Papers

AI papers from arXiv and journals, each with a summary and why it matters · 7705 papers
Nov 8
PapersBuildersSelected12:10
KOSMOS Self-Scheduling Kernels Speed Up Training on AMD GPUs

Fixed-role fused kernels leave idle windows because a predetermined communication-compute split cannot track actual data readiness; KOSMOS replaces static schedules with demand-ordered queues that adapt at runtime. Engineers training large MoE models on AMD Instinct hardware can apply it directly through Megatron-LM, where it outperforms DeepEP by up to 1.83x on DeepSeek-V3 and Qwen3-235B proxies.

arXiv: DeepSeek@Alexander Magro et al. (4)Source
Nov 6
PapersBuildersSelected13:40
Survey Maps Mechanistic Interpretability Methods for AI Safety

The survey's value lies in its proposed evaluation framework and its six-way taxonomy of research orientations, which gives structure to a field that has grown mostly through scattered techniques. Researchers and safety teams can use the released notebooks to reproduce methods hands-on rather than reading only about them. The comparison between mechanistic evidence and behavioural evaluation is a useful lens for anyone deciding whether to invest in interpretability tools over standard safety testing.

arXiv cs.LG@Fazl Barez et al. (27)Source
PapersBuildersSelected12:56
Strong-to-Weak OPD Matches GRPO Accuracy With 50x Fewer FLOPs

The practical lesson is that the teacher choice decides whether OPD is worth the compute: self-distillation from a model's own GRPO checkpoint is far cheaper than rerunning reinforcement learning, while generic strong teachers are not. Practitioners training domain-specialised models can consolidate separate GRPO experts via multi-teacher OPD at a fraction of the cost. Readers comparing RL and distillation budgets will find a reusable evaluation framework here.

arXiv: DeepSeek@Zeyu Leo Liu et al. (10)Source
PapersUS vs ChinaSelected12:54
CoRE Benchmark Exposes LLM Data-Cleaning Reliance on Known Entities

Benchmark scores on data cleaning may measure memorised entity knowledge rather than repair skill, so leaderboard rankings deserve scepticism when the underlying rows are public. Teams using LLMs to clean operational data should test on renamed surrogates or fresh records before trusting reported accuracy. The released controlled tables and offline replay scoring code make it straightforward to rerun the protocol on in-house data.

Event
arXiv: DeepSeek@Jingru Yang et al. (6)2 sourcesSource
PapersBuildersMulti-source12:01
Chinese contentTCSExam 基准测试:评估前沿 LLM 在理论计算机科学研究中的能力

理论计算机科学版的 LLM 考试来了:GPT-6 Astra 和 Opus 5.5 分别拿下两个版本第一,还真能推进一些开放猜想的特殊情形,做研究的人可以看看这套用 Lean 验证的评测框架。

Event
arXiv cs.AI@Yifan Luo et al. (5)11 sourcesSource