A string of dots lifts GPT-6's 4-hop reasoning from 10% to 50%, Redwood Research finds

新智元 · wechat · 2026-09-25

Redwood Research reports that appending meaningless filler tokens to prompts dramatically boosts GPT-6 Astra's reasoning—while other models barely respond.

Key results: On 4-hop serial reasoning, Astra's accuracy jumped from 10% to 50%; on modified old AIME problems, from 60% to 90%. Controls (Opus4.5, Opus5, GPT-5.6-Sol) showed little to no effect.
Details: Gains peak around 8,192 filler tokens; fillers help only when placed after the question; counting or repetition works similarly.
Setup caveat: The API reported zero reasoning tokens, but that only reflects interface logging, not the model's internal computation.
Safety implication: If much of a model's computation never appears in its chain-of-thought, CoT-based monitoring may miss critical behavior—and no-CoT evals may underestimate capability. The team recommends adding filler-token tests to future no-reasoning evals.

Original post →

A string of dots lifts GPT-6's 4-hop reasoning from 10% to 50%, Redwood Research finds

More from Models