To Align Long, Plan Short: Simulating Diverse Profile-Aligned Students by Realizing Short-Term Plans in Long-Horizon Learning Environments

Synthesis: Radmehr, Zhao, Güreş, Singla and Käser built ALPS, a hierarchical simulated student that plans profile-aligned task segments and then generates individual actions, rather than choosing each action independently from a behavior profile. Tested in an inquiry-based physics game and a pharmacy simulation, under both data-sufficient supervision and a data-scarce setting where only three real trajectories per profile were available, ALPS beat turn-level generation on behavioral density and on judge-rated profile consistency while coverage stayed lower. The authors' claim is narrow and useful: hierarchical planning holds a LLM simulator to an assigned profile over long horizons, which turn-level prompting does not, even when real data are scarce.

Key Findings

1. Hierarchical planning beat turn-level generation in both data settings. ALPS consistently outperformed turn-level generation on density and judge pass rates, with comparable judge agreement under scarce supervision.
2. Coverage remained the weak metric. The paper reports coverage as lower than turn-level generation's in places, and names it the metric that SFT adaptation most improves.
3. Prompted ALPS roughly tripled judge pass rates for GPT-OSS. Without fine-tuning, ALPS significantly improved density and judge pass rate over turn-level generation in both environments for that backbone.
4. Turn-level simulators collapsed on hard profiles. Turn-level GPT-OSS failed completely on at least one difficult or long-horizon profile in each environment, driving worst-profile scores to zero.
5. Supervised fine-tuning was the adaptation that worked. Every improvement surviving correction came from adding SFT; no comparison involving reward-ranked fine-tuning (RAFT) was significantly positive.
6. The planner and generator do different jobs. Ablating the planner returned prompted-level density and coverage, while ablating the generator reduced coverage — the planner maintains alignment, the generator broadens coverage.
7. The human-aligned judge agreed with blind human profile labels on every annotated trajectory. Agreement was κ = 1.00 on profile labels, and 0.83 and 0.94 on two hard-signature checks.

Why long horizons break turn-level simulation

The problem ALPS addresses is drift. In an open-ended environment, student behavior is a pattern across a long trajectory — exploring, committing to a strategy, abandoning it, returning — and a simulator that re-decides the next action at every turn has no representation of that pattern to hold onto. The authors' diagnostic is direct: as trajectories unfold, ALPS becomes more profile-aligned while turn-level generation stagnates or declines toward chance.

The fix is a hierarchy that mirrors how the profile is described: plan a segment that belongs to the assigned profile, then generate actions that realize it. Fewer decisions about alignment, made at a coarser grain, resist drift better than many decisions made at the finest grain.

Data scarcity is the point

The scarce setting is the one practitioners face. Three real trajectories per profile plus a teacher LLM for the rest of the supervision is a realistic budget for a classroom-scale study, and ALPS holding comparable judge pass rates under it is the result that matters for anyone who cannot collect thousands of logged sessions.

The caveat the paper reports honestly is coverage: a simulator that behaves consistently like the profile it was assigned may still not represent the spread of real students within that profile, which is what coverage measures and what the generator ablation addresses.

Where the model comparison landed

The backbone matters. With Qwen3, the architecture comparisons were not significantly positive, and removing three long examples made ALPS significantly outperform turn-level generation — evidence that added context overloads an 8B model and degrades segment planning. GPT-OSS showed the opposite pattern, with ALPS helping substantially.

That asymmetry is a deployment constraint rather than a footnote: a small model may not be able to run the hierarchy at all, and a course or platform choosing a backbone should test the configuration rather than the architecture in the abstract.

What this means for practice

• Simulate trajectories, not turns, when testing a platform. A simulator that re-decides every action drifts off its profile, which makes results from long sessions unreliable.
• Budget for a cheap supervision mix. Three real trajectories per profile plus a teacher model held up in this study, which is within reach of a single course rather than a data team.
• Check coverage, not just consistency. A faithful-looking simulator that under-represents the students an intervention targets will make the intervention look better than it is.
• Validate the backbone on your environment. Architecture gains disappeared or reversed with a smaller model and long context, so configuration testing is part of the work.

Limitations

• Two environments, both simulations. An inquiry-based physics game and a pharmacy simulation; transfer to other platforms is untested.
• Coverage stayed below turn-level generation in places. The authors report it as the metric the approach does not fully recover, and SFT mainly addresses it.
• Three seeds per condition. Results are reported as means over three seeds, which limits how finely differences can be resolved.
• Scarce-setting supervision is partly synthetic. In the data-scarce condition a teacher LLM supplied supervision beyond three real trajectories per profile, so the simulator is partly trained on model-generated behavior.
• Quality is judged by an LLM calibrated against human labels. Profile consistency is scored by a judge agreeing with blind human annotations, not by human rating of every trajectory.

Citation

Radmehr, B., Zhao, J., Güreş, F. B., Singla, A., & Käser, T. (2026). To Align Long, Plan Short: Simulating Diverse Profile-Aligned Students by Realizing Short-Term Plans in Long-Horizon Learning Environments. arXiv:2611.02002 [cs.AI].