Mudith Jayasekara(@mudithj):Excited about this RLM work as it shows: (i) Clear benefits of harness design that solves for the particular constraints of the task (every model showed significant uplift with an RLM harness). RLMs here solve for the 80M tokens sometimes required to be traversed per task (ii) RL works and the recipe can translate (with some careful monitoring). We RL'd smaller models like Qwen 122ba10b and larger ones like GLM 5.3 (iii) RL shapes useful behaviours even when not directly rewarded. In this case, dataroom coverage went from 62% to 96% which intuitively is the right behaviour being learned. We're just getting started etc. etc. @nikogrupen @gabepereyra @ItsJulioPereyra @winstonweinberg and broader research team are cranking.

2026.09.08 19:24

Excited about this RLM work as it shows: (i) Clear benefits of harness design that solves for the particular constraints of the task (every model showed significant uplift with an RLM harness). RLMs here solve for the 80M tokens sometimes required to be traversed per task (ii) RL works and the recipe can translate (with some careful monitoring). We RL'd smaller models like Qwen 122ba10b and larger ones like GLM 5.3 (iii) RL shapes useful behaviours even when not directly rewarded. In this case, dataroom coverage went from 62% to 96% which intuitively is the right behaviour being learned. We're just getting started etc. etc. @nikogrupen @gabepereyra @ItsJulioPereyra @winstonweinberg and broader research team are cranking.

Harvey@harvey

2026.09.08 17:11

We partnered with @baseten to post-train recursive language model (RLM) agents for M&A diligence. We found that model-harness co-optimization meaningfully improves agent performance in long-horizon environments, bringing us closer to agents that can complete M&A diligence end-to-end. We started by building an RLM harness that lets a root agent search a data room, delegate document review to sub-agents, and orchestrate the sub-agents' work to produce a final diligence memo. By distributing document review across sub-agents, the RLM harness can work through thousands of documents in data rooms containing up to 80M tokens. On LAB Diligence, our synthetic M&A diligence benchmark, the RLM harness raised average rubric criteria pass rate across models from 23% to 62%. Despite the root agent processing just 1-4% of the tokens in diligence (with sub-agents processing 96-99%), the choice of root agent had a much larger effect on performance than choice of sub-agents. This led us to focus on the root agent, training two models separately to improve diligence performance. First, we trained a GLM-5.2 root model in the RLM harness using self-distillation SFT. We selected GLM-5.2 diligence runs with high data room coverage and fine-tuned the model on these trajectories. This increased benchmark performance from 46% to 60%. Second, we trained a base Qwen3.5-122B-A10B using GRPO within the RLM harness, keeping the sub-agent models fixed. This raised rubric criteria pass rate from 30% to 63% on 50 held-out data rooms. In both cases we saw major qualitative shifts in agent behavior. The trained root agent learned to delegate more thoroughly and write its memo incrementally as sub-agent findings came back. Data room coverage rose from 62% to 96%. And the root agent learned to scale the size of its agent team proportionally to the size of the data room. We are now scaling RL training with GLM-5.3, a frontier open-weight model, as the root agent. A small amount of training gives an increase in pass rate from 51% to 59%. @nikogrupen walks through these experiments with more on harness design, training setup and results, and what we're exploring next:

커뮤니티로 전달