Fetching from the wire…
Infra2026-08-25 · source-backed
Activation-Weighted Seeded Residual Coding encodes the residual between true and quantized weights using deterministic seed-generated bases, storing seed selectors, low-bit coefficients and scales instead of an explicit codebook, with activation statistics prioritizing the errors that move layer outputs (arXiv 2608.23144). Adding 0.162 bits per weight to an INT4 RTN backbone closed 88.2% of the perplexity gap, 78.9% of the KL gap and 71.3% of the accuracy gap to BF16. The sidecar is 0.8% of the BF16 weight payload and beat sparse, low-rank and vector-quantized codecs at matched size.
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...
A controlled study ran five Qwen models over eight cases against a DWSIM simulator, 120 slots per arm, with one instruction as the only difference: request a fresh simulation after a substantive modification. No hard gate. Re-verification happened in 94 of 120 guided slots aga...
Eight model-selection strategies across before-generation routing and after-generation voting/judging on hard scientific benchmarks: expanding the candidate pool frequently degraded performance below the single top-performing base model. The only strategy that reliably beat a...
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...
arXiv 2609.09769 argues that relying on the static issue description biases reasoning toward the narrow scope of that text. Adding dynamic behavioral analysis gets 72.8% function-localization accuracy while staying cost-competitive, and resolves 7 issues the top baselines miss...
Community 4-bit quantizations of Qwen3.8-27B kept the 48 GDN layers' decay and write-strength gates at 8 or 16 bits, on the intuition that recurrence errors accumulate. Minima pushes NVFP4 W4A4 through all 496 linear layers including GDN and matches BF16 within seed noise acro...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.