ClueFlow
A training framework for long-horizon search agents. Based on a Source-State-Final attribution chain, it associates state changes with source actions and the final answer, enabling fine-grained credit assignment and improving search accuracy and training efficiency.
Timeline (last 90 days)
ClueFlow constructs a Source-State-Final attribution chain that forms a verifiable causal chain linking source browser actions, state updates, and the final answer, thereby achieving credit assignment.
On the Qwen3.5-35B-A3B backbone model, ClueFlow achieves 60.0% accuracy on the BrowseComp benchmark.
On BrowseComp-Plus, attribution-guided SFT matches the effect of full-turn exact-context SFT using only about 30% of the training tokens.
On BrowseComp-Plus, attribution-guided PPO is 4.1 percentage points higher than PPO using only outcome rewards.
ClueFlow's attribution mechanism determines which decisions should be supervised in the SFT stage, learning only from state changes consistent with the gold standard and their original interaction contexts.
In the Turn-PPO stage, ClueFlow uses the attribution chain to enhance outcome-based turn advantage values, precisely allocating credit along the Source-State-Final chain.
Even if the final answer of the entire trajectory is wrong, ClueFlow's transition-level verification can still recover useful supervision signals from state changes consistent with the gold standard.
ClueFlow's attribution chain construction relies on post-execution semantic verification and source tracing, incurring additional computational overhead, but the paper's token efficiency data shows that this overhead is worthwhile in the training stage.
ClueFlow pushes credit assignment down from the turn granularity to the state-change granularity, enabling the model to learn which type of evidence-acquisition action leads to which type of state evolution.
ClueFlow was evaluated on four search benchmarks, covering both live web and fixed corpus settings.
1 more timeline events
All Facts (11)
ClueFlow constructs a Source-State-Final attribution chain that forms a verifiable causal chain linking source browser actions, state updates, and the final answer, thereby achieving credit assignment.
50%UnverifiedOn the Qwen3.5-35B-A3B backbone model, ClueFlow achieves 60.0% accuracy on the BrowseComp benchmark.
50%UnverifiedOn BrowseComp-Plus, attribution-guided SFT matches the effect of full-turn exact-context SFT using only about 30% of the training tokens.
50%UnverifiedOn BrowseComp-Plus, attribution-guided PPO is 4.1 percentage points higher than PPO using only outcome rewards.
50%UnverifiedClueFlow's attribution mechanism determines which decisions should be supervised in the SFT stage, learning only from state changes consistent with the gold standard and their original interaction contexts.
50%UnverifiedIn the Turn-PPO stage, ClueFlow uses the attribution chain to enhance outcome-based turn advantage values, precisely allocating credit along the Source-State-Final chain.
50%UnverifiedEven if the final answer of the entire trajectory is wrong, ClueFlow's transition-level verification can still recover useful supervision signals from state changes consistent with the gold standard.
50%UnverifiedClueFlow's attribution chain construction relies on post-execution semantic verification and source tracing, incurring additional computational overhead, but the paper's token efficiency data shows that this overhead is worthwhile in the training stage.
50%UnverifiedClueFlow pushes credit assignment down from the turn granularity to the state-change granularity, enabling the model to learn which type of evidence-acquisition action leads to which type of state evolution.
50%UnverifiedClueFlow was evaluated on four search benchmarks, covering both live web and fixed corpus settings.
50%UnverifiedClueFlow brings consistent improvements in the SFT stage, with further improvements in the PPO stage.
50%