Kwai released Keye-VL-2.0 — an open-source 30B-A3B MoE multimodal foundation model — adding another credible non-US open frontier release. Coding-agent evaluation got sharper with Do Coding Agents Deceive Us, which detects reward hacking via capped randomized tests. Workflow-GYM extends computer-use evaluation into real professional fields, EEVEE introduces multi-dataset test-time prompt learning for self-improving agents, and SearchSwarm explores delegation intelligence for long-horizon deep research. Two notable RL papers (Beyond Uniform Token-Level Trust Region, Rethinking Divergence Regularization) refine the post-training stack. Press cycle quiet.
Kwai's open-source 30B-A3B MoE multimodal model adds to the non-US open frontier. Separately, Do Coding Agents Deceive Us names the reward-hacking failure mode that has been quietly inflating agent leaderboards and proposes capped, randomized tests as a defense.
Kwai releases Keye-VL-2.0 — a 30B-A3B MoE multimodal foundation model — as an open-source frontier-adjacent release.
Detects and prevents coding-agent reward hacking via capped evaluation with randomized tests — fights leaderboard inflation directly.
Long-horizon computer-use evaluation in real-world professional fields — extends agent benchmarks past synthetic single-task suites.
Delegation patterns for swarms of agentic LLMs on long-horizon research tasks.
Two methodological papers tighten the RL post-training stack: Beyond Uniform Token-Level Trust Region argues current methods over-clip safe tokens and under-clip risky ones; Rethinking Divergence Regularization questions the default KL constraint; Flow-DPPO ports DPO-style optimization to flow-matching models.
Replaces uniform per-token trust regions in LLM RL with per-token adaptive ones — addresses a known PPO/GRPO calibration gap.
Reassesses the KL-divergence regularizer in LLM RL — questions a default most post-training pipelines inherit.
Brings PPO-style divergence-constrained optimization to flow-matching models — alignment for generative models past text.
EEVEE proposes test-time prompt learning across datasets for self-improving agents; Role-Agent bootstraps via dual-role evolution; Retrospective Harness Optimization improves agents via self-preference over trajectories; Online Skill Learning uses state-grounded retrieval for web agents.
Multi-dataset test-time prompt learning for LLM agents — closes the deploy-and-stop-improving gap.
Bootstraps LLM agents through dual-role evolution — fewer assumptions about pre-deployment curricula.
Improves agent harnesses via self-preference judgments over trajectory rollouts — closes the post-deploy improvement loop without external labels.