AITOPAI热报
PapersMulti-sourceSelected

Microsoft Paper Co-Evolves RL Harness and Qwen3.5 Model

DAIR.AI@dair_ai
Post· 3 sources
View the post
Why we picked it

Co-evolving the harness and the model appears to unlock gains that optimizing either alone cannot reach, suggesting the harness itself is a first-class lever in RL training. Teams training coding agents can copy the loop to squeeze more performance from mid-size open models like Qwen3.5-35B-A3B. The distillation trick of using Opus 4.8 as the harness proposer also offers a cheaper route to borrow frontier-model judgment.

A Microsoft-led paper introduces a method that alternates between optimizing the agent harness and training the model with reinforcement learning. On Qwen3.5-35B-A3B, average resolve rate across three coding benchmarks rises from 57.8% to 64.8%. In each round, the policy reads its own failures, proposes harness changes, and is then trained with RL under the harness it selected; tasks that remain hard become the search set for the next harness round. The joint gain exceeds the harness-only and policy-only gains combined, and each optimized component keeps its improvement when paired with the other's original version. Using Claude Opus 4.8 as the harness proposer adds further gains, transferring a frontier model's harness judgment to a smaller policy.

Original · DAIR.AI

Great work from Microsoft and colleagues.

This one is on the hot new topic of co-evolving the harness and the model.

(bookmark it)

They raise Qwen3.5-35B-A3B's average resolve rate on three coding benchmarks from 57.8% to 64.8% by alternating the two.

In each round, the policy reads its own failures and proposes harness changes. Then it is trained with RL under the harness it picked. Tasks that stay hard during training become the search set for the next harness round.

The joint gain is larger than the harness-only gain and the policy-only gain added together. Each optimized part also keeps its improvement when paired with the other's original version.

Using Opus 4.8 as the harness proposer in the same loop adds more on top, which makes this a practical way to transfer a frontier model's harness judgment into a smaller policy.

https://t.co/FlYU8zQBCz

Chat with Paper: https://t.co/PrpXkk1qey

#Qwen3.5#reinforcement learning#coding assistants#agents

Keep reading

Recent stories with the same tag or section
PapersBuildersMulti-sourceSelected04:44
Stanford paper splits agent failures between harness fixes and fine-tuning

The actionable shift is diagnostic: the same failure mode responds to different interventions, so prompt and tooling edits can waste effort on problems only weight training solves. Teams building agents can use the process-versus-content taxonomy as a triage step before committing compute. The Qwen3.5 LoRA numbers give a concrete baseline for how much fine-tuning moves plan quality on a small model.

#Stanford#Qwen3.5#agents#fine-tuning
Event
rohanpaul_ai@rohanpaul_ai6 sourcesSource
PapersBuildersMulti-sourceSelected11:48
NVIDIA paper: tools weaken multimodal model safety refusals

Plain-chat safety evals can overstate how safe an agentic system actually is, since tool use measurably erodes refusal behaviour across every model tested. Teams building multimodal agents should benchmark safety with tools attached and consider re-inserting the original request before generation. The paper is on arXiv with a chat interface at academy.dair.ai for quicker exploration.

#NVIDIA#multimodal#AI safety#agents
Event
elvis@omarsar011 sourcesSource
ModelsBuildersMulti-sourceSelected73°04:51
Overmind fine-tunes Qwen3.5 9B to cut contract hallucinations

The key idea is that observability and training share one dataset: the traces showing where an agent fails become the fine-tuning data and the eval set. Teams running agents on specialised documents can replace expensive frontier APIs with a small model tuned to their own workload. The comparison against GPT5.6 Luna on clause-citation accuracy gives a concrete baseline for anyone weighing cost against reliability in production.

#Overmind#Qwen3.5#GPT5.6 Luna#fine-tuning
Event
rohanpaul_ai@rohanpaul_ai3 sourcesSource
PapersBuildersMulti-sourceSelected06:46
Shanghai AI Laboratory SkillGym trains better agents from verified skill runs

Training on verified execution traces can bake retrieval-dependent prompt skills into model weights, reducing reliance on the agent finding and following instructions at runtime. Teams building coding agents can replicate the recipe by turning their own skill libraries into checkable sandbox tasks. The work comes from a Chinese lab and benchmarks on Claude Code with Qwen, making it a useful reference point against prompt-based skill systems like Claude Skills.

#Shanghai AI Laboratory#Qwen#SkillGym#agents
Event
rohanpaul_ai@rohanpaul_ai3 sourcesSource
ModelsBuildersSelected16:56
Aleph Alpha's Kolibri: 78B open-weight model built for German

A European lab reaching benchmark parity with Chinese and American open models in a bilingual German-English model suggests small-model training capability has spread well beyond the usual hubs. The Apache 2.0 license makes it directly usable for on-premise deployments, which matters for German public-sector and enterprise buyers with data-sovereignty requirements. Teams comparing open models should weigh its German-language edge against its weaker SWE-Bench Verified result.

#Aleph Alpha#Kolibri#open-source models#Qwen
Event
kimmonismus@kimmonismus2 sourcesSource
PapersBuildersMulti-sourceSelected16:15
ProVer method for agent RL credit assignment

The approach addresses a fundamental limitation in current RL methods where all trajectory tokens receive equal credit, making it impossible to identify decisive steps. Developers working on agent systems can implement this method to improve training efficiency, particularly when using smaller judge models, potentially reducing computational costs while maintaining performance gains.

#ProVer#Qwen3.5#agent RL#credit assignment
Event
elvis@omarsar03 sourcesSource