Don't Mask the Environment: Observation Supervision Changes How Agents Explore Under RL

Papers

arxiv:2609.20715

Published on Sep 17

· Submitted by

Juzheng Zhang on Sep 18

· Amazon

Upvote

40

Authors:

,

,

,

,

Abstract

Agent trajectories record what an agent does and what happens next. Yet standard supervised fine-tuning (SFT) applies loss only to agent-authored action tokens, using environment observations as context but not as prediction targets. We ask whether this convention provides the best initialization for subsequent reinforcement learning. We introduce ActObs, which also supervises the observation tokens already present in each trajectory. Although deployed agents never generate observations, learning to predict them encourages the policy to model action consequences without adding data, parameters, sequence tokens, or forward passes. The methods perform similarly after SFT but diverge after GRPO. On Qwen3-4B, GRPO from ActObs achieves higher pass@k at every evaluated sampling budget than its action-only counterpart on Terminal-Bench 2.0. On Qwen3-8B, it trades some pass@1 reliability for higher pass@k (+3.4 pp at pass@16) and solves more distinct tasks. The advantage extends to cross-domain code editing on aider-polyglot (+4.2 pp at pass@1 at 4B), whose tasks are unseen during SFT and RL. ActObs retains more entropy during RL while requiring less policy movement, leaving the final policy closer to its SFT initialization. Our analysis traces this difference to SFT: action and observation gradients rapidly become orthogonal, while action-only training leaves a large residual observation gradient and degrades environment prediction below the base model. Joint supervision prevents this one-sided specialization, preserving consequence prediction and preparing the policy for downstream exploration.

View arXiv page View PDF Add to collection

Community

juzhengz

Paper submitter 5 days ago

Supervising both actions and observations in agent trajectories produces a stronger initialization for downstream reinforcement learning, improving exploration and cross-domain generalization.

librarian-bot

4 days ago

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on HF Mirror checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

O96a

2 days ago

Does the observation-supervision signal survive scale, or is this a small-model effect? The exploration gains you show are real, but I'd want to see the added loss on a 7B+ model — SFT objectives get diluted fast when you're stacking auxiliary losses. Also curious about training cost: does the extra supervision slow convergence enough to eat the exploration benefit? For anyone deploying this, the question is whether the exploration improvement is worth the extra training complexity, or if it's just a nice trick for small runs.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment

Upvote

40

Get this paper in your agent:

hf papers read 2609.20715

Don't have the latest CLI?

curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2609.20715 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2609.20715 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2609.20715 in a Space README.md to link it from this page.

Collections including this paper 1

← 返回资讯列表