Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-03-16 · source-backed
Uses confidence metrics and steering vectors to dynamically guide reasoning depth — reduces output redundancy while improving accuracy across 9 benchmarks, 4 model sizes, no training cost. Drop-in for o1-style reasoning models. arXiv
Each link below shares sources, entities, or timing with this story.
IndexCache identifies that DeepSeek Sparse Attention reduces core attention to O(Lk) but its lightning indexer retains O(L²) complexity. By exploiting high cross-layer index stability to reuse token selection indices, IndexCache delivers significant wall-clock throughput impro...
PCAS: Policy Compiler for Secure Agentic Systems — The first paper to provide measured enforcement results for agent policy compliance (48% to 93%). Uses dependency graphs and Datalog-derived policy language with a reference monitor intercepting all actions. Three case studies...
arXiv 2609.20734 shows a pretrained model's decoding states already carry information predictive of whether a global read will help, before the read happens. Training only a small recall head to invoke global attention selectively, with pretrained weights untouched and the ful...
arXiv 2609.09134 evolved a harness with a weak model, watched a stronger expert use that harness better, then tried the obvious fix. It cost 4 to 30 points on Qwen3-Coder and Gemma 4, even though the same procedure helps under the unevolved harness. The diagnosis is model-harn...
SMITH (arXiv 2608.24571) points at a real gap in existing tool-creation systems: they prompt a frozen LLM at inference time, so the model writing a schema gets no signal about whether it can invoke that schema. SMITH alternates build rollouts and use rollouts inside one RL pol...
The attack iteratively pulls hidden chain-of-thought from black-box reasoning models using API-returned fidelity signals, reaching 66.4% near-verbatim extraction on open-source LRMs (trace length within 10% of target, 90%+ tokens matching exactly), generalizing to unseen datas...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.