Let the agent solve the task first, then collect its successful execution traces and feed them back to the model — that is a mainstream pipeline for training tool-use agents today. A paper submitted to arXiv on October 24 (2610.33568) pushes back against half of that recipe in its opening line: successful tool-use trajectories are not necessarily useful supervision. Listed under the ant-international organization and submitted to Hugging Face by first author Yihao Hu on October 27, it ranks as the #1 paper of the day with 330 upvotes.
Two hidden failures inside successful trajectories
The paper splits "successful but useless" into two failure modes. First, insufficient compositional coverage: the observation-action dependencies that matter in a multi-step task are simply not represented, so the model memorizes an action sequence that happened to work. Second, ungrounded decisions: at decision time, the interaction history visible to the learner does not justify the action — the model is effectively asked to memorize an answer it cannot derive. The first erases structure; the second rewards arbitrariness.
Don't collect trajectories — construct interactions
Trajectory2Gym replaces trajectory collection with interaction construction. It locates under-supported compositional structures, builds executable interactions around them, and derives supervision only from grounded decisions in the realized executions. Two mechanisms support this: structural-support selection allocates the resulting training experience, while parent-normalized supervised fine-tuning treats one semantic interaction — not one rendered row — as the unit of training weight. A counter-intuitive detail: the official repository states plainly that the experiments use ordinary assistant-token cross-entropy, not an RL objective. The gains come from reorganized supervision, not from a fancier training algorithm.
Five Qwen3 scales, gains at every one
Across five Qwen3 scales — 1.7B, 4B, 8B, 14B, and 30B-A3B — the equal-weight mean of tau-bench, tau2-bench, and BFCL rises by 16.9 to 29.8 percentage points, improving at every tested scale. The gains are strongest in stateful, multi-turn interaction, where later decisions depend on earlier observations and feedback — exactly what naive trajectory distillation tends to lose. In the Hugging Face discussion, a reader walks through the paper's ablation tables: parent normalization alone adds 8.04 points to the Qwen3-8B mean, with tau2-bench going from 52.50 to 74.17; support depth 2 reaches a 72.28 mean, beating depths 3 and 4. The 30B-A3B study uses an expanded corpus.
A genuinely complete open-source release
The code ships under Apache-2.0, with a method-to-code map in the README. The dataset mis2/Trajectory2gym contains 34,772 semantic interactions in seven compressed JSONL shards (about 125 MB), with checksums, provenance, and license notices. All five checkpoints are public — 1.7B (float32, 8.1 GB), 4B (17.6 GB), 8B (32.8 GB), 14B (bfloat16, 29.5 GB), and 30B-A3B (61.1 GB) — unquantized and unretrained for release; the README candidly notes the reported scores are inherited experiment results, not a fresh rerun. Paper: arXiv:2610.33568; code: GitHub - yihaohu0118/Trajectory2Gym.
So what
What this paper really hits is the link the data-flywheel narrative skips: the assumption that if agents run long enough and successful trajectories pile up high enough, capability will grow on its own. The five-scale controlled comparison says otherwise — the ceiling is set not by the number of trajectories but by whether their dependency structures are represented and whether each decision is derivable from visible history. For teams building agent distillation pipelines, this is an immediately actionable reminder: before scaling up volume, audit how many actions in your trajectories are ungrounded. Constructing learnable interactions may sit closer to the essence of scaling than collecting successful trajectories.