Also Worth Noting - 2026-08-01

Also Worth Noting

02 [Open-source] Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering Recursive self-improvement research has lacked a concrete, executable testbed , OpenMLE is the first full-stack open system that changes that. It packages verifiable task environments with execution feedback (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo) into one coherent stack. On top of it, Frontis-MA1 (35B) is post-trained as a meta-evolution agent organized around four atomic program-evolution operators. Teams studying RSI now have a shared, reproducible foundation instead of arguing about capability in the abstract. link

03 [Training] CoRT: Counterfactual Replay for Token-Level Rubric-Guided Policy Optimization GRPO's scalar reward broadcast actively discards the structured information rubrics carry , every token in a response gets the same advantage signal regardless of which criteria it actually satisfies. CoRT fixes this by replaying counterfactual token sequences to derive token-level credit weights that reflect where each rubric criterion is grounded in the output. The fix requires no changes to the reward model itself. Teams running rubric-conditioned GRPO pipelines can adopt token-level credit weighting without rebuilding their reward infrastructure. link

04 [Eval] DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space A fixed rubric becomes the binding constraint once a solver improves past the dimensions it measures , apparent progress stalls not because the model stopped improving but because the evaluation criteria stopped keeping pace. DecoEvo co-evolves the solver and its rubric generator in text space, decoupling their scores so rubric updates are not gamed by the current solver's performance. Neither model weights nor the reward model are touched; all optimization happens on inspectable natural-language artifacts. Teams running open-ended text-space optimization should treat static rubrics as a ceiling worth monitoring. link

05 [Inference] ReToken: One Token to Improve Vision-Language Models for Visual Retrieval A single learnable embedding acting as a retrieval target lifts Qwen3VL-8B by 13.4 points on Visual Haystacks and InternVL3.5 by 12.4 points, without processing all visual tokens at once. ReToken selects a sparse set of query-relevant tokens from a pre-filled visual KV cache, sidestepping the GPU memory wall that dense visual attention hits as distractor count grows. Training uses only a small image-QA dataset. For multimodal inference pipelines bottlenecked on long visual context, one extra learned token replaces the need for architectural overhauls. link

06 [Agent] Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale Bulk synthetic environment generation moves the bottleneck from quantity to quality , Echoverse finds that returns come from three specific properties: behavioral depth per environment, targeting the interactions an agent actually fails on, and environments that improve alongside the model. The system composes these properties into stateful, login-gated-style environments that can be broken and reset at training scale. Environment count, the common scaling lever, turns out to be the wrong variable. Teams building computer-use agents should audit environment depth before generating more of them. link