In-Context Learning Operates as Concept Subspace Learning

arXiv cs.LG Papers
in-context-learning concept-subspace llm interpretability mechanistic-analysis representation-learning transformer

Summary

This paper proposes that in-context learning in LLMs operates through low-dimensional concept subspaces, where task-relevant information concentrates in a small fraction of the representation space, supported by experiments on Llama-3-8B and Qwen2.5-7B.

arXiv:2605.18830v1 Announce Type: new Abstract: Regression and Bayesian accounts of in-context learning (ICL) explain how demonstrations can induce predictors, while mechanistic analyses often identify compact activation directions that steer prompted behavior. However, it remains unclear whether structured demonstrations induce low-dimensional concept inference. We study this question through a concept-subspace view of ICL, in which tasks vary only along intrinsic concept coordinates, although inputs are observed in a high-dimensional ambient space. For ridge and least-squares ICL proxies, prediction decomposes exactly into concept-coordinate regression and off-subspace leakage. Under block-diagonal or near-block-diagonal covariance assumptions, the leading estimation and nuisance-sensitivity terms scale with the dimension of the concept subspace, while residual effects are controlled by cross-subspace coupling. This separation gives a mechanistic prediction: recoverable task information should concentrate in a low-dimensional, task-aligned activation subspace. On CounterFact-derived multi-relation prompts with Llama-3-8B, a 68--73-dimensional subspace of the 4096-dimensional residual stream restores 78.8% of the clean--corrupted accuracy gap, whereas patching the complementary subspace restores 0%. Concept swaps redirect predictions toward injected relations, while random and cross-task matched-rank controls are largely ineffective. Additional experiments on Qwen2.5-7B and a controlled cross-lingual rule task show the same qualitative pattern. These results support concept subspaces as compact, task-aligned mediators of recoverable ICL behavior in structured task families, without implying full-circuit recovery.
Original Article

Similar Articles

Context-Aware RL for Agentic and Multimodal LLMs

Hugging Face Daily Papers

Introduces ContextRL, a reinforcement learning approach that teaches LLMs to identify which context supports an answer, achieving gains on agentic and multimodal benchmarks.

When Does Continual Learning Require Learning

arXiv cs.LG

This paper proposes a unified framework for continual learning in LLMs, disentangling change along space (new domains) and time (data drift). It evaluates various methods including prompting, supervised learning, reinforcement learning, and context compression under realistic sequential settings.

Learning, Fast and Slow: Towards LLMs That Adapt Continually

Hugging Face Daily Papers

A fast-slow learning framework for LLMs combines fixed slow weights with optimized fast context weights, achieving up to 3x better sample efficiency and reduced catastrophic forgetting in continual learning scenarios.