Post
105
Trajectory2Gym raises five Qwen3 models (1.7B to 30B-A3B) by 16.9 to 29.8 points on the equal-weight mean of Tau1, Tau2 and BFCL. Plain SFT, no RL.
A successful tool-use trajectory is not automatically good supervision. In a pool of 34,772 certified candidates, 40.9% contain at most two supervised decisions and only 10.0% have a certified decision after an authentic error.
So the method builds interactions around under-represented multi-step dependencies. An action becomes a target only if it executed validly and is supported by what the agent could see at that step (g_t = e_t v_t). Failed calls stay in context with no imitation loss; the grounded recovery that follows becomes the target.
📊 The numbers
• Post-trained 4B: 69.17 on Tau2, vs 28.33 for the untrained 30B-A3B
• Post-trained 30B-A3B: 84.17 on Tau2, vs reported 84.76 for GPT-5.2 (high) and 85.31 for Claude-4.5-Opus (high)
• EnvFactory-8B nearly ties the 8B model on BFCL (77.55 vs 78.62) but trails it by 47.5 points on Tau2
• Biggest ablation gain (+8.04 mean): parent normalization, weighting each interaction equally instead of each rendered row (Tau2: 52.50 to 74.17)
⚠️ Caveats
• Ablations are single, non-additive runs on Qwen3-8B
• BFCL Live Simple drops 6 to 13 points at four of five scales
• The code alone does not reproduce the reported numbers; the public pool is not the per-scale training sets
🛠️ Try it
• The Apache-2.0 code, five checkpoints and the 34,772-interaction pool are public
Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents (2610.33568)
https://github.com/yihaohu0118/Trajectory2Gym
In your agent SFT data, how many imitated actions could the agent justify from what it saw?
• Medium: https://medium.com/p/ce1123e4e18b
• Substack: https://sajjadpakdamansavoji.substack.com/p/trajectory2gym-a-4b-qwen3-beats-the
• X: https://x.com/saavoji/status/2115470777499656285
• Website: https://sajjadpsavoji.github.io/blog#trajectory2gym
A successful tool-use trajectory is not automatically good supervision. In a pool of 34,772 certified candidates, 40.9% contain at most two supervised decisions and only 10.0% have a certified decision after an authentic error.
So the method builds interactions around under-represented multi-step dependencies. An action becomes a target only if it executed validly and is supported by what the agent could see at that step (g_t = e_t v_t). Failed calls stay in context with no imitation loss; the grounded recovery that follows becomes the target.
📊 The numbers
• Post-trained 4B: 69.17 on Tau2, vs 28.33 for the untrained 30B-A3B
• Post-trained 30B-A3B: 84.17 on Tau2, vs reported 84.76 for GPT-5.2 (high) and 85.31 for Claude-4.5-Opus (high)
• EnvFactory-8B nearly ties the 8B model on BFCL (77.55 vs 78.62) but trails it by 47.5 points on Tau2
• Biggest ablation gain (+8.04 mean): parent normalization, weighting each interaction equally instead of each rendered row (Tau2: 52.50 to 74.17)
⚠️ Caveats
• Ablations are single, non-additive runs on Qwen3-8B
• BFCL Live Simple drops 6 to 13 points at four of five scales
• The code alone does not reproduce the reported numbers; the public pool is not the per-scale training sets
🛠️ Try it
• The Apache-2.0 code, five checkpoints and the 34,772-interaction pool are public
Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents (2610.33568)
https://github.com/yihaohu0118/Trajectory2Gym
In your agent SFT data, how many imitated actions could the agent justify from what it saw?
• Medium: https://medium.com/p/ce1123e4e18b
• Substack: https://sajjadpakdamansavoji.substack.com/p/trajectory2gym-a-4b-qwen3-beats-the
• X: https://x.com/saavoji/status/2115470777499656285
• Website: https://sajjadpsavoji.github.io/blog#trajectory2gym
