Latest

Every published item, in order · 7652 items
Oct 26
Oct 23
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
PapersBuildersSelected10:54
Study Finds Speculative Decoding Acceptance Does Not Decay Over Long Chain-of-Thought

The finding overturns the widely repeated narrative that acceptance rate falls as chains grow, which has driven draft-model co-design and position-aware training spending. Teams building inference acceleration should redirect effort from positional engineering toward entropy at decision points, where the paper localizes the real failure. The MATH-500 baseline and open diagnostic protocol are directly reusable for benchmarking any draft-target pairing.

arXiv: DeepSeek@Zhiming MengSource
PapersBuildersSelected10:54
Behavioral Counterfactuals Method Traces How Prompts Sway Model Reasoning

Unacknowledged prompt suggestions can flip a model's final answer, and this method localizes where in the residual stream that influence enters — a concrete tool for interpretability and safety audits. Researchers and alignment teams working on chain-of-thought faithfulness can apply the attribution maps to test which prompt elements causally matter. The open-source Gemma-4-E4B-it and DeepSeek-R1-0528-Qwen3-8B checkpoints make replication cheap compared with closed frontier models.

arXiv: DeepSeek@Shawn Im et al. (4)Source
PapersBuildersSelected10:49
Simulated Control Evals Predict AI Safety With Lower Error

Control evals have so far been limited to purpose-built settings whose transfer to real deployment is questionable; SCEs let safety teams test safeguards against deployment-specific environments cheaply, without executing every task. The 3.6 vs 10.1 percentage-point RMSE gap against BashArena is the concrete benchmark to cite when arguing simulation fidelity. AI safety researchers and red-team builders can apply the technique by generating simulated tool responses for their own target environments.

Event
arXiv cs.AI@Jeffrey Olmo et al. (3)Source

Only the latest 2000 items are listed; for older ones see the daily archive