A Mechanistic Lens on Semantic Conflicts: Using Activation Patching to Understand LLM Behavior

2026-07-06

AI Summary

This study investigates how large language models (LLMs) resolve conflicts between code logic and semantic cues—such as comments or identifiers—when they diverge. By constructing 45 token-aligned Python triplets that systematically manipulate either semantic hints or functional logic, the authors combine behavioral evaluation with residual stream activation patching to conduct the first controlled, systematic analysis of how LLMs integrate contradictory semantic information. The findings reveal that semantic conflicts substantially degrade model accuracy, with misleading semantic cues often dominating model outputs. Crucially, causal signals driving these decisions are concentrated in a small subset of residual stream states and are aggregated prior to output generation—a pattern that also extends internally within generative tasks. This work establishes a general mechanistic framework for analyzing how LLMs handle semantic conflicts in code.

Abstract

Large language models (LLMs) are increasingly used in software-engineering tasks processing executable code and non-executable semantic cues such as comments or identifiers. These two sources can conflict when semantic cues suggest different program behavior than the code itself. It remains unclear how such semantic conflicts affect LLM behavior and which source dominates their outputs. We present the first controlled, mechanistic study of LLM behavior under semantic conflicts. To this end, we construct 45 Python snippet triplets that isolate conflicts by varying either semantic cues or implementation while keeping token-aligned pairs for causal intervention. We evaluate four open-weight LLMs on two tasks (output prediction and unit-test generation) using behavioral performance measures and residual-stream activation patching to identify token-layer states that causally contribute to behavioral differences between aligned and conflicting inputs. Our results show that semantic conflicts significantly reduce execution-grounded correctness in both tasks and that all tested LLMs often follow misleading semantic cues. Residual-stream activation patching reveals a consistent pattern for final-output prediction: The changed cue/code region and a small set of intermediate tokens carry most of the recoverable causal signal before aggregation near the output readout. For unit-test generation, this pattern extends beyond the prompt, showing that conflict-related information is recoverable at generated sites before producing expected values. Overall, our findings show that semantic conflicts affect program comprehension and downstream tasks, with relevant information concentrated in a small number of causally active residual-stream states, and demonstrate a framework for mechanistically analyzing how LLMs integrate code-related information under controlled semantic variations.

Problem

• semantic conflicts
• large language models
• code comprehension
• executable code
• semantic cues

Innovation

• semantic conflicts
• activation patching
• mechanistic interpretability
• code comprehension
• residual stream