Weaker Teacher DeepSeek-V3.2 Beats Claude Opus 4.6 at Training Smarter AI Agents
A new paper reveals that weaker teacher models can produce better-trained agents, and that how a trajectory is structured matters more than who generated it.
Takeaways
- Pedagogical paradox: Students fine-tuned on DeepSeek-V3.2 trajectories outperform those trained on higher-scoring Claude Opus 4.6 data on Terminal-Bench 2.0.
- Environment-Grounded Supervision (EGS): Trajectories that expose inspect-act-verify loops teach generalizable reasoning, not just action memorization.
- Terminal-Lego pipeline: Transforms real-world issues into Dockerized, verifiable agentic tasks with automated EGS-based trajectory filtering.
- 30x data efficiency: Qwen3-32B reaches 24.3% on Terminal-Bench 2.0 with only 15.3k trajectories, matching results from pipelines using 500k+ samples.
- Harness Engineering framing: The paper argues that designing the environment-agent interaction structure is the primary lever for reproducible agent training.
- No public release yet: Code and model checkpoints are not confirmed available in the current preprint; the framework is reproducible using open tools like Terminal-Bench 2.0.
There is a deeply held assumption in agent post-training: if you want to teach a smaller model to be a better agent, you should use the strongest available model to generate the training data. The stronger the teacher, the better the student. A new paper called "What Makes Interaction Trajectories Effective for Training Terminal Agents?" challenges that assumption head-on, and the results are surprising enough to reframe how the field thinks about data curation for agentic fine-tuning.
The Problem With "Stronger Is Better"
Training terminal agents -- models that autonomously execute multi-step tasks inside a command-line environment -- requires high-quality interaction trajectories. These are sequences of actions, observations, and outcomes that a model generates while solving a task in a live environment. The standard playbook is to use the most capable frontier model you can afford to generate these trajectories, then fine-tune a smaller student model on them.
But this approach conflates two separate things: the teacher's raw benchmark performance and the actual pedagogical value of its trajectories. Prior approaches have largely sidestepped the environment bottleneck, with some distilling behavior from stronger proprietary models through supervised fine-tuning, inheriting the ceiling of the teacher and requiring expensive API access. Nobody had cleanly separated the question of who generates the data from how that data is structured.
Terminal-Lego and the Benchmark
The researchers built their experiments around Terminal-Bench 2.0, the current standard for evaluating terminal agents. Terminal-Bench 2.0 is a carefully curated hard benchmark composed of 89 tasks in computer terminal environments inspired by problems from real workflows, where each task features a unique environment, human-written solution, and comprehensive tests for verification. Tasks span diverse domains including protein assembly for synthesis, debugging asynchronous code, and resolving security vulnerabilities.
To generate training data at scale, the team built Terminal-Lego, a pipeline that transforms multi-domain real-world issues into environment-verified agentic tasks. The name echoes SWE-Lego, a similar data-curation framework for software engineering agents. The key design goal was scalability with verifiability: every generated task is executed inside a Docker container, and the agent's solution is checked by automated tests, ensuring the trajectory actually reflects a correct problem-solving process.