Focused on frontier AI technologies, empowering AI researchers' progress.
The author tests Qwen 3.8 27B by generating a data-analysis tool in five minutes and a 12306-like booking page that looks real but lacks backend integration, while noting Cerebras acceleration hits 1950 tok/s but requires foreign payment.
ByteDance Seed and TokenWave introduce three benchmarks—ASPIRE, S³Gym, and HarnessDev—to evaluate whether AI agents can autonomously form goals, learn from experience, and retain improvements, showing that current agents struggle to translate self-assessment into lasting capability gains.
Bespoke Labs' AutoResearchExam benchmarks nine frontier models on 29 open-ended ML research tasks for 24 hours each, revealing that validation-set gains often fail to transfer to hidden tests — exposing benchmark overfitting where models like GPT-5.6 Sol spend 55.5% of rounds on hyperparameter tuning, while Claude Fable 5.1 ultimately leads on time-weighted hidden-test AUARC despite Astra's early advantage.
Huawei GTS's NetCanvas provides LLM agents with an interactive visual topology as external working memory, enabling them to 'watch' network topology during troubleshooting; on CTBench, pass rates jump from 30.3% to 54.5% (63.6% with priors), dual-firewall/ECMP tasks soar from ~10% to ~90%, with zero regressions and 26–45% token cost reduction.
ByteDance Seed and TokenWave introduce ASPIRE, S³Gym, and HarnessDev — three benchmarks that test whether AI agents can autonomously select learning goals, distill experience into improved decisions, and persistently upgrade their own execution systems without human-provided verification.
This paper introduces DynHD, a hallucination detection method for diffusion large language models that analyzes denoising dynamics across token and temporal dimensions, achieving 84% AUROC on LLaDA-8B and Dream-7B without repeated sampling.
Yann LeCun's ECCV 2026 keynote argues pixel prediction is fundamentally impossible, advocating JEPA which predicts in abstract representation spaces rather than pixels, and contends that language-only models cannot achieve human-level intelligence without world models learned from visual and sensorimotor data.
Alibaba and Zhejiang University's Astar learns from AI systems' own Git history to propose evolution strategies, achieving 54–68% single-shot success rates versus 31% for GPT-5.5 and 32% for human experts, and delivering 23.6% offline HitRate and 4.86% GMV gains in Lazada's ad system through 20 fully automated iterations.
Analysis of 42,123 ICLR papers (2017–2026) reveals that chasing hot research topics yields higher acceptance rates initially, but the advantage vanishes as fields become crowded; for PhD students with short time horizons, following momentum is rational, but growth increasingly goes into rejections rather than acceptances.
Researchers from Beihang University and Nanyang Technological University propose PhyFilter, a lightweight plug-and-play module that uses physical differential structures and real-time feedback to correct learning residuals, allowing robots to generalize across unseen terrains and disturbances without massive datasets.
