CoT may be a misleading proxy: models can pass richer messages via KV cache
stochasticchasm · x · 2026-09-04
The author argues that even if chain-of-thought looks like a decent proxy for a model's internal process, it can be misleading: models can pass messages through the KV cache that are far more complex than what appears in token space, so readable CoT may not reflect what's happening inside the model.
Related event: Looped Transformer Rumors Spark Fierce Debate Over CoT Monitorability(7 posts)→
More from Research
- Alibaba-NLP CORE: boosting compositional reasoning in MLLM embeddings via reranker distillation — Alibaba-NLP · 2026-09-04
- On-policy distillation improves for hundreds of steps from a single query — it's algorithm-starved, not data-starved — Thinking-Space · 2026-09-04
- WorldReward: a vision-language reward model for evaluating camera-conditioned world models — Yibin Wang · 2026-09-04
- Salesforce research: Random eviction of reasoning tokens matches selective KV cache compression — Salesforce · 2026-09-04
- Google Trends quietly redraws samples daily, shrinking significance in 70% of replicated econ papers — RexDouglass · 2026-09-04
- Immunologist builds research-grade flow cytometry software with GPT-6 Astra, cancels all commercial subscriptions — DeryaTR_ · 2026-09-04