Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-24 · high
A training-free method retrieves reasoning memories at prefill, recovering accuracy lost to compression while running faster than standard chain-of-thought.
Why now: It's an August 2026 arXiv paper, arriving as compressed reasoning chains trade accuracy for speed across the field.
A training-free method recovers 21 to 29 points of reasoning accuracy lost to Chain-of-Draft compression, per a paper posted to arXiv.
That gap costs anyone running compressed reasoning chains in production, where every dropped point of accuracy is a real error downstream. Chain-of-Draft cuts tokens and latency, but the shorter chains lose accuracy, and this method recovers it without retraining the model.
The method builds memories from historical reasoning traces that summarize patterns, key constraints, and critical operations. It retrieves those memories at prefill to guide the compressed chain, without touching the underlying model's weights. It recovered 21.4 points on GSM8K, 28.0 on MATH, 29.5 on BBH, and a smaller 6.61 on MMLU-Sci. It also ran 1.14 to 1.49 times faster than standard, uncompressed chain-of-thought.
The paper doesn't say how these memories perform outside math and logic benchmarks, or how retrieval scales as the memory store grows. Both are open questions before anyone relies on this in a live deployment.
Each link below shares sources, entities, or timing with this story.
DeepMind's Delegation Capability Tokens paper (arXiv) is the most important agent security paper since the MCP specification. It formally solves the delegation problem: how do agents safely give other agents scoped permissions? The cryptographic caveat system enables least-pri...
For non-reasoning models on math and logic, sampling multiple CoT paths and majority-voting lifts GSM8K by +17.9% over a single greedy pass. Highest-ROI accuracy lever for reasoning-heavy prompts. But skip it entirely on native reasoning models, they already reason internally...
Reasoning models engage in "performative" CoT — the model's final answer is decodable from activations far earlier than visible CoT suggests. Activation probing enables up to 80% token reduction on MMLU. Critical for safety monitoring: visible reasoning may not reflect actual...
arXiv 2608.04735 points out that monitorability evals overwhelmingly use *explicit* influence, where the prompt tells the model to hide a side task, and monitors catch 60-94% across seven frontier extended-thinking models. Swap in subtle contextual bias and detection drops 41-...
Rewriting an agent's reasoning to read as good-faith engineering while copying every command and output verbatim, so the exploit itself is unchanged, defeats a held-out CoT monitor almost completely. The sharper point for anyone running a monitor in production: headline accura...
31. arXiv — Chain-of-Draft ---
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.