$K$-MSHC: Unmasking Minimally Sufficient Head Circuits in Large Language Models with Experiments on Syntactic Classification Tasks

• URL: http://arxiv.org/abs/2505.12268v2
• Date: Thu, 05 Jun 2025 01:45:59 GMT
• Title: $K$-MSHC: Unmasking Minimally Sufficient Head Circuits in Large Language Models with Experiments on Syntactic Classification Tasks
• Authors: Pratim Chowdhary, Peter Chin, Deepernab Chakrabarty,
• Abstract summary: We introduce the $(bmK, epsilon)$-Minimum Sufficient Head Circuit, a methodology to identify minimal sets of attention heads crucial for classification tasks.<n>Applying our Search-K-MSHC algorithm to Gemma-9B, we analyze three syntactic task families: grammar acceptability, arithmetic verification, and arithmetic word problems.<n>Our findings reveal distinct task-specific head circuits, with grammar tasks predominantly utilizing early layers, word problems showing pronounced activity in both shallow and deep regions, and arithmetic verification demonstrating a more distributed pattern across the network.
• Score: 3.767957313558699
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Understanding which neural components drive specific capabilities in mid-sized language models ($\leq$10B parameters) remains a key challenge. We introduce the $(\bm{K}, \epsilon)$-Minimum Sufficient Head Circuit ($K$-MSHC), a methodology to identify minimal sets of attention heads crucial for classification tasks as well as Search-K-MSHC, an efficient algorithm for discovering these circuits. Applying our Search-K-MSHC algorithm to Gemma-9B, we analyze three syntactic task families: grammar acceptability, arithmetic verification, and arithmetic word problems. Our findings reveal distinct task-specific head circuits, with grammar tasks predominantly utilizing early layers, word problems showing pronounced activity in both shallow and deep regions, and arithmetic verification demonstrating a more distributed pattern across the network. We discover non-linear circuit overlap patterns, where different task pairs share computational components at varying levels of importance. While grammar and arithmetic share many "weak" heads, arithmetic and word problems share more consistently critical "strong" heads. Importantly, we find that each task maintains dedicated "super-heads" with minimal cross-task overlap, suggesting that syntactic and numerical competencies emerge from specialized yet partially reusable head circuits.

Related papers

• Operationalising the Superficial Alignment Hypothesis via Task Complexity [49.93635747700126]
  We propose a new metric called task complexity: the length of the shortest program that achieves a target performance on a task.<n>Our results highlight that task adaptation often requires surprisingly little information -- often just a few kilobytes.
  arXiv  Detail & Related papers  (2026-02-17T18:59:39Z)
• Teaching Transformers to Solve Combinatorial Problems through Efficient Trial & Error [18.209374823705446]
  We focus on the paradigmatic task of Sudoku and achieve state-of-the-art accuracy (99%) compared to prior neuro-symbolic approaches.<n>Our method integrates imitation learning of simple Sudoku rules with an explicit Depth-First Search (DFS) exploration strategy.
  arXiv  Detail & Related papers  (2025-09-26T07:57:34Z)
• Do Attention Heads Compete or Cooperate during Counting? [0.12116854758481393]
  We present an in-depth mechanistic interpretability analysis of training small transformers on an elementary task, counting.<n>We ask whether the attention heads behave as a pseudo-ensemble, all solving the same subtask, or they perform different subtasks, meaning that they can only solve the original task in conjunction.
  arXiv  Detail & Related papers  (2025-02-10T17:21:39Z)
• An Analysis under a Unified Fomulation of Learning Algorithms with Output Constraints [5.10832476049103]
  Neural networks (NN) perform well in diverse tasks, but sometimes produce nonsensical results to humans. Injecting human knowledge by reducing output constraints during training can improve model performance and reduce constraint violations. We propose new algorithms to integrate the information of main task and constraint injection.
  arXiv  Detail & Related papers  (2024-06-03T12:58:29Z)
• An Examination on the Effectiveness of Divide-and-Conquer Prompting in Large Language Models [28.139780691709266]
  We provide a theoretic analysis to divide-and-conquer prompting strategy and help us identify the specific tasks where DaC prompting can bring performance boost with theoretic guarantee. We present two cases (large integer arithmetic and fact verification) where experimental results align with our theoretic analysis.
  arXiv  Detail & Related papers  (2024-02-08T02:37:30Z)
• DCR: Divide-and-Conquer Reasoning for Multi-choice Question Answering with LLMs [9.561022942046279]
  We propose Divide and Conquer Reasoning (DCR) to enhance the reasoning capability of large language models (LLMs) We first categorize questions into two subsets based on confidence score ($mathcalCS$), which is estimated by statistical frequency of generated answers. In particular, we first categorize questions into two subsets based on confidence score ($mathcalCS$), which is estimated by statistical frequency of generated answers.
  arXiv  Detail & Related papers  (2024-01-10T14:38:46Z)
• Self-paced Weight Consolidation for Continual Learning [39.27729549041708]
  Continual learning algorithms are popular in preventing catastrophic forgetting in sequential task learning settings. We propose a self-paced Weight Consolidation (spWC) framework to attain continual learning.
  arXiv  Detail & Related papers  (2023-07-20T13:07:41Z)
• Object-Centric Multi-Task Learning for Human Instances [8.035105819936808]
  We explore a compact multi-task network architecture that maximally shares the parameters of the multiple tasks via object-centric learning. We propose a novel query design to encode the human instance information effectively, called human-centric query (HCQ) Experimental results show that the proposed multi-task network achieves comparable accuracy to state-of-the-art task-specific models.
  arXiv  Detail & Related papers  (2023-03-13T01:10:50Z)
• Unsupervised Learning of Initialization in Deep Neural Networks via Maximum Mean Discrepancy [74.34895342081407]
  We propose an unsupervised algorithm to find good initialization for input data. We first notice that each parameter configuration in the parameter space corresponds to one particular downstream task of d-way classification. We then conjecture that the success of learning is directly related to how diverse downstream tasks are in the vicinity of the initial parameters.
  arXiv  Detail & Related papers  (2023-02-08T23:23:28Z)
• Are Deep Neural Networks SMARTer than Second Graders? [85.60342335636341]
  We evaluate the abstraction, deduction, and generalization abilities of neural networks in solving visuo-linguistic puzzles designed for children in the 6--8 age group. Our dataset consists of 101 unique puzzles; each puzzle comprises a picture question, and their solution needs a mix of several elementary skills, including arithmetic, algebra, and spatial reasoning. Experiments reveal that while powerful deep models offer reasonable performances on puzzles in a supervised setting, they are not better than random accuracy when analyzed for generalization.
  arXiv  Detail & Related papers  (2022-12-20T04:33:32Z)
• Fast Inference and Transfer of Compositional Task Structures for Few-shot Task Generalization [101.72755769194677]
  We formulate it as a few-shot reinforcement learning problem where a task is characterized by a subtask graph. Our multi-task subtask graph inferencer (MTSGI) first infers the common high-level task structure in terms of the subtask graph from the training tasks. Our experiment results on 2D grid-world and complex web navigation domains show that the proposed method can learn and leverage the common underlying structure of the tasks for faster adaptation to the unseen tasks.
  arXiv  Detail & Related papers  (2022-05-25T10:44:25Z)
• Learning to Detect Critical Nodes in Sparse Graphs via Feature Importance Awareness [53.351863569314794]
  The critical node problem (CNP) aims to find a set of critical nodes from a network whose deletion maximally degrades the pairwise connectivity of the residual network. This work proposes a feature importance-aware graph attention network for node representation. It combines it with dueling double deep Q-network to create an end-to-end algorithm to solve CNP for the first time.
  arXiv  Detail & Related papers  (2021-12-03T14:23:05Z)
• Multi-task Supervised Learning via Cross-learning [102.64082402388192]
  We consider a problem known as multi-task learning, consisting of fitting a set of regression functions intended for solving different tasks. In our novel formulation, we couple the parameters of these functions, so that they learn in their task specific domains while staying close to each other. This facilitates cross-fertilization in which data collected across different domains help improving the learning performance at each other task.
  arXiv  Detail & Related papers  (2020-10-24T21:35:57Z)
• Machine Number Sense: A Dataset of Visual Arithmetic Problems for Abstract and Relational Reasoning [95.18337034090648]
  We propose a dataset, Machine Number Sense (MNS), consisting of visual arithmetic problems automatically generated using a grammar model--And-Or Graph (AOG) These visual arithmetic problems are in the form of geometric figures. We benchmark the MNS dataset using four predominant neural network models as baselines in this visual reasoning task.
  arXiv  Detail & Related papers  (2020-04-25T17:14:58Z)

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.