SJTU's ABSeeker: 4B Search Agent Matches 30B Models via Step-Level Credit Assignment
SJTU · hf · 2026-08-06
Shanghai Jiao Tong University introduced Answer-Backtracked Credit Assignment (ABC), a framework designed to address the issue of uniformly treating all actions within a trajectory when training long-horizon search agents.
- Core Mechanism: It recovers intermediate clues by backtracking from the final answer and scores each search step. This converts sparse trajectory-level outcomes into dense step-level rewards, rewarding useful actions while suppressing erroneous or redundant ones.
- Training Application: The authors developed ABC-SFT (which reweights loss) and ABC-GRPO (which uses step-level scores as rewards).
- Performance: Trained on Qwen3.5-4B with only 8.5k examples, ABSeeker achieves outstanding results on BrowseComp. With context management, scores reach 55.3% (English) and 52.9% (Chinese), significantly outperforming same-scale (4B) agents and matching the performance of much larger (30B) models.
More from coding & agent
- Ending AI Slop: Engineering Fuzzy Tasks into Clear Ground Truths — _ScottCondron · 2026-08-06
- AWS Bedrock Launches Native Web Search for OpenAI Models — DigitalColmer · 2026-08-06
- Dev Uses Claude Opus to Write C Code Driving ESP32 S3 Hardware — petewoodbridge · 2026-08-06
- Beyond Generated Video: Using Agents to Automate Product Demo Shoots — socialwithaayan · 2026-08-06
- Paper Proposes Token-Native Storage Architecture for AI Agents — bclavie · 2026-08-06
- What STT do you use for production voice agents? Devs say LLM often blamed, but issues lie in voice pipeline — potqtocake · 2026-08-06