Research & policy
Academic papers, official research, regulatory material, patents, and standards are grouped together with their evidence labels intact.
arXiv
Nov 05, 2026
Measured coupling trades a small accuracy loss for faster diffusion decoding
This preprint replaces similarity-based token coupling with probes measuring how committing one token changes confidence at another. Across three models and four benchmarks, the authors report fewer decoding steps and faster execution in most tested settings, with an average accuracy cost of 1.1 points. The method needs no extra model or training.
- The authors report an average 1.1-point accuracy cost for their tested decoding acceleration.
Why it mattersThe paper offers a separate latency–accuracy trade-off from exact speculative sampling. Compare quality-matched cost and hardware utilization; the reported average does not establish an improvement for every workload.
arXiv
Nov 05, 2026
SpecCTMC accelerates exact diffusion sampling without a draft model
SpecCTMC proposes parallel speculative updates followed by acceptance and correction for continuous-time diffusion language-model sampling. The authors claim preservation of the reference joint sampling distribution and report 1.73–4.38 times fewer sequential model evaluations on two models across five benchmarks. Evaluation-count reduction is not an end-to-end production speedup.
- The preprint reports fewer sequential evaluations while retaining the underlying continuous-time sampling distribution.
Why it mattersA distribution-preserving acceleration path could make diffusion inference easier to assess. Replicate wall-clock latency, parallel hardware cost and memory demand before translating the result into serving economics.
arXiv
Nov 05, 2026
ReproBridge tests whether reusable execution artifacts improve research reproduction
The ReproBridge preprint packages refined code, environment setup, launchers and reproduction guides into reusable research artifacts. Across four models, its authors report reproduction success rising from 76.4% to 92.6% and API spending falling 19.7%, on 1,029 tasks drawn from 387 experiments in 36 papers. These are results within the authors’ benchmark.
- Authors report higher benchmark reproduction success and lower API spending after refining the released artifacts.
Why it mattersPreserving execution knowledge may improve reliability more cheaply than repeatedly rediscovering it. Check artifact selection, auditor independence and transfer to an unseen research portfolio before forecasting production savings.
arXiv
Nov 05, 2026
Scientist World turns published methods into executable research environments
Scientist World releases 1,000 verified environments that reconstruct research questions and baseline methods from papers. Its authors report that evaluated AI scientists exceed those reconstructed baselines in 74% of trials and release 6,400 hours of research traces. This compares against selected executable baselines; it does not establish superiority to human scientists in open-ended research.
- The released environments evaluate agents relative to reconstructed paper baselines.
Why it mattersExecutable research environments could support reproducible evaluation and training. Audit reconstruction fidelity, task selection, leakage and held-out performance before accepting the headline success rate as general scientific capability.