Redwood Research: Filler tokens boost GPT-6-Astra reasoning, complicating monitoring
Redwood Research reports that GPT-6-Astra performs significantly better on serial reasoning tasks when prompts are padded with meaningless filler tokens, improving from ~10% to ~50% on 4-hop natural facts reasoning and from ~60% to ~90% on old AIME problems. The finding raises interpretability concerns because the model can perform substantial cognition without verbalizing it in chain-of-thought, making monitoring harder.
Coverage timeline
Redwood ResearchDylan Xu
We measure GPT-6-Astra’s capabilities when its prompt is padded with a variable number of meaningless “filler” tokens (e.g., dots) and it is told to answer immediately 1 without reasoning. On tasks designed to require lots of serial cognition, Astra performs significantly better with filler tokens than without (e.g., improving from ~10% to ~50% on 4-hop natural facts reasoning). On more general benchmarks, filler tokens also modestly improve Astra’s performance (e.g., improving from ~60% to ~90% on old AIME problems). This is concerning because it means Astra can perform significant cognition that it doesn’t verbalize in its chain-of-thought, making it harder to monitor. We first measure Astra’s performance on “N-hop natural facts” : a task that asks the model to retrieve some natural language facts in succession, similar to Ryan Greenblatt’s filler token eval (but with more hops). An example question in this benchmark is the following: Thanks for reading Redwood Research blog! Subscri
Related stories
SenseTime-led study: GPT-6-Astra scores near-perfect on benchmarks but fails real-robot tasks
A research report by SenseTime, Daxia Robotics, Ropedia, NTU Singapore, and the University of Tokyo analyzes GPT-6-Astra's embodied intelligence across 12 simulation benchmarks and 8 real-robot platforms. The model achieves near-perfect scores on traditional benchmarks (e.g., 100% on LIBERO-Long++, 99.3% on LIBERO-Pro) but fails on real-robot tasks, such as taking over 4 hours and $55 to place a pen in a holder, and failing 11 insertion attempts. The report concludes GPT-6-Astra has largely crossed the embodied intelligence threshold but remains far from practical deployment.
1 reports · 1 sources
Goodfire researcher argues neural networks are largely interpretable, most neurons understandable
Nick Cammarata of Goodfire published an essay arguing that neural networks are far more interpretable than previously believed, with most neurons in models up to trillions of parameters understandable. He attributes polysemanticity in early models to fixable issues rather than superposition, and suggests mechanistic interpretability may be solved soon.
.png)
LatePost Embodied-AI Quarterly: GPT-6 Astra Raises Question of Whether General Models Will Absorb Robotics
LatePost's Q3 2026 embodied-AI quarterly report examines whether general multimodal foundation models like OpenAI's GPT-6 Astra will absorb embodied intelligence, noting that Astra can directly output robot end-effector poses and outperform some embodied models on certain tasks. The report also covers AMD's planned $8.2 billion acquisition of World Labs, Unitree's IPO, and FCC restrictions on new Chinese robot imports, prompting a rethinking of the boundary between physical AI and general AI.
1 reports · 1 sources
GPT-6 Astra Helps Prove Remaining Cases of 57-Year-Old Circle Packing Problem for N up to 20
A math enthusiast named nor used GPT-6 Astra to rigorously prove the remaining cases (N=15, 16, 17, 18, 20) of the 'packing circles in a circle' problem, extending verified results to N≤20. Previously, only N≤14 and N=19 had been proven. The problem, posed in 1969, asks for the minimal radius of a large circle that can contain N unit circles without overlap.
1 reports · 1 sources
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe