whether artificial neural networks can learn model-theoretic interpretation functions that generalize systematically to out-of-training-sample sentences, framing interpretation as an encoding task from discrete linguistic input to continuous truth-conditional representations. We extend Frank et al. (2009) with entity-level semantic representations, modern architectures (GRU, LSTM, Attention with AbsPE/RoPE), principled competing event generation, extended systematicity tests (∼350 vs. ∼80 sentences), and a two- dimensional difficulty analysis disaggregating results by modifier complexity. Across 140 trained models (7 architectures), we find that capacity-matched architectures perform comparably on easy tests, but gated recurrent networks (GRU and LSTM) significantly outperform transformer architectures on the hardest compositional generalization test (Basic Event), while ungated SRN does not—indicating that the gating mechanism is a critical factor. Entity vectors significantly improve scores on Basic Event across most architectures, with gated architectures benefiting most, validating formal semantics’ treatment of entities as important theoretical primitives. The extended test set reveals that systematicity difficulty has two dimensions: the type of systematicity test (as in Frank et al. 2009), and the number of modifiers being composed.
A first decision for any automated natural language processing system is the granularity of the input units. Traditionally, characters or words have been used, but recently, subwords have become the standard. In this paper, we investigate trends in input processing steps and discuss common shortcomings in this foundational first step of model design. We start by providing an overview of currently used tokenizers, showing that there is only minimal variety, with three highly similar designs dominating current models, and many of the tokenizers being exact duplicates. Next, we reconsider Unicode normalization strategies. Previous work has recommended applying consistent normalization; however, we argue that this removes signal and we show how this can harm performance for language classification. Finally, we take a closer look at UTF-8 character encoding, the very first layer of representation used in many language models. We argue that UTF-8 is not optimized for efficiency, nor for fairness across languages, and propose proof of concept alternatives focused on fairness and efficiency. Based on our findings, we recommend future work to 1) put more thought into subword segmentation and explore more diversity, 2) apply normalization only when beneficial 3) consider alternative character encodings for models operating on the byte-level.
Despite significant progress in alignment, large language models (LLMs) remain vulnerable to adversarial attacks that elicit harmful behaviors. Activation steering techniques offer a promising inference-time intervention approach, but existing methods suffer from critical limitations: activation addition requires careful coefficient tuning and is sensitive to layer-specific norm variations, while directional ablation provides only binary control. Recent work on Angular Steering introduces continuous control via rotation in a 2D subspace, but its practical implementation violates norm preservation, causing distribution shift and generation collapse, particularly in models below 7B parameters. We propose Selective Steering, which addresses these limitations through two key innovations: (1) a mathematically rigorous norm-preserving rotation formulation that maintains activation distribution integrity, and (2) discriminative layer selection that applies steering only where feature representations exhibit opposite-signed class alignment. Experiments across nine models demonstrate that Selective Steering achieves 5.5 higher attack success rates than prior methods while maintaining zero perplexity violations and approximately 100% capability retention on standard benchmarks. Our approach provides a principled, efficient framework for controllable and stable LLM behavior modification.
Current vision-language benchmarks predominantly feature well-structured questions with clear, explicit prompts. However, real user queries are often informal and underspecified. Users naturally leave much unsaid, relying on images to convey context. We introduce HAERAE-Vision, a benchmark of 653 real-world visual questions from Korean online communities (0.76% survival from 86K candidates), each paired with an explicit rewrite, yielding 1,306 query variants in total. Evaluating 39 VLMs, we find that even state-of-the-art models (GPT-5, Gemini 2.5 Pro) achieve under 50% on the original queries. Crucially, query explicitation alone yields 8 to 22 point improvements, with smaller models benefiting most. We further show that even with web search, under-specified queries underperform explicit queries without search, revealing that current retrieval cannot compensate for what users leave unsaid. Our findings demonstrate that a substantial portion of VLM difficulty stem from natural query under-specification instead of model capability, highlighting a critical gap between benchmark evaluation and real-world deployment.
Electronic health record (EHR) question answering is often handled by LLM-based pipelines that are costly to deploy and do not explicitly leverage the hierarchical structure of clinical data. Motivated by evidence that medical ontologies and patient trajectories exhibit hyperbolic geometry, we propose HypEHR, a compact Lorentzian model that embeds codes, visits, and questions in hyperbolic space and answers queries via geometry-consistent cross-attention with type-specific pointer heads. HypEHR is pretrained with next-visit diagnosis prediction and hierarchy-aware regularization to align representations with the ICD ontology. On two MIMIC-IV-based EHR-QA benchmarks, HypEHR approaches LLM-based methods while using far fewer parameters. Our code is publicly available at https://github.com/yuyuliu11037/HypEHR.
The advent of LLMs has given rise to generative search, a new search paradigm in which LLMs retrieve information from the web related to a query and synthesize it into a single, coherent response. This paradigm differs fundamentally from traditional web search, where results are returned as a ranked list of independent web pages. In this paper, we ask: Along what dimensions does generative search differ from traditional search?We conduct a systematic comparison between Google organic search and five generative search systems from three providers: Google, OpenAI, and Perplexity. Our analysis reveals substantial variation among engines in their reliance on internal v.s. external knowledge, source diversity, and stability. While generative systems often achieve topical coverage comparable to traditional search, they do so using markedly different retrieval footprints and synthesis strategies. We further show that the outputs of generative search can vary across time and executions, raising new challenges for robustness. Our findings demonstrate that generative search introduces new dimensions that are not captured by existing evaluation paradigms, motivating the development of evaluations that explicitly account for retrieval behavior, synthesis, and stability in generative search systems.
While memory is a core component in agent systems, its behavioral impact in complex, long-horizon domains like machine learning engineering (MLE) remains poorly understood. Unlike short, reactive exchanges, MLE agents solve tasks through cycles of experimentation and improvement where past errors can inform future success. This paper presents a systematic study dissecting how memory influences agent behavior and performance across diverse MLE challenges. We first introduce a dynamic coding memory designed to capture and reuse debugging experiences, and integrate it into two representative agent paradigms: a sequential, chain-based agent that mirrors human-like iterative refinement, and a parallel, tree-based agent that performs broad, self-exploratory search in the code space. Our central finding is that the role of memory is contingent on the agent’s underlying architecture. For chain-based agents, memory proves highly beneficial, enabling them to avoid recurring mistakes and engage in more coherent, iterative refinement, which significantly improves reliability and task success. In contrast, for tree-based search agents, memory introduces a critical trade-off: it enhances procedural stability at the cost of constraining search diversity, which can prematurely narrow exploration and lead to suboptimal final solutions. These findings reveal a fundamental trade-off between procedural reliability and solution innovation modulated by memory, offering insights for designing more effective and robust MLE agents.
Exploiting Tree Structure for Credit Assignment in Reinforcement Learning with Large Language Models
PDF ↗Reinforcement learning has shown strong promise for strengthening the reasoning ability of large language models (LLMs), but sparse, delayed rewards over long chains make token-level credit assignment a central challenge. Actor–critic methods like PPO provide token-level credit but require training a value network alongside the policy, which introduces complexity and can encourage overfitting. Critic-free alternatives such as GRPO avoid this burden but rely on sequence-level outcomes, distributing a single reward uniformly across tokens and ignoring structural differences between responses. We propose Prefix-to-Tree (P2T), which organizes the sampled responses of a prompt into a prefix tree and computes nonparametric prefix values by aggregating descendant outcomes. Building on this idea, we develop TEMPO (Tree-Estimated Mean Prefix Value for Policy Optimization), a critic-free algorithm that enriches GRPO with branch-aware temporal-difference (TD) corrections. Across Qwen3-1.7B and Qwen3-4B, TEMPO consistently improves both convergence and final performance over PPO and GRPO on in-distribution benchmarks (MATH, MedQA) and out-of-distribution settings (GSM-HARD, AMC23, MedMCQA, MMLU-Medical), achieving higher validation accuracy within comparable wall-clock time.