Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-26 · high
Crase replaces the agent loop with a fixed 1.5-hop citation walk, cutting compute to about a third of rivals' cost.
Why now: Crase's arXiv paper is a rare head-to-head measurement of agent autonomy against a narrower, fixed alternative.
A new retrieval system called Crase skips the agent loop for scholarly search. It was tested against deep research agents on proprietary models over a 500K-paper arXiv corpus. It delivered up to 3x the recall@50 at about a third of the cost.
Rather than turning a language model loose to search, read, and judge relevance on its own, it runs four fixed steps. One search-engine query pulls seed papers. The system then expands 1.5 hops through the citation graph and prunes any citation edge whose claim lacks entailment support. What's left gets ranked by a recency-aware random walk.
The candidate set, the reason each paper survives pruning, and the stopping condition are all decided in advance. Someone can trace exactly why a given paper made the cut.
That's a direct hit on the assumption that more agent autonomy always means better search. Citation graphs already encode the relationships a literature search needs, so an agent rediscovering that structure step by step burns tokens without adding accuracy. Anyone building retrieval over a corpus with known relationships, citations, imports, dependency graphs, should ask a question first. Is an open-ended agent solving a problem a fixed traversal already handles?
Each link below shares sources, entities, or timing with this story.
When an agent consolidates an external observation into long-term memory, attach platform-controlled metadata recording the source's trust level, then gate tool execution by matching action risk against supporting-memory authority. Laundered memories hit a 1.000 attack success...
arXiv 2607.26935 argues the human-vs-bot label space can't represent agent traffic: an MLP binary classifier misroutes 39.1% of real agent sessions as human, a SAINT transformer 34.5%, while adding an explicit third class yields agent F1 = 1.000 across all 30 runs. Against a f...
Two thirds. Not two thirds of a contrived jailbreak set. Two thirds of realistic malicious issue requests, against the exact three tools most of the people reading this run daily. Ankur Singh, Jinqiu Yang, and Tse-Hsun Chen built IssueTrojanBench across four attack categories...
arXiv 2607.06283 attacks the problem that growing skill libraries make selection harder. It decomposes on both the task and skill side, builds a DAG with intermediate task states as nodes and candidate skills as edges, then cross-encodes over candidates per task interval. On A...
Li, Huo, and Johnson show that one-way message flow between agents produces neither mimicry nor solo behavior but an entirely novel dynamical state, at identical temperature settings. It's conceptual rather than quantitative, but the implication for orchestrator-worker fan-out...
Wafer.ai ran K3 at TP8 on a single MI355X node: 952 tok/s aggregate, 118 tok/s single-stream decode, ~13k tok/s steady-state prefill after fixing missing PyTorch sampling functions and AITER MLA prefill kernels via SGLang and ROCm. Against a two-node TP16 B200 deployment that'...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.