AI Agents

SAO tackles a real weakness in asynchronous RL for agents: GRPO's group-wise sampling doesn't fit async training well, hurting stability. Their fix is single-rollout sampling — one rollout per prompt — plus value-model training designs and strict double-sided token-level clipping. They report stable training over 1,000 steps and gains over GRPO variants on SWE-Bench Verified, BeyondAIME, and IMOAnswerBench, with a notable edge in a simulated online-learning setting where the task distribution shifts.

The core finding is damning for a whole genre of eval: across 1,538 tasks from EffiBench, Enamel, EvalPerf, and Mercury, run 30 times each with statistical testing, only 6.11% of the benchmark's "performant" implementations are significantly faster than the canonical solutions. Manual review of 308 non-significant tasks found 209 held real improvements the original tests never exposed. Their fix is a three-agent framework — generate, diagnose, repair — that builds deterministic performance-oriented tests. Rigorous methodology, uncomfortable conclusion.

UI2App measures whether a model can recover an app's behavior from screenshots alone — no text, no behavioral hints — across 327 screenshots in 45 state-coherent sets for runnable multi-route web apps. It scores executability, navigation reachability, visual fidelity, and an interaction inference metric that credits any valid implementation. The headline result: the visual-fidelity leader scores just 7.5 on interaction inference, trailing the interaction leader by 5.2x. Looking right and working right are decoupled capabilities.

ProjAgent adds "procedural similarity" as a retrieval signal for repository-level code generation — finding functions that implement similar logic even when identifiers and domains differ, which lexical, structural, and semantic retrieval miss. It decomposes the target into reasoning steps, retrieves procedurally similar functions per step, and adds a conservative static-analysis repair loop using compiler feedback. On REPOCOD it hits 41.14% Pass@1, beating retrieval-based baselines. The insight — that behavioral similarity is orthogonal to surface similarity — is clean and the result backs it.

This targets visual token explosion in long multimodal dialogue: instead of re-feeding all prior images and text into one context window, it externalizes visual info into an Episodic Visual Memory and selectively reactivates relevant episodes. Three components handle abstraction, cross-turn retrieval, and planning, with RL optimizing the abstraction and retrieval policies over programmatically generated conversations. The 8B agent hits 91.4% retrieval accuracy over 20-turn sessions, beating 32B baselines by +8.2% while cutting per-turn latency from 23.1s to 12.7s.

This runs LLM inference across a peer-to-peer mesh built on iroh, splitting work over machines connected without central coordination. The engineering is concrete and the HN interest (265 points) is real, but the agent angle is indirect — it's an inference-distribution substrate, not an agent runtime. If you're thinking about spreading long-running agent workloads across heterogeneous or edge hardware without a datacenter, the transport layer here is worth understanding.

Mindwalk replays a coding agent's session on a 3D map of your codebase, letting you watch which files it touched, in what order, and where it went sideways. It's a direct answer to the debugging pain of long agent runs where the trace is a wall of tool calls. Spatializing the session over the actual repo structure is a genuinely different lens than reading logs. Early Show HN artifact, so expect rough edges.