Do Models Read What They Write? Causal Registers in Scratchpad Reasoning

2026-06-28

AI Summary

This study investigates whether large language models genuinely leverage their explicitly generated intermediate states for subsequent reasoning, thereby validating the efficacy of process supervision. To this end, the authors design a controlled state-tracking task that requires models to produce intermediate states before final answers and employ causal interventions by editing internal representations to test whether subsequent predictions adhere to known state-transition rules. The work provides the first empirical evidence of the causal role of intermediate states and introduces the concept of a “causal register,” emphasizing that such states must substantively participate in computation. Experiments show that Qwen2.5-Coder-7B achieves 80%–91% prediction accuracy under state edits—significantly outperforming models that output only final answers or rely on pretrained baselines—with consistent results replicated across multiple model families.

Abstract

A central hope behind process supervision is that models can expose intermediate variables that matter for their later behavior. For this to help with alignment, a scratchpad must be tied to the computation: when the model writes a state, later steps should compute from that state. To test this requirement, we use a controlled state-tracking task with a known update rule, comparing models trained to report only the final state with models trained to write intermediate states before giving the final answer. At evaluation, we edit the internal representation of one written state while leaving the visible scratchpad text fixed. Because the transition rule is known, the edit has a single correct downstream consequence. In Qwen2.5-Coder-7B, the state-writing model predicts the next phase bit implied by the edited state on 80% and 91% of held-out examples across the two task variants, while pretrained and final-answer-only controls remain near baseline. Additional controls rule out generic next-token steering and copying another continuation: the prediction depends on both the edited state and the current move. The same causal-use pattern replicates across model families. Together, these results suggest a sharper goal for scratchpad oversight: not just to make intermediate reasoning legible, but to train written states that the model uses as part of its computation.

Problem

• scratchpad reasoning
• causal registers
• process supervision
• state tracking
• model alignment

Innovation

• causal registers
• scratchpad reasoning
• process supervision
• state tracking
• intermediate computation