Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-09-23 · source-backed
arXiv 2609.26621 tested six 1.1B-7B models across four families. One token flip cascades into a different trajectory. The flip is governed by the top-two logit margin at the LM head, not accumulated layer error, and broader FP32 compute makes agreement worse. Recomputing only the LM head in FP32 when the margin is small adds 22-36 points of exact agreement on A10G. Anyone treating temperature 0 as reproducible in evals or regression tests is measuring something else.
Each link below shares sources, entities, or timing with this story.
arXiv 2607.28495 tests the assumption that replaying identical token prefixes reconstructs the decoder state that produced them. On a Qwen2.5-derived system with a matched 200-item experiment, retained live cache versus one-shot prefill of identical integer tokens diverge on 1...
An independent reproduction tested the claim that Orthrus's intra-model consensus produces the same output sequence as the autoregressive backbone (arXiv 2609.15504). Under BF16, exact trajectory matching happens in 45% of cases for the authors' checkpoint and 43% for an indep...
A 15-author Huawei team argues kernel-generation benchmarks are almost entirely CUDA and Triton, leaving less-documented hardware with no shared yardstick (arXiv 2607.20518). CANN Bench covers 53 operators and 1,060 test cases in four difficulty tiers, from elementwise primiti...
The proposed elements include agent memory and memory access records, actual versus potential autonomy level, and tool usage, none of which existing AI incident frameworks capture. The experts also flagged the reporting pipeline as its own attack surface, since incident record...
Researchers loaded five systems with a revoked policy and its replacement, then measured retrieval and downstream action across nine policy scenarios, nine models and six defense conditions. Wherever the revocation label was visible to the retrieval layer, the revoked fact cam...
koala73/worldmonitor, an AGPL-3.0 real-time geopolitical dashboard, sits at 84,987 stars with 12,783 forks and was pushed today, but its last two tags are v2.5.23 and v2.5.21, both from March 1. Six months of daily commits, nothing cut. Egonex-AI/Understand-Anything is at 81,0...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.