Length Effects in Chain-of-Thought Are Conditional on Explicit Answer-Cue Availability: Experimental Data and Analysis Code

This dataset accompanies the paper "Length Effects in Chain-of-Thought Are Conditional on Explicit Answer-Cue Availability." We study chain-of-thought (CoT) trace consumption under controlled perturbations on GSM8K (n = 188 problems), independently manipulating trace length (L ∈ {5, 10, 15, 20}), answer-cue availability (present vs. absent), and trace corruption (c ∈ {0, 0.4, 0.8}). Key findings:- Length effects are strictly cue-conditional: with cues present, accuracy saturates at ceiling (ΔL ≈ 0 pp); without cues, strong length benefits emerge (ΔL = +14 to +16 pp)- Models are not blind extractors: incorrect cues are rejected, correct cues under corruption rescue performance differentially (Claude 99.5% vs. GPT-4o 75.9%)- Corruption sensitivity—the threshold for abandoning extraction—explains cross-model divergence This repository contains:- Raw experimental data (model responses for all conditions)- Generated reasoning traces at multiple length levels- Corruption manipulation scripts- Statistical analysis code- Figure generation scripts Models evaluated: Claude Sonnet 4 (claude-sonnet-4-20250514), GPT-4o (gpt-4o-2024-08-06)