FocalRL-4B

FocalRL-4B is a deep research agent based on Qwen3.5-4B, trained with FocalRL for long-horizon web research. It combines full-trajectory outcome rewards with rubric-scored local action repair to improve credit assignment across extended search and browsing trajectories.

This repository provides FocalRL-4B model weights in Hugging Face Transformers format.

Comparison of baseline, GRPO, and FocalRL across deep research and software engineering benchmarks

FocalRL compared with an SFT baseline and GRPO on Qwen3.5-9B Deep Research, and with the base model and GRPO on Qwen3.5-4B SWE-bench Verified. BrowseComp in this figure uses the first 200 tasks.

BrowseComp accuracy versus total model parameters, including FocalRL-4B at 41.3% and FocalRL-9B at 57.9%

Accuracy on the full BrowseComp benchmark (1,266 tasks) versus total model parameters, including FocalRL-4B (41.3%) and FocalRL-9B (57.9%).

Focal Credit Assignment

FocalRL trains long-horizon agents through local action repair. It collects full trajectories, identifies consequential errors in failed attempts, and branches from the corresponding pre-error states. Short continuations are scored against task-specific repair rubrics, and the policy is optimized jointly with full-trajectory outcome rewards and local repair rewards.

Each local continuation spans at most five assistant turns. Repair rewards are 0 for repeating the identified error, 0.5 for avoiding it, and 1 for avoiding it and making the specified correction.

Performance

Accuracy (%) across five Deep Research benchmarks, compared with 4B search agents:

Model Parameters BrowseComp BrowseComp-ZH xbench-2505 xbench-2510 GAIA-text
DR-Venus 4B 29.1 37.7 74.7 40.7 64.4
AgentCPM-Explore 4B 24.1 29.1 70.0 34.0 63.9
ABSeeker 4B 37.3 39.1 77.0 46.0 81.6
FocalRL-4B 4B 41.3 46.1 78.0 48.0 82.5

Bold and underlined scores indicate the best and second-best results in the 4B category.

Code and Usage

Use this checkpoint as the policy model with the FocalRL Deep Research agent. The agent uses web search and page fetching to gather evidence, with browser summarization for fetched content.

License

Apache-2.0.

Downloads last month
5
Safetensors
Model size
5B params
Tensor type
BF16
·
Video Preview
loading

Model tree for XingYing-stack/FocalRL-4B

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(1,486)
this model