JET
A fine-tuning method for repository-level question-answering agents. It truncates trajectories at the point where information is sufficient and inserts rationales with high loss weights, training the agent to stop calling tools once enough information has been gathered and thereby mitigating over-exploration.
Timeline (last 90 days)
The paper "JET: Just-Enough Exploration Tuning for Efficient Tool-Calling Agents in Repository-Level Question Answering" proposes a fine-tuning method named JET, aimed at training agents to stop exploring and answer the question when information is sufficient.
The JET method truncates training trajectories: it stops at the step where information becomes sufficient and inserts a rationale explaining why to stop.
During fine-tuning, JET assigns a higher loss weight to the inserted rationale to strengthen the agent's learning of the stopping condition.
Through truncation and rationale augmentation, JET essentially introduces an emphasis on key decisions within the SFT framework, thereby alleviating SFT's shortcomings.
JET requires no environment interaction and achieves efficiency optimization solely through supervisory signals.
The paper conducted experiments on two repository-level question answering benchmarks using two open-source 8-billion-parameter models.
On self-generated trajectories, JET reduces the number of tool calls by 11% to 33% while largely maintaining task scores.
In comparison with SFT, JET achieves 18% to 28% fewer tool calls than SFT on expert trajectories while retaining most of SFT's score improvement.
JET's effectiveness in reducing tool calls is consistent across different models and benchmarks, demonstrating generalization ability.
JET's method may be generalizable to other scenarios requiring multi-step tool calls, such as web navigation and API calling.
All Facts (10)
The paper "JET: Just-Enough Exploration Tuning for Efficient Tool-Calling Agents in Repository-Level Question Answering" proposes a fine-tuning method named JET, aimed at training agents to stop exploring and answer the question when information is sufficient.
50%UnverifiedThe JET method truncates training trajectories: it stops at the step where information becomes sufficient and inserts a rationale explaining why to stop.
50%UnverifiedDuring fine-tuning, JET assigns a higher loss weight to the inserted rationale to strengthen the agent's learning of the stopping condition.
50%UnverifiedThrough truncation and rationale augmentation, JET essentially introduces an emphasis on key decisions within the SFT framework, thereby alleviating SFT's shortcomings.
50%UnverifiedJET requires no environment interaction and achieves efficiency optimization solely through supervisory signals.
50%UnverifiedThe paper conducted experiments on two repository-level question answering benchmarks using two open-source 8-billion-parameter models.
50%UnverifiedOn self-generated trajectories, JET reduces the number of tool calls by 11% to 33% while largely maintaining task scores.
50%UnverifiedIn comparison with SFT, JET achieves 18% to 28% fewer tool calls than SFT on expert trajectories while retaining most of SFT's score improvement.
50%UnverifiedJET's effectiveness in reducing tool calls is consistent across different models and benchmarks, demonstrating generalization ability.
50%UnverifiedJET's method may be generalizable to other scenarios requiring multi-step tool calls, such as web navigation and API calling.
50%