论文检索

Theory · Learning Theory

Jonas von Berg, Adalbert Fono, Massimiliano Datres, Sohir Maskey, Gitta Kutyniok

The relationship between overparameterization, stability, and generalization remains incompletely understood in the setting of discontinuous classifiers. We address this gap by establishing a generalization bound for finite function classes that improves inversely with _class stability_, defined as the expected distance to the decision boundary in the input domain (margin). Interpreting class stability as a quantifiable notion of robustness, we derive as a corollary a _law of robustness_ for classification that extends the results of Bubeck and Selke beyond smoothness assumptions to discontinuous functions. In particular, any interpolating model with $p \approx n$ parameters on $n$ data points must be _unstable_, implying that substantial overparameterization is necessary to achieve high stability. We obtain analogous results for (parameterized) infinite function classes by analyzing a stronger robustness measure derived from the margin in the co-domain, which we refer to as the _normalized co-stability_. Experiments support our theory: stability increases with model size and correlates with test performance, while traditional norm-based measures remain largely uninformative.

Social Aspects · Fairness, Equity, Justice and Safety

Rui Guo, Weibin Zeng, Fuzhang Wu, Yan Kong, sicheng shen, Yanjun WU, Weiming Dong

The widespread deployment of large language models (LLMs) has made the reliable detection of AI-generated text a crucial task. However, existing zero-shot detectors typically rely on proxy models to approximate probability distributions of unknown source models at a single token level. Such approaches limit detection effectiveness and make the results highly sensitive to the choice of proxy models. In contrast, supervised classifiers are often detected as black boxes, sacrificing interpretability in the detection process. To address these limitations, we propose a novel detection framework that identifies LLM-generated text by approximating **H**ierarchical **L**inguistic **D**istributions--**HLD-Detector**. Specifically, we leverage n-grams to capture the feature distribution of human-written and machine-generated text across the word, syntactic, and semantic levels, and perform LLM-generated text detection by comparing these distributions under the Bayesian theory. By progressively modeling the linguistic distribution from shallow-level (token/word), then medium-level (syntactic), and ultimately high-level (semantic representations), our method mitigates the shortcomings of previous single feature level detectors, improving both robustness and overall performance. Additionally, HLD-Detector requires only a small amount of offline corpus for distribution estimation, instead of relying on online approximation with large proxy models, resulting in significantly lower computational overhead. Extensive experiments have verified the superiority of our method in detection tasks such as multi-llm and multi-domain scenarios, achieving the current SOTA performance.

Applications · Language, Speech and Dialog

Parthiv Chatterjee, Asish Batha, Tashvi patel, Sourish Dasgupta, Tanmoy Chakraborty

Document summarization helps readers focus on the "content-of-interest", a *subjective* and *time-variant* quantity. Capturing this *dynamic subjectivity* requires modeling how user preferences evolve over time, thereby demanding *personalized summarization*. Recent news recommendation and summarization models often assume that preferences follow a *memoryless or short-memory random walk* on interaction graphs, i.e., a Markovian diffusion seeded at the latest interaction or compressed into a short hidden state or prompt. We ask whether such a hypothesis also holds for personalized summarization. To test this, we propose **Walk2Pers**, a lightweight encoder–decoder framework that extends the walk view with *action-conditioned geometric steps*, decomposed into (i) a *magnitude* controlling shift strength and (ii) an *orientation* capturing continuity vs. novelty. The process is mediated by dual memory lanes that reinforce consistent interests while suppressing disinterest, and is augmented with a drift term for summary requests. We show theoretically that such structured walks approximate first-order action-conditioned kernels, and empirically validate the hypothesis on PENS, OpenAI-Reddit, and PersonalSum. Using PerSEval, a personalization metric with strong human correlation, Walk2Pers outperforms specialized personalized summarizers by an average of $0.41 \uparrow$, and strong LLM baselines (DeepSeek-R1-14B, LLaMA-2-13B, Mistral-7B, Zephyr-7B) by $0.22 \uparrow$. Analyses further confirm cross-domain robustness ($0.19 \uparrow$ over the best LLM) and stability on long histories. Together, these results support viewing personalized summarization as an *action-biased geometric walk with memory*, offering both interpretability and efficiency.

Gabriela Moisescu-Pareja, Gavin McCracken, Harley Wiltzer, Colin Daniels, Vincent Létourneau, Doina Precup, Jonathan Love

The Clock and Pizza interpretations, associated with architectures differing in either uniform or learnable attention, were introduced to argue that different architectural designs can yield distinct circuits for modular addition. In this work, we show that this is not the case, and that both the uniform and trainable attention architectures implement the same algorithm via topologically and geometrically equivalent representations. Our methodology goes beyond the interpretation of individual neurons and weights. Instead, we identify all of the neurons corresponding to each learned representation and then study the collective group of neurons as one entity. This method reveals that each learned representation is a manifold that we can study utilizing tools from topology. Based on this insight, we can statistically analyze the learned representations across hundreds of circuits to demonstrate the similarity between learned modular addition circuits that arise naturally from common deep learning paradigms.

Theory · Interpretability and Visualization

Dong-Kyum Kim, Minsung Kim, Jea Kwon, Nakyeong Yang, Meeyoung Cha

The reversal curse—a language model's inability to infer an unseen fact "B is A" from a learned fact "A is B"—is widely considered a fundamental limitation. We show that this is not an inherent failure but an artifact of how models encode knowledge. Our results demonstrate that training from scratch on synthetic relational knowledge graphs leads to the emergence of a bilinear relational structure within the models' hidden representations. This structure alleviates the reversal curse and facilitates inference of unseen reverse facts. Crucially, this bilinear geometry is foundational for consistent model editing: updates to a single fact propagate correctly to its reverse and logically dependent relations. In contrast, models lacking this representation suffer from the reversal curse and fail to generalize model edits, leading to logical inconsistencies. Our results establish that training on a relational knowledge dataset induces the emergence of bilinear internal representations, which in turn support language models in behaving in a logically consistent manner after editing. This suggests that the efficacy of language model editing depends not only on the choice of algorithm but on the underlying representational geometry of the knowledge itself.

Social Aspects · Accountability, Transparency and Interpretability

Nakyeong Yang, Dong-Kyum Kim, Jea Kwon, Minsung Kim, Kyomin Jung, Meeyoung Cha

Large language models trained on web-scale data can memorize private or sensitive knowledge, raising significant privacy risks. Although some unlearning methods mitigate these risks, they remain vulnerable to "relearning" during subsequent training, allowing a substantial portion of forgotten knowledge to resurface. In this paper, we show that widely used unlearning methods cause shallow alignment: instead of faithfully erasing target knowledge, they generate spurious unlearning neurons that amplify negative influence to hide it. To overcome this limitation, we introduce Ssiuu, a new class of unlearning methods that employs attribution-guided regularization to prevent spurious negative influence and faithfully remove target knowledge. Experimental results confirm that our method reliably erases target knowledge and outperforms strong baselines across two practical retraining scenarios: (1) adversarial injection of private data, and (2) benign attack using an instruction-following benchmark. Our findings highlight the necessity of robust and faithful unlearning methods for safe deployment of language models.

Computer Vision · Vision Models & Multimodal

Yingzhao Jian, Zhongan Wang, Yi Yang, Hehe Fan

In this paper, we explore how to empower general-purpose Vision-Language Models (VLMs) to control humanoid agents. General-purpose VLMs (e.g., GPT-4) exhibit strong open-world generalization, and remove the need for additional fine-tuning data. To build such an agent, two key components are required: (1) an embodied instruction compiler, which enables the VLM to observe the scene and translate high-level user instructions into low-level control parameters; and (2) a motion executor, which generates human motions from these parameters while adapting to real-time physical feedback. We present BiBo, a VLM-driven humanoid agent composed of an embodied instruction compiler and a diffusion-based motion executor. The compiler interprets user instructions in context with the environment, and leverages a chain of visual question answering (VQA) to guide the VLM in specifying control parameters (e.g., motion captions, locations). The diffusion executor extends future joint trajectories from prior motion, conditioned on both control parameters and environmental feedback. Experiments demonstrate that BiBo achieves an interaction task success rate of 90.2\% in open environments, and improves the precision of text-guided motion execution by 16.3\% over prior methods. BiBo handles not only basic interaction but also diverse motions, and even dancing while striking at a sandbag. The code will be released upon publication.

Reinforcement Learning · Multi-agent

Zhixi Cai, Fucai Ke, Kevin Leo, Sukai Huang, Maria de la Banda, Peter Stuckey, Hamid Rezatofighi

Recent vision-language models have strong perceptual ability but their implicit reasoning is hard to explain and easily generates hallucinations on complex queries. Compositional methods improve interpretability, but most rely on a single agent or hand-crafted pipeline and cannot decide when to collaborate across complementary agents or compete among overlapping ones. We introduce MATA (Multi-Agent hierarchical Trainable Automaton), a multi-agent system presented as hierarchical finite-state automaton for visual reasoning whose top-level transitions are chosen by a trainable hyper agent. Each agent corresponds to a state in the hyper automaton, and runs a small rule-based sub-automaton for reliable micro-control. All agents read and write a shared memory, yielding transparent execution history. To supervise the hyper agent’s transition policy, we build transition-trajectory trees and transform to memory-to-next-state pairs, forming the MATA-SFT-90K dataset for supervised finetuning (SFT). The finetuned LLM as the transition policy understands the query and the capacity of agents, and it can efficiently choose the optimal agent to solve the task. Across multiple visual reasoning benchmarks, MATA achieves the state-of-the-art results compared with monolithic and compositional baselines. The code and dataset are available at https://github.com/ControlNet/MATA.

Deep Learning · Generative Models and Autoencoders

Todd Nief, David Reber, Sean Richardson, Ari Holtzman

When an LLM learns a new fact during finetuning (e.g., new movie releases, newly elected pope, etc.), where does this information go? Are entities enriched with relation information immediately, or do models recall information just-in-time before a prediction? Or, are "all of the above" true, with LLMs implementing multiple redundant heuristics? Existing localization approaches (e.g., activation patching) are ill-suited for this analysis because they usually replace parts of the residual stream, thus overriding previous information. To fill this interpretability gap, we propose dynamic weight grafting, an analysis technique that selectively grafts subsets of weights from a finetuned model onto a pretrained model. Using this technique, we show two separate pathways for retrieving finetuned relation information: 1) "enriching" the residual stream with relation information while processing the tokens that correspond to an entity (e.g., "Zendaya" in "Zendaya co-starred with Timothée Chalamet" and 2) "recalling" this information at the final token position before generating a target fact. In some cases, models need information from both of these pathways to correctly generate finetuned facts while, in other cases, either the "enrichment" or "recall" pathway alone is sufficient. We localize the "recall" pathway to model components---finding that "recall" occurs via both task-specific attention mechanisms and an entity-specific extraction step in the feedforward networks of the final layers before prediction. By targeting model components and parameters, as opposed to just activations, we are able to understand the mechanisms by which finetuned knowledge is retrieved during generation.

Social Aspects · Fairness, Equity, Justice and Safety

James Oldfield, Philip Torr, Ioannis Patras, Adel Bibi, Fazl Barez

Monitoring large language models' (LLMs) activations is an effective way to detect harmful requests before they lead to unsafe outputs. However, traditional safety monitors often require the same amount of compute for every query. This creates a trade-off: expensive monitors waste resources on easy inputs, while cheap ones risk missing subtle cases. We argue that safety monitors should be flexible--costs should rise only when inputs are difficult to assess, or when more compute is available. To achieve this, we introduce Truncated Polynomial Classifiers (TPCs), a natural extension of linear probes for dynamic activation monitoring. Our key insight is that polynomials can be trained and evaluated progressively, term-by-term. At test-time, one can early-stop for lightweight monitoring, or use more terms for stronger guardrails when needed. TPCs provide two modes of use. First, as a safety dial: by evaluating more terms, developers and regulators can "buy" stronger guardrails from the same model. Second, as an adaptive cascade: clear cases exit early after low-order checks, and higher-order guardrails are evaluated only for ambiguous inputs, reducing overall monitoring costs. On two large-scale safety datasets (WildGuardMix and BeaverTails), for 4 models with up to 30B parameters, we show that TPCs compete with or outperform MLP-based probe baselines of the same size, all the while being more interpretable than their black-box counterparts. Our anonymous code is available at https://github.com/james-oldfield/tpc/.

General Machine Learning · Representation Learning

Kosio Beshkov, Anders Malthe-Sorenssen

While protein language models (PLMs) are one of the most promising avenues of research for future de novo protein design, the way in which they transform sequences to hidden representations, as well as the information encoded in such representations is yet to be fully understood. Several works have attempted to propose interpretability tools for PLMs, but they have focused on understanding how individual sequences are transformed by such models. Therefore, the way in which PLMs transform the whole space of sequences along with their relations is still unknown. In this work we attempt to understand this transformed space of sequences by identifying protein structure and representation with square-root velocity (SRV) representations and graph filtrations. Both approaches naturally lead to a metric space in which pairs of proteins or protein representations can be compared with each other. We analyze different types of proteins from the SCOP dataset and show that the Fréchet radius and effective dimension of the SRV shape space follows a non-linear pattern as a function of the layers in ESM2 models of different sizes. Furthermore, we use graph filtrations as a tool to study the context lengths at which models encode the structural features of proteins. We find that PLMs preferentially encode immediate as well as local relations between amino acids, but start to degrade for larger context lengths. The most structurally faithful encoding tends to occur close to, but before the last layer of the models, indicating that training a folding model ontop of these layers might lead to improved folding performance.

Computer Vision · Image and Video Generation

Wootaek Jeong, Junghyo Sohn, Jee Seok Yoon, Heung-Il Suk

Text-guided image editing requires more than prompt following—it demands a principled understanding of what to modify versus what to preserve. We investigate the internal guidance mechanism of diffusion models and reveal that the guidance signal follows a structured semantic hierarchy. We formalize this insight as the Semantic Scale Hypothesis: the magnitude of the guidance difference vector ($\Delta\boldsymbol{\epsilon}$) directly encodes the semantic scale of edits. Crucially, this phenomenon is theoretically grounded in Tweedie’s formula, which links score prediction to the variance of the underlying data distribution. Low-variance regions, such as objects, yield large-magnitude differences corresponding to structural edits, whereas high-variance regions, such as backgrounds, yield small-magnitude differences corresponding to stylistic adjustments. Building on this principle, we introduce Prism-Edit, a training-free, plug-and-play module that decomposes the guidance signal into semantic layers, enabling selective and interpretable control. Extensive experiments—spanning direct visualization of the semantic hierarchy, generalization across foundation models, and integration with state-of-the-art editors—demonstrate that Prism-Edit achieves precise, robust, and controllable editing. Our findings establish semantic scale as a foundational axis for understanding and advancing diffusion-based image editing.

Reinforcement Learning · Multi-agent

Iris Xu, Guangtao Zeng, Zexue He, Charles Jin, Aldo Pareja, Dan Gutfreund, Chuang Gan, Zhang-Wei Hong

Large language models (LLMs) have shown strong reasoning and coding capabilities, yet they struggle to generalize to real-world software engineering (SWE) problems that are long-horizon and out-of-distribution. Existing systems often rely on a single agent to handle the entire workflow—interpreting issues, navigating large codebases, and implementing fixes—within one reasoning chain. Such monolithic designs force the model to retain irrelevant context, leading to spurious correlations and poor generalization. Motivated by how human engineers decompose complex problems, we propose structuring SWE agents as orchestrators coordinating specialized sub-agents for sub-tasks such as localization, editing, and validation. The challenge lies in discovering effective hierarchies automatically: as the number of sub-agents grows, the search space becomes combinatorial, and it is difficult to attribute credit to individual sub-agents within a team. We address these challenges by formulating hierarchy discovery as a multi-armed bandit (MAB) problem, where each arm represents a candidate sub-agent and the reward measures its helpfulness when collaborating with others. This framework, termed Bandit Optimization for Agent Design (BOAD), enables efficient exploration of sub-agent designs under limited evaluation budgets. On SWE-bench-Verified, BOAD outperforms single-agent and manually designed multi-agent systems. On SWE-bench-Live, featuring more recent and out-of-distribution issues, our 36B system ranks second on the leaderboard at the time of evaluation, surpassing larger models such as GPT-4 and Claude. These results demonstrate that automatically discovered hierarchical multi-agent systems significantly improve generalization on challenging long-horizon SWE tasks. Code is available at https://github.com/iamxjy/BOAD-SWE-Agent.

Applications · Everything Else

Haochen Luo, HoTin Ko, Jiandong Chen, David Sun, Yuan Zhang, Chen Liu

Formulaic alpha factor mining (FAFM) is a central problem in quantitative investment, where interpretable formulas are designed to extract predictive signals from historical financial series. With the emergence of large language models (LLMs), recent studies have begun to explore their roles in FAFM, yet their capabilities across different tasks and configurations remain unclear. In this work, we introduce AlphaBench, the first systematic benchmark for evaluating LLMs in FAFM. AlphaBench covers three core tasks, including factor generation, factor evaluation, and factor searching, which are all popular tasks integrated in the workflow of quantitative researchers. Beyond task-level evaluation, we further analyze how different LLM settings, including model type, prompting paradigm, and reasoning strategy, influence performance. Our experiments on a range of open-source and closed-source models reveal that LLMs hold strong potential in automating factor mining, while also facing persistent challenges in robustness, search efficiency, and practical usability. The project is available at: https://alphabench.cc/

Social Aspects · Accountability, Transparency and Interpretability

Eric Enouen, sainyam galhotra

Concept Bottleneck Models (CBMs) use a set of human-interpretable concepts to predict the final task label, enabling domain experts to not only validate the CBM's predictions, but also intervene on incorrect concepts at test time. However, these interventions fail to address systemic misalignment between the CBM and the expert's reasoning, such as when the model learns shortcuts from biased data. To address this, we present a general interpretable debugging framework for CBMs that follows a two-step process of *Removal* and *Retraining*. In the *Removal* step, experts use concept explanations to identify and remove any undesired concepts. In the *Retraining* step, we introduce **CBDebug**, a novel method that leverages the interpretability of CBMs as a bridge for converting concept-level user feedback into sample-level auxiliary labels. These labels are then used to apply supervised bias mitigation and targeted augmentation, reducing the model’s reliance on undesired concepts. We evaluate our framework with both real and automated expert feedback, and find that **CBDebug** significantly outperforms prior retraining methods across multiple CBM architectures (PIP-Net, Post-hoc CBM) and benchmarks with known spurious correlations.

Social Aspects · Accountability, Transparency and Interpretability

Mélissa Tamine, Benjamin Heymann, Maxime Vono, Patrick Loiseau

Semivalue–based data valuation uses cooperative‐game theory intuitions to assign each data point a value reflecting its contribution to a downstream task. Still, those values depend on the practitioner’s choice of utility, raising the question: *How robust is semivalue-based data valuation to changes in the utility?* This issue is critical when the utility is set as a trade‐off between several criteria and when practitioners must select among multiple equally valid utilities. We address this by introducing the notion of a dataset’s *spatial signature*: given a semivalue, we embed each data point into a lower-dimensional space in which any utility becomes a linear functional, making the data valuation framework amenable to a simpler geometric picture. Building on this, we propose a practical methodology centered on an explicit robustness metric that informs practitioners whether and by how much their data valuation results will shift as the utility changes. We validate this approach across diverse datasets and semivalues, demonstrating strong agreement with rank‐correlation analyses and offering analytical insight into how choosing a semivalue can amplify or diminish robustness.

Deep Learning · Everything Else

Michał Zawalski, Meriem Boubdir, Klaudia Bałazy, Besmira Nushi, Pablo Ribalta

We present Contamination Detection via Context (CoDeC), a practical and accurate method to detect and quantify training data contamination in large language models. CoDeC distinguishes between data memorized during training and data outside the training distribution by measuring how in-context learning affects model performance. We find that in‑context examples typically boost confidence for unseen datasets but may reduce it when the dataset was part of training, due to disrupted memorization patterns. Experiments show that CoDeC produces interpretable contamination scores that clearly separate seen and unseen datasets, and reveals strong evidence of memorization in open-weight models with undisclosed training corpora. The method is simple, automated, and both model- and dataset-agnostic, making it easy to integrate with benchmark evaluations.

Theory · Interpretability and Visualization

Zhuowen Liang, Xiaotian LIN, Zhengxuan Zhang, Yuyu Luo, Haixun Wang, Nan Tang

Large language models (LLMs) are widely applied to data analytics over documents, yet direct reasoning over long, noisy documents remains brittle and error-prone. Hence, we study document question answering (QA) that consolidates dispersed evidence into a structured output (e.g., a table, graph, or chunks) to support reliable, verifiable QA. We propose a two-pillar framework, LiteCoST, to achieve both high accuracy and low latency with small language models (SLMs). Pillar 1: Chain-of-Structured-Thought (CoST). We introduce a CoST template, a schema-aware instruction that guides a strong LLM to produce both a step-wise CoST trace and the corresponding structured output. The process induces a minimal structure, normalizes entities/units, aligns records, serializes the output, and verifies/refines it, yielding auditable supervision. Pillar 2: SLM fine-tuning. The compact models are trained on LLM-generated CoST data in two stages: Supervised Fine-Tuning for structural alignment, followed by Group Relative Policy Optimization (GRPO) incorporating triple rewards for answer/format quality and process consistency. By distilling structure-first behavior into SLMs, this approach achieves LLM-comparable quality on multi-domain long-document QA using 3B/7B SLMs, while delivering 2–4× lower latency than GPT-4o and DeepSeek-R1 (671B). The code is available at https://github.com/HKUSTDial/LiteCoST.

Applications · Neuroscience, Cognitive Science

Jathurshan Pradeepkumar, Xihao Piao, Zheng Chen, Jimeng Sun

Foundation models are reshaping EEG analysis, yet an important problem of EEG tokenization remains a challenge. This paper presents TFM-Tokenizer, a novel tokenization framework that learns a vocabulary of time-frequency motifs from *single-channel* EEG signals and encodes them into discrete tokens. We propose a dual-path architecture with time–frequency masking to capture robust motif representations, and it is model-agnostic, supporting both lightweight transformers and existing foundation models for downstream tasks. Our study demonstrates three key benefits: *Accuracy:* Experiments on four diverse EEG benchmarks demonstrate consistent performance gains across both single- and multi-dataset pretraining settings, achieving up to $11\%$ improvement in Cohen’s Kappa over strong baselines. *Generalization:* Moreover, as a plug-and-play component, it consistently boosts the performance of diverse foundation models, including BIOT and LaBraM. *Scalability:* By operating at the single-channel level rather than relying on the strict 10–20 EEG system, our method has the potential to be device-agnostic. Experiments on ear-EEG sleep staging, which differs from the pretraining data in signal format, channel configuration, recording device, and task, show that our tokenizer outperforms baselines by $14\%$. A comprehensive token analysis reveals strong class-discriminative, frequency-aware, and consistent structure, enabling improved representation quality and interpretability. Code is available at https://github.com/Jathurshan0330/TFM-Tokenizer.

Shuhang Chen, Yunqiu Xu, Junjie Xie, Aojun Lu, Tao Feng, ZEYING HUANG, ZHANG NING, Yi Sun, Yi Yang, Hangjie Yuan

Despite significant progress, multimodal large language models continue to struggle with visual mathematical problem solving. Some recent works recognize that visual perception is a bottleneck in visual mathematical reasoning, but their solutions are limited to improving the extraction and interpretation of visual inputs. Notably, they all ignore the key issue of whether the extracted visual cues are faithfully integrated and properly utilized in subsequent reasoning. Motivated by this, we present CogFlow, a novel cognitive-inspired three-stage framework that incorporates a knowledge internalization stage, explicitly simulating the hierarchical flow of human reasoning: perception$\Rightarrow$internalization$\Rightarrow$reasoning. In line with this hierarchical flow, we holistically enhance all its stages. We devise Synergistic Visual Rewards to boost perception capabilities in parametric and semantic spaces, jointly improving visual information extraction from symbols and diagrams. To guarantee faithful integration of extracted visual cues into subsequent reasoning, we introduce a Knowledge Internalization Reward model in the internalization stage, bridging perception and reasoning. Moreover, we design a Visual-Gated Policy Optimization algorithm to further enforce the reasoning is grounded with the visual knowledge, preventing models seeking shortcuts that appear coherent but are visually ungrounded reasoning chains. Moreover, we contribute a new dataset MathCog for model training, which contains samples with over 120K high-quality perception-reasoning aligned annotations. Comprehensive experiments and analysis on commonly used visual mathematical reasoning benchmarks validate the superiority of the proposed CogFlow. Project page: https://shchen233.github.io/cogflow.