Unveiling Language Skills via Path-Level Circuit Discovery

• URL: http://arxiv.org/abs/2410.01334v2
• Date: Mon, 16 Dec 2024 03:33:36 GMT
• Title: Unveiling Language Skills via Path-Level Circuit Discovery
• Abstract summary: We propose a novel path-level circuit discovery framework capturing how behaviors emerge through interconnected linear chain.<n>Our framework is constructed upon a fully-disentangled linear combinations of memory circuits'' decomposed from the original model.<n>In contrast to circuit graph from existing works, we focus on the complete paths of a generic skill rather than on the fine-grained responses to individual components of the input.
• Score: 31.608080868988825
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Circuit discovery with edge-level ablation has become a foundational framework for mechanism interpretability of language models. However, its focus on individual edges often overlooks the sequential, path-level causal relationships that underpin complex behaviors, thus potentially leading to misleading or incomplete circuit discoveries. To address this issue, we propose a novel path-level circuit discovery framework capturing how behaviors emerge through interconnected linear chain and build towards complex behaviors. Our framework is constructed upon a fully-disentangled linear combinations of ``memory circuits'' decomposed from the original model. To discover functional circuit paths, we leverage a 2-step pruning strategy by first reducing the computational graph to a faithful and minimal subgraph and then applying causal mediation to identify common paths of a specific skill, termed as skill paths. In contrast to circuit graph from existing works, we focus on the complete paths of a generic skill rather than on the fine-grained responses to individual components of the input. To demonstrate this, we explore three generic language skills, namely Previous Token Skill, Induction Skill and In-Context Learning Skill using our framework and provide more compelling evidence to substantiate stratification and inclusiveness of these skills.

Related papers

• SkillTrace: Traversing a Query-Skill Graph for Composable LLM Agents [22.759579282367614]
  We introduce SkillTrace, which organizes user queries into a semantic hierarchy, matches skill queries and candidates, and propagates over the skill dependencies.<n> Experiments on SkillsBench and ALFWorld demonstrate that SkillTrace achieves state-of-the-art performance, reaching a success rate of 53.17%.
  arXiv  Detail & Related papers  (2026-08-03T15:07:11Z)
• SkillGraph: Skill-Augmented Reinforcement Learning for Agents via Evolving Skill Graphs [70.1970574147839]
  We propose SKILLGRAPH, a framework that represents reusable skills as nodes in a directed graph.<n>SKILLGRAPH retrieves not just individual skills, but an ordered skill subgraph that can guide multi-step decision making.<n> Experiments show that SKILLGRAPH achieves state-of-the-art performance against memory-augmented RL methods.
  arXiv  Detail & Related papers  (2026-05-12T12:21:49Z)
• ADAG: Automatically Describing Attribution Graphs [50.460651620833055]
  We introduce textbfADAG, an end-to-end pipeline for describing these attribution graphs which is fully automated.<n>We then introduce a novel clustering algorithm for grouping features, and an LLM explainer--simulator setup which generates and scores natural-language explanations of the functional role of these feature groups.<n>We run our system on known human-analysed circuit-tracing tasks and recover interpretable circuits, and further show ADAG can find steerable clusters which are responsible for a harmful advice jailbreak in Llama 3.1 8B Instruct.
  arXiv  Detail & Related papers  (2026-04-08T21:34:37Z)
• STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization [87.77475595961154]
  We present textbfSkill textbfTraining with textbfAugmented textbfRotation (textbfSTAR), a framework that advances both skill learning and composition to complete complex behaviors.
  arXiv  Detail & Related papers  (2025-06-04T11:54:42Z)
• Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates [31.608080868988825]
  We introduce three types of logic gates: AND, OR, and ADDER gates, and decompose the circuit into combinations of these logical gates.<n>We propose a framework that combines noising-based and denoising-based interventions, which can be easily integrated into existing circuit discovery methods.
  arXiv  Detail & Related papers  (2025-05-15T07:35:14Z)
• Position-aware Automatic Circuit Discovery [59.64762573617173]
  We identify a gap in existing circuit discovery methods, treating model components as equally relevant across input positions. We propose two improvements to incorporate positionality into circuits, even on tasks containing variable-length examples. Our approach enables fully automated discovery of position-sensitive circuits, yielding better trade-offs between circuit size and faithfulness compared to prior work.
  arXiv  Detail & Related papers  (2025-02-07T00:18:20Z)
• Navigating Shortcuts, Spurious Correlations, and Confounders: From Origins via Detection to Mitigation [21.21130450731374]
  Clever Hans behavior, spurious correlations, or confounders, present a significant challenge in machine learning and AI. Research in this area remains fragmented across various terminologies, hindering the progress of the field as a whole. We introduce a unifying taxonomy by providing a formal definition of shortcuts and bridging the diverse terms used in the literature.
  arXiv  Detail & Related papers  (2024-12-06T16:10:13Z)
• Circuit Compositions: Exploring Modular Structures in Transformer-Based Language Models [22.89563355840371]
  We study the modularity of neural networks by analyzing circuits for highly compositional subtasks within a language model. Our results indicate that functionally similar circuits exhibit both notable node overlap and cross-task faithfulness.
  arXiv  Detail & Related papers  (2024-10-02T11:36:45Z)
• Transformer Circuit Faithfulness Metrics are not Robust [0.04260910081285213]
  We measure circuit 'faithfulness' by ablating portions of the model's computation. We conclude that existing circuit faithfulness scores reflect both the methodological choices of researchers as well as the actual components of the circuit. The ultimate goal of mechanistic interpretability work is to understand neural networks, so we emphasize the need for more clarity in the precise claims being made about circuits.
  arXiv  Detail & Related papers  (2024-07-11T17:59:00Z)
• Functional Faithfulness in the Wild: Circuit Discovery with Differentiable Computation Graph Pruning [14.639036250438517]
  We introduce a comprehensive reformulation of the task known as Circuit Discovery, along with DiscoGP. DiscoGP is a novel and effective algorithm based on differentiable masking for discovering circuits.
  arXiv  Detail & Related papers  (2024-07-04T09:42:25Z)
• A Mathematical Theory for Learning Semantic Languages by Abstract Learners [9.139188656944429]
  We develop a mathematical theory to explain the emergence of learned skills, taking the learning process into account. We demonstrate the emergence of learned skills when the ratio of the number of training texts to the number of skills exceeds a certain threshold. We use site percolation analysis to derive the conditions for the existence of a giant component in the skill association graph.
  arXiv  Detail & Related papers  (2024-04-10T13:50:46Z)
• Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models [55.19497659895122]
  We introduce methods for discovering and applying sparse feature circuits. These are causally implicatedworks of human-interpretable features for explaining language model behaviors.
  arXiv  Detail & Related papers  (2024-03-28T17:56:07Z)
• Rethinking Mutual Information for Language Conditioned Skill Discovery on Imitation Learning [36.624923972563415]
  We propose an end-to-end imitation learning approach known as Language Conditioned Skill Discovery (LCSD) We utilize vector quantization to learn discrete latent skills and leverage skill sequences of trajectories to reconstruct high-level semantic instructions. Our approach exhibits enhanced generalization capabilities towards unseen tasks, improved skill interpretability, and notably higher rates of task completion success.
  arXiv  Detail & Related papers  (2024-02-27T13:53:52Z)
• SkillDiffuser: Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution [75.2573501625811]
  Diffusion models have demonstrated strong potential for robotic trajectory planning. generating coherent trajectories from high-level instructions remains challenging. We propose SkillDiffuser, an end-to-end hierarchical planning framework.
  arXiv  Detail & Related papers  (2023-12-18T18:16:52Z)
• Are Structural Concepts Universal in Transformer Language Models? Towards Interpretable Cross-Lingual Generalization [27.368684663279463]
  We investigate the potential for explicitly aligning conceptual correspondence between languages to enhance cross-lingual generalization. Using the syntactic aspect of language as a testbed, our analyses of 43 languages reveal a high degree of alignability. We propose a meta-learning-based method to learn to align conceptual spaces of different languages.
  arXiv  Detail & Related papers  (2023-10-19T14:50:51Z)
• SkillNet-X: A Multilingual Multitask Model with Sparsely Activated Skills [51.74947795895178]
  This paper proposes a general multilingual multitask model, named SkillNet-X. We define several language-specific skills and task-specific skills, each of which corresponds to a skill module. We evaluate SkillNet-X on eleven natural language understanding datasets in four languages.
  arXiv  Detail & Related papers  (2023-06-28T12:53:30Z)
• GIF: A General Graph Unlearning Strategy via Influence Function [63.52038638220563]
  Graph Influence Function (GIF) is a model-agnostic unlearning method that can efficiently and accurately estimate parameter changes in response to a $epsilon$-mass perturbation in deleted data. We conduct extensive experiments on four representative GNN models and three benchmark datasets to justify GIF's superiority in terms of unlearning efficacy, model utility, and unlearning efficiency.
  arXiv  Detail & Related papers  (2023-04-06T03:02:54Z)
• Language Is Not All You Need: Aligning Perception with Language Models [110.51362453720458]
  We introduce Kosmos-1, a Multimodal Large Language Model (MLLM) that can perceive general modalities, learn in context, and follow instructions. We train Kosmos-1 from scratch on web-scale multimodal corpora, including arbitrarily interleaved text and images, image-caption pairs, and text data. Experimental results show that Kosmos-1 achieves impressive performance on (i) language understanding, generation, and even OCR-free NLP. We also show that MLLMs can benefit from cross-modal transfer, i.e., transfer knowledge from language to multimodal, and from multimodal to language
  arXiv  Detail & Related papers  (2023-02-27T18:55:27Z)
• Grounding Language with Visual Affordances over Unstructured Data [26.92329260907805]
  We propose a novel approach to efficiently learn language-conditioned robot skills from unstructured, offline and reset-free data. We exploit a self-supervised visuo-lingual affordance model, which requires as little as 1% of the total data with language. We find that our method is capable of completing long-horizon, multi-tier tasks in the real world, while requiring an order of magnitude less data than previous approaches.
  arXiv  Detail & Related papers  (2022-10-04T21:16:48Z)
• Joint Language Semantic and Structure Embedding for Knowledge Graph Completion [66.15933600765835]
  We propose to jointly embed the semantics in the natural language description of the knowledge triplets with their structure information. Our method embeds knowledge graphs for the completion task via fine-tuning pre-trained language models. Our experiments on a variety of knowledge graph benchmarks have demonstrated the state-of-the-art performance of our method.
  arXiv  Detail & Related papers  (2022-09-19T02:41:02Z)
• VEM$^2$L: A Plug-and-play Framework for Fusing Text and Structure Knowledge on Sparse Knowledge Graph Completion [14.537509860565706]
  We propose a plug-and-play framework VEM2L over sparse Knowledge Graphs to fuse knowledge extracted from text and structure messages into a unity. Specifically, we partition knowledge acquired by models into two nonoverlapping parts. We also propose a new fusion strategy proved by Variational EM algorithm to fuse the generalization ability of models.
  arXiv  Detail & Related papers  (2022-07-04T15:50:21Z)
• On Neural Architecture Inductive Biases for Relational Tasks [76.18938462270503]
  We introduce a simple architecture based on similarity-distribution scores which we name Compositional Network generalization (CoRelNet) We find that simple architectural choices can outperform existing models in out-of-distribution generalizations.
  arXiv  Detail & Related papers  (2022-06-09T16:24:01Z)
• Multi-level Contrastive Learning for Cross-lingual Spoken Language Understanding [90.87454350016121]
  We develop novel code-switching schemes to generate hard negative examples for contrastive learning at all levels. We develop a label-aware joint model to leverage label semantics for cross-lingual knowledge transfer.
  arXiv  Detail & Related papers  (2022-05-07T13:44:28Z)
• Exposing Cross-Lingual Lexical Knowledge from Multilingual Sentence Encoders [85.80950708769923]
  We probe multilingual language models for the amount of cross-lingual lexical knowledge stored in their parameters, and compare them against the original multilingual LMs. We also devise a novel method to expose this knowledge by additionally fine-tuning multilingual models. We report substantial gains on standard benchmarks.
  arXiv  Detail & Related papers  (2022-04-30T13:23:16Z)
• LISA: Learning Interpretable Skill Abstractions from Language [85.20587800593293]
  We propose a hierarchical imitation learning framework that can learn diverse, interpretable skills from language-conditioned demonstrations. Our method demonstrates a more natural way to condition on language in sequential decision-making problems.
  arXiv  Detail & Related papers  (2022-02-28T19:43:24Z)
• Hierarchical Skills for Efficient Exploration [70.62309286348057]
  In reinforcement learning, pre-trained low-level skills have the potential to greatly facilitate exploration. Prior knowledge of the downstream task is required to strike the right balance between generality (fine-grained control) and specificity (faster learning) in skill design. We propose a hierarchical skill learning framework that acquires skills of varying complexity in an unsupervised manner.
  arXiv  Detail & Related papers  (2021-10-20T22:29:32Z)
• A Survey on Extraction of Causal Relations from Natural Language Text [9.317718453037667]
  Cause-effect relations appear frequently in text, and curating cause-effect relations from text helps in building causal networks for predictive tasks. Existing causality extraction techniques include knowledge-based, statistical machine learning(ML)-based, and deep learning-based approaches.
  arXiv  Detail & Related papers  (2021-01-16T10:49:39Z)
• Probing Pretrained Language Models for Lexical Semantics [76.73599166020307]
  We present a systematic empirical analysis across six typologically diverse languages and five different lexical tasks. Our results indicate patterns and best practices that hold universally, but also point to prominent variations across languages and tasks.
  arXiv  Detail & Related papers  (2020-10-12T14:24:01Z)
• GraphOpt: Learning Optimization Models of Graph Formation [72.75384705298303]
  We propose an end-to-end framework that learns an implicit model of graph structure formation and discovers an underlying optimization mechanism. The learned objective can serve as an explanation for the observed graph properties, thereby lending itself to transfer across different graphs within a domain. GraphOpt poses link formation in graphs as a sequential decision-making process and solves it using maximum entropy inverse reinforcement learning algorithm.
  arXiv  Detail & Related papers  (2020-07-07T16:51:39Z)
• Structure-Augmented Text Representation Learning for Efficient Knowledge Graph Completion [53.31911669146451]
  Human-curated knowledge graphs provide critical supportive information to various natural language processing tasks. These graphs are usually incomplete, urging auto-completion of them. graph embedding approaches, e.g., TransE, learn structured knowledge via representing graph elements into dense embeddings. textual encoding approaches, e.g., KG-BERT, resort to graph triple's text and triple-level contextualized representations.
  arXiv  Detail & Related papers  (2020-04-30T13:50:34Z)
• Zero-Shot Cross-Lingual Transfer with Meta Learning [45.29398184889296]
  We consider the setting of training models on multiple languages at the same time, when little or no data is available for languages other than English. We show that this challenging setup can be approached using meta-learning. We experiment using standard supervised, zero-shot cross-lingual, as well as few-shot cross-lingual settings for different natural language understanding tasks.
  arXiv  Detail & Related papers  (2020-03-05T16:07:32Z)
• Automated Relational Meta-learning [95.02216511235191]
  We propose an automated relational meta-learning framework that automatically extracts the cross-task relations and constructs the meta-knowledge graph. We conduct extensive experiments on 2D toy regression and few-shot image classification and the results demonstrate the superiority of ARML over state-of-the-art baselines.
  arXiv  Detail & Related papers  (2020-01-03T07:02:25Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.