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

Bahar Radmehr; Jiaxu Zhao; Fatma Betül Güreş; Adish Singla; Tanja Käser

cs.AINov 1, 2026

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

Authors: Bahar Radmehr, Jiaxu Zhao, Fatma Betül Güreş, Adish Singla, Tanja Käser

Organizations: EPFL, Lausanne, Switzerland · MPI-SWS, Saarbrücken, Germany

Abstract

Simulated students offer a scalable way to test learning platforms and AI tutors before deployment with real students and to augment scarce data for studying learning at scale. Existing LLM-based simulators primarily target short, language-centered interactions. In open-ended learning environments, however, student behavior emerges from patterns across long-horizon interaction trajectories, where turn-level prompting can drift from an assigned behavior profile. We introduce ALPS (Align Long-term by Planning Short-term), a hierarchical simulated student conditioned on observable, environment-specific behavior profiles. Rather than selecting every action independently, ALPS alternates between planning profile-aligned task segments and generating the corresponding actions. We evaluate ALPS in both data-sufficient and data-scarce settings, where supervision comes either from real student trajectories or from a teacher LLM. Across both settings, we compare in-context learning (ICL), supervised fine-tuning (SFT), reward-ranked fine-tuning (RAFT), and their sequential combination (SFT followed by RAFT). Experiments in an inquiry-based physics game and a pharmacy simulation evaluate behavioral density and coverage against held-out real data, and profile consistency using a human-aligned LLM judge. Across both settings, ALPS consistently outperforms turn-level generation, achieving comparable density and judge pass rates under scarce supervision, although coverage remains lower. These results show that hierarchical planning enables faithful, profile-aligned simulation over long interaction horizons, even with limited real student data.

Figures & tables

Figure 1: Overview of the proposed simulation framework. (1) Real student trajectories are clustered into environment-specific behavior profiles. (2) Conditioned on a target profile, ALPS generates long trajectories through a hierarchical planner-generator architecture. (3) When real trajectories are available, ALPS is adapted under both data-sufficient and data-scarce settings using ICL, SFT, and RAFT.

	Model	Architecture	Adaptation	Signal Source	Clipped Density $\uparrow$		Clipped Coverage $\uparrow$		Judge pass rate $\uparrow$
					mean	worst	mean	worst	mean	worst
PharmaSim	GPT-OSS	Turn-level	ICL	3-shot	22.5 ±0.3	0.0 ±0.0	2.2 ±0.7	0.0 ±0.0	39.6 ±0.2	0.0 ±0.0
	GPT-OSS	ALPS	ICL	3-shot	76.0 ±2.3 ∗	47.9 ±4.5	9.7 ±0.3 ∗	3.0 ±0.0	95.6 ±0.8 ∗	86.9 ±1.5
	Qwen3	Turn-level	ICL	3-shot	37.4 ±0.6	0.0 ±0.0	4.2 ±1.3	2.0 ±1.4	46.6 ±0.3	1.0 ±0.8
			RAFT	in-profile	21.1 ±2.5	0.0 ±0.0	13.7 ±6.0	0.0 ±0.0	28.1 ±2.2	0.0 ±0.0
			SFT	$\mathcal{D}^{\text{suff.}}$	37.7 ±4.3	8.6 ±5.2	43.0 ±6.5	6.3 ±2.8	53.2 ±5.6	23.8 ±11.3

Table 1: Exp1: Data-sufficient setting. Simulation quality in PharmaSim and TugLet , reported as percentages (mean ±std over three seeds). Mean averages performance across behavior profiles, and worst reports the lowest-performing profile.

\mathcal{D}^{\text{suff.}}

denotes the real profile-labeled trajectories used for SFT, and in-profile denotes trajectories accepted by the RAFT reward and used for fine-tuning. Bold and underlined values indicate the best and second-best results, respectively, computed separately for GPT-OSS and Qwen3 within each environment. Almost all column-wise best result is achieved by an ALPS configuration.

\ast

indicates that ALPS significantly outperforms its turn-level counterpart (

p<0.05

	Model	Architecture	Adaptation	Signal Source	Clipped Density $\uparrow$		Clipped Coverage $\uparrow$		Judge pass rate $\uparrow$
					mean	worst	mean	worst	mean	worst
PharmaSim	GPT-OSS	Turn-level	ICL	$\mathcal{D}^{\text{scarce}}$	22.9 ±0.2	0.0 ±0.0	1.9 ±0.2	0.0 ±0.0	39.7 ±0.1	0.0 ±0.0
	GPT-OSS	ALPS	ICL	$\mathcal{D}^{\text{scarce}}$	83.7 ±0.7 ∗	18.4 ±3.3	25.9 ±0.2 ∗	3.0 ±0.0	90.5 ±0.7	67.0 ±2.2
	Qwen3	Turn-level	ICL	$\mathcal{D}^{\text{scarce}}$	36.8 ±0.4	0.0 ±0.0	9.5 ±8.7	0.7 ±0.9	46.3 ±0.8	0.7 ±0.5
			RAFT	TR	42.0 ±1.4	0.0 ±0.0	14.0 ±7.2	2.0 ±0.0	66.3 ±2.9	0.7 ±0.5
			SFT	TD	30.0 ±4.8	0.0 ±0.0	2.4 ±0.2	0.7 ±0.9	46.3 ±0.7	0.7 ±0.5

Table 2: Exp2: Data-Scarce Setting. Simulation quality in PharmaSim and TugLet when only three real profile-labeled trajectories per profile are available and a teacher LLM provides additional supervision. Results are reported as percentages (mean ±std over three seeds). Mean averages performance across behavior profiles, and worst reports the lowest-performing profile.

\mathcal{D}^{\text{scarce}}

denotes the three real trajectories per profile, TD teacher-generated demonstrations, and TR the teacher reward. Bold values are computed separately for GPT-OSS and Qwen3 within each environment. Almost all column-wise best values are achieved by an ALPS configuration.

\ast

indicates that ALPS significantly outperforms its turn-level counterpart (

p<0.05

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2: User interface of the two learning environments. Left ( TugLet ): an inquiry-based game where learners alternate between Explore for testing setups and Challenge for predicting outcomes. Right ( PharmaSim ): a simulated pharmacy scenario where learners gather information by discussing with characters and consulting the compendium, documents, shelf, and pharmacist.

	TugLet	PharmaSim
Action types	3	7
Behavior profiles	6	5
Students ( $\mathcal{D}^{\text{suff.}}$ / $\mathcal{D}^{\text{test}}$ )	140 / 109	140 / 94
Trajectories in $\mathcal{D}^{\text{scarce}}$	18	15
Median trajectory length	41	31
90th percentile length	124	39

Appendix

Table 3: Environment-level quantities. The action budget is the generation limit referenced in Section 3.3 ; trajectories that reach it are completed with the deterministic procedure described below. Segment counts are measured on the real trajectories in

\mathcal{D}^{\text{suff.}}

Profile	Exploration volume	Strong-explore ratio	Phase concentration	Session length	Distinguishing behavior
P1: Thorough Learner	substantial	highest	middle (L3–L5)	moderate	most middle-phase exploration
P2: Efficient Learner	modest	second highest	opening (L1), then tapering	short	brief but high-quality exploration
P3: Challenge-First	lowest	near zero	almost entirely opening (L1)	short	advances mainly through Challenge attempts
P4: Early Burst	large initial burst	lowest	opening (L1)	moderate	largest opening-phase burst
P5: Persistent Grinder	modest	low	mostly early and middle, with sparse late exploration	long	continues exploring during the final grind
P6: Stubborn Grinder	lowest per segment	near zero	sparse throughout	long	highest Challenge-question volume

Appendix

Table 4: Qualitative summary of the six TugLet behavior profiles. Properties are stated relative to the other profiles rather than through fixed numeric thresholds. The complete model-facing descriptions and hard and soft signatures will be provided in our code release.

Group	Features
Discussion effort	mother share, baby share, and number of discussion actions
Research volume	numbers of document, shelf-product, and compendium actions
Research relevance	fraction of relevant actions for each research tool (3)
Hint usage	number of pharmacist hints
Topic variety	topic coverage and repetition rate for each character (4)

Appendix

Table 5: The 14 PharmaSim behavioral features organized into five groups.

Profile	Discussion	Hints	Coverage and repetition	Distinguishing behavior
P1: Hint-Heavy Deep Engagement	both characters, somewhat mother-weighted	highest; distributed across segments	broad coverage with limited repetition	substantial engagement over multiple varied segments
P2: Mother-Only	mother only	not defining	narrow, short discussion	never questions the baby and remains within one segment
P3: Hint-Seeking Wide Coverage	both characters, mother first	always at least one	broad coverage with almost no repetition	routinely checks understanding with the pharmacist
P4: Independent Balanced Learner	balanced across both characters	none	broadest coverage with the least repetition	works independently and targets relevant documents
P5: Baby-Only	baby only	not defining	very limited discussion	never questions the mother and ends after one short segment

Appendix

Table 6: Qualitative summary of the five PharmaSim behavior profiles. Our code release will provide the complete model-facing descriptions and hard and soft behavioral signatures.

	TugLet		PharmaSim
Profile	$\mathcal{D}^{\text{suff.}}$	$\mathcal{D}^{\text{test}}$	$\mathcal{D}^{\text{suff.}}$	$\mathcal{D}^{\text{test}}$
P1	20	19	8	5
P2	44	39	17	11
P3	40	25	42	29
P4	8	5	65	43
P5	12	10	8	6

Appendix

Table 7: Students per behavior profile in the development set

\mathcal{D}^{\text{suff.}}

and held-out test set

\mathcal{D}^{\text{test}}

. Each student contributes one trajectory. Every profile contains at least five test students. Following Section 3.4 , density and coverage use

N{=}1

neighbor for the profiles with fewer than ten test students ( TugLet P4, PharmaSim P1 and P5) and

N{=}3

otherwise.

Environment	Quantity	R1	R2	R3
TugLet	Training seed	4096	4097	4098
	Data-split seed	42	43	44
	Generation seed	42	142	242
PharmaSim	Training seed	4096	1234	5678
	Data-split seed	42	43	44
	Generation seed	42	142	242

Appendix

Table 8: Random seeds for the three replicas.

Setting	TugLet	PharmaSim
Base model	Qwen3-8B
LoRA $r$ / $\alpha$ / dropout	32 / 64 / 0
LoRA variant	rsLoRA
Target modules	all attn. & MLP proj.
Optimizer	AdamW
Learning rate	$2\times 10^{-5}$

Appendix

Table 9: Fine-tuning hyperparameters. The TugLet planner used gradient accumulation to preserve an effective batch size of 16.

	Model	Architecture	Adaptation	Signal Source	P1	P2	P3	P4	P5
Clipped Density $\uparrow$
PharmaSim	GPT-OSS	Turn-level	ICL	3-shot	0.0 ±0.0	12.5 ±1.4	0.0 ±0.0	0.0 ±0.0	100.0 ±0.0
			ICL	0-shot numeric	1.7 ±1.2	31.3 ±4.3	0.0 ±0.0	0.0 ±0.0	100.0 ±0.0
			ICL	0-shot	0.0 ±0.0	31.5 ±4.4	0.0 ±0.0	0.0 ±0.0	100.0 ±0.0
		ALPS	ICL	3-shot	100.0 ±0.0	68.6 ±3.5	58.5 ±9.9	52.7 ±11.2	100.0 ±0.0
			ICL	0-shot numeric	100.0 ±0.0	61.0 ±2.0	66.6 ±4.4	100.0 ±0.0	100.0 ±0.0

Appendix

Table 10: Exp1: Data-sufficient setting, per profile ( PharmaSim ). Simulation quality per behavior profile, as percentages (mean ±std over three seeds, 100 trajectories per profile per seed). Rows follow Table 1 , and each row averaged over the five profiles reproduces the corresponding mean column there. Rows in the GPT-OSS block are the prompted GPT-OSS reference; all other rows are Qwen3. The ICL rows add two ablated prompts to the 3-shot condition of the main table: 0-shot numeric (numeric profile descriptions, no examples) and 0-shot (neither). † Turn-level RAFT produced no trajectories for profiles 1 and 4 in any replica, so no adapter was trained for them; those cells ( italic ) are filled per replica with the prompted turn-level baseline from the 0-shot numeric condition.

	Model	Architecture	Adaptation	Signal Source	P1	P2	P3	P4	P5	P6
Clipped Density $\uparrow$
TugLet	GPT-OSS	Turn-level	ICL	3-shot	2.0 ±2.0	33.7 ±9.6	65.2 ±3.3	21.7 ±12.5	83.6 ±11.1	86.2 ±9.3
			ICL	0-shot numeric	6.5 ±1.3	36.0 ±5.2	70.6 ±2.7	28.3 ±7.2	4.8 ±0.3	0.0 ±0.0
			ICL	0-shot	3.4 ±2.0	26.6 ±3.6	68.0 ±1.4	75.8 ±15.0	19.0 ±10.3	3.4 ±3.3
		ALPS	ICL	3-shot	42.4 ±6.5	100.0 ±0.0	100.0 ±0.0	100.0 ±0.0	100.0 ±0.0	61.3 ±4.0
			ICL	0-shot numeric	21.0 ±2.4	83.9 ±5.7	100.0 ±0.0	100.0 ±0.0	93.8 ±8.8	63.6 ±1.9

Appendix

Table 11: Exp1: Data-sufficient setting, per profile ( TugLet ). Simulation quality per behavior profile, as percentages (mean ±std over three seeds, 100 trajectories per profile per seed). Rows follow Table 1 , and each row averaged over the six profiles reproduces the corresponding mean column there. The ICL rows add two ablated prompts to the 3-shot condition of the main table: 0-shot numeric (numeric profile descriptions, no examples) and 0-shot (neither). Turn-level RAFT produced no trajectories for profiles 5 and 6 in any replica, so no adapter was trained for them; those cells ( italic ) are filled per replica with the prompted turn-level baseline from the 0-shot numeric condition.

	Model	Architecture	Adaptation	Signal Source	P1	P2	P3	P4	P5
Clipped Density $\uparrow$
PharmaSim	GPT-OSS	Turn-level	ICL	$\mathcal{D}^{\text{scarce}}$	0.0 ±0.0	14.7 ±1.1	0.0 ±0.0	0.0 ±0.0	100.0 ±0.0
		Turn-level	ICL	0-shot	0.0 ±0.0	31.5 ±4.4	0.0 ±0.0	0.0 ±0.0	100.0 ±0.0
		ALPS	ICL	$\mathcal{D}^{\text{scarce}}$	100.0 ±0.0	100.0 ±0.0	18.4 ±3.3	100.0 ±0.0	100.0 ±0.0
		ALPS	ICL	0-shot	59.2 ±5.1	100.0 ±0.0	16.9 ±6.8	100.0 ±0.0	100.0 ±0.0
	Qwen3	Turn-level	ICL	$\mathcal{D}^{\text{scarce}}$	0.0 ±0.0	84.0 ±1.9	0.0 ±0.0	0.0 ±0.0	100.0 ±0.0

Appendix

Table 12: Exp2: Data-scarce setting, per profile ( PharmaSim ). Simulation quality per behavior profile, as percentages (mean ±std over three seeds, 100 trajectories per profile per seed). The ICL rows report the

\mathcal{D}^{\text{scarce}}

prompt used in the main table and a 0-shot variant without the three example trajectories. † Turn-level RAFT produced no trajectories for profile 1, and turn-level SFT and SFT

\rightarrow

RAFT none for profiles 1, 3 and 4, in any replica; those cells ( italic ) are filled per replica with the matched prompted turn-level baseline.

	Model	Architecture	Adaptation	Signal Source	P1	P2	P3	P4	P5	P6
Clipped Density $\uparrow$
TugLet	GPT-OSS	Turn-level	ICL	$\mathcal{D}^{\text{scarce}}$	2.4 ±0.5	26.6 ±4.9	64.8 ±2.5	70.8 ±12.3	6.4 ±1.5	17.0 ±19.1
		Turn-level	ICL	0-shot	3.4 ±2.0	26.6 ±3.6	68.0 ±1.4	75.8 ±15.0	19.0 ±10.3	3.4 ±3.3
		ALPS	ICL	$\mathcal{D}^{\text{scarce}}$	33.2 ±3.4	38.4 ±5.5	97.3 ±0.3	100.0 ±0.0	84.5 ±3.8	0.0 ±0.0
		ALPS	ICL	0-shot	23.8 ±4.4	34.5 ±5.4	89.4 ±3.0	18.3 ±5.9	90.5 ±4.4	0.0 ±0.0
	Qwen3	Turn-level	ICL	$\mathcal{D}^{\text{scarce}}$	9.3 ±4.1	45.1 ±1.6	15.5 ±0.7	0.0 ±0.0	0.0 ±0.0	8.1 ±3.0

Appendix

Table 13: Exp2: Data-scarce setting, per profile ( TugLet ). Simulation quality per behavior profile, as percentages (mean ±std over three seeds, 100 trajectories per profile per seed). Rows follow Table 2 , and each row averaged over the six profiles reproduces the corresponding mean column there. The ICL rows report the

\mathcal{D}^{\text{scarce}}

prompt used in the main table and a 0-shot variant with neither the three example trajectories nor numeric profile descriptions.

Model	Architecture	Numeric	Examples	Clipped Density $\uparrow$		Clipped Coverage $\uparrow$		Judge pass rate $\uparrow$
				mean	worst	mean	worst	mean	worst
GPT-OSS	Turn-level	–	–	26.3 ±0.9	0.0 ±0.0	4.0 ±1.3	0.0 ±0.0	37.3 ±0.3	0.0 ±0.0
		$\checkmark$	–	26.6 ±0.6	0.0 ±0.0	3.2 ±1.2	0.0 ±0.0	37.7 ±0.3	0.0 ±0.0
		–	$\checkmark$	22.9 ±0.2	0.0 ±0.0	1.9 ±0.2	0.0 ±0.0	39.7 ±0.1	0.0 ±0.0
		$\checkmark$	$\checkmark$	22.5 ±0.3	0.0 ±0.0	2.2 ±0.7	0.0 ±0.0	39.6 ±0.2	0.0 ±0.0
	ALPS	–	–	75.2 ±1.7	16.9 ±6.8	6.7 ±0.9	3.0 ±0.0	76.3 ±1.1	58.0 ±2.9

Appendix

Table 14: Prompt-variant comparison ( PharmaSim ). The two prompt factors are crossed: Numeric indicates numeric profile descriptions (explicit count ranges per behavior), Examples indicates three worked example trajectories in context. Values are percentages (mean ±std over three seeds); mean averages across the five behavior profiles and worst reports the lowest-scoring.

Model	Architecture	Numeric	Examples	Clipped Density $\uparrow$		Clipped Coverage $\uparrow$		Judge pass rate $\uparrow$
				mean	worst	mean	worst	mean	worst
GPT-OSS	Turn-level	–	–	32.7 ±3.6	1.6 ±1.8	11.8 ±6.6	2.6 ±2.0	19.5 ±0.3	0.0 ±0.0
		$\checkmark$	–	24.4 ±1.1	0.0 ±0.0	7.0 ±0.3	1.0 ±1.4	21.3 ±0.1	0.0 ±0.0
		–	$\checkmark$	31.3 ±5.4	2.4 ±0.5	11.7 ±6.7	3.0 ±0.0	20.5 ±1.2	0.0 ±0.0
		$\checkmark$	$\checkmark$	48.7 ±4.0	2.0 ±2.0	48.8 ±7.8	3.0 ±0.0	20.6 ±0.3	0.0 ±0.0
	ALPS	–	–	42.7 ±0.7	0.0 ±0.0	9.1 ±1.2	0.0 ±0.0	30.2 ±1.8	0.0 ±0.0

Appendix

Table 15: Prompt-variant comparison ( TugLet ). The two prompt factors are crossed: Numeric indicates numeric profile descriptions (explicit count ranges per behavior), Examples indicates three worked example trajectories in context. Values are percentages (mean ±std over three seeds); mean averages across the six behavior profiles and worst reports the lowest-scoring.

Figure 3: Profile alignment along the trajectory (GPT-OSS-120B). Target-profile vote share of the in-profile classifier, outlier-masked, against position in the trajectory, for PharmaSim (top, chance

=1/5

) and TugLet (bottom, chance

=1/6

). Real students (black) align increasingly with their profile as the trajectory unfolds. Turn-level generation (gray, dashed) stagnates or decays toward chance, whereas ALPS (teal) tracks the real trend.

	Clipped Density		Clipped Coverage		Judge
Configuration	mean	worst	mean	worst	mean	worst
PharmaSim
ALPS (full)	69.7 $\pm$ 3.1	33.4 $\pm$ 5.8	58.1 $\pm$ 2.8	21.2 $\pm$ 5.0	75.8 $\pm$ 3.8	41.6 $\pm$ 20.4
- generator adapter	67.0 $\pm$ 5.3	23.7 $\pm$ 4.5	49.7 $\pm$ 8.4	17.1 $\pm$ 3.8	71.2 $\pm$ 5.6	40.4 $\pm$ 17.7
- planner adapter	35.3 $\pm$ 0.9	4.5 $\pm$ 2.2	6.8 $\pm$ 0.8	3.0 $\pm$ 0.0	67.0 $\pm$ 2.8	35.9 $\pm$ 2.2
TugLet

Appendix

Table 16: Adapter ablation of ALPS (SFT

\rightarrow

RAFT) in the data-sufficient setting. “- generator adapter” keeps only the planner adapter and “- planner adapter” only the generator adapter; the removed component runs prompted. Percentages, mean

{}_{\pm\text{std}}

over 3 replicas; mean and worst aggregate over profiles; bold marks the best value per column within each environment.

	Contrast	Metric	$n$ (inf.)	$\Delta$	95% CI	$p$	$p_{\text{Holm}}$
PharmaSim Exp1: architecture under prompting (2 contrasts)
	E1.1 GPT-OSS: ALPS vs. Turn	density	15 (12)	$+53.5$	$[+36.9,+69.5]$	$.0004$	$<$ .001
		coverage	15 (14)	$+7.5$	$[+4.2,+10.9]$	$.0008$	.0015
		judge	15 (12)	$+56.0$	$[+31.9,+78.8]$	$.0004$	$<$ .001
	E1.2 Qwen3: ALPS vs. Turn	density	15 (12)	$-6.6$	$[-19.8,+4.7]$	$.2622$	$.2622$
		coverage	15 (10)	$+4.0$	$[+1.0,+6.9]$	$.0677$	$.0677$

Appendix

Table 17: Exp1: data-sufficient test results, PharmaSim . Wilcoxon signed-rank over paired profile

\times

seed cells, exact null. Zero differences are retained in

n

( zero_method=pratt ), since ties arise from ceiling effects where both arms saturate a metric on a profile. Tests are one-sided where the claim is directional and two-sided for the RAFT contrasts and the reference comparison.

n

gives total pairs and, in parentheses, non-tied pairs.

\Delta

is the mean paired difference in percentage points with a percentile-bootstrap 95% CI (

2{\times}10^{4}

resamples). Holm correction is applied within each (contrast block, metric) family. Bold marks

p_{\text{Holm}}<.05

. A ‡ on the

n

column marks at least 40% tied pairs, and a § on

\Delta

marks

|\Delta|<0.05

. Rows with either flag should be read from

\Delta

and the non-tied count rather than from

p

	Contrast	Metric	$n$ (inf.)	$\Delta$	95% CI	$p$	$p_{\text{Holm}}$
TugLet Exp1: architecture under prompting (2 contrasts)
	E1.1 GPT-OSS: ALPS vs. Turn	density	18 (18)	$+35.2$	$[+18.6,+50.9]$	$.0006$	.0013
		coverage	18 (15)	$-13.9$	$[-36.6,+9.0]$	$.4325$	$.8650$
		judge	18 (18)	$+49.1$	$[+35.3,+62.1]$	$<.0001$	$<$ .001
	E1.2 Qwen3: ALPS vs. Turn	density	18 (18)	$-32.7$	$[-51.6,-14.6]$	$.9962$	$.9962$
		coverage	18 (17)	$-55.6$	$[-76.1,-34.5]$	$.9997$	$.9997$

Appendix

Table 18: Exp1: data-sufficient test results, TugLet . Wilcoxon signed-rank over paired profile

\times

seed cells, exact null. Zero differences are retained in

n

n

gives total pairs and, in parentheses, non-tied pairs.

\Delta

is the mean paired difference in percentage points with a percentile-bootstrap 95% CI (

2{\times}10^{4}

resamples). Holm correction is applied within each (contrast block, metric) family. Bold marks

p_{\text{Holm}}<.05

. A ‡ on the

n

column marks at least 40% tied pairs, and a § on

\Delta

marks

|\Delta|<0.05

. Rows with either flag should be read from

\Delta

and the non-tied count rather than from

p

	Contrast	Metric	$n$ (inf.)	$\Delta$	95% CI	$p$	$p_{\text{Holm}}$
PharmaSim Exp2: teacher guidance, Qwen3 (8 contrasts)
	E2.1 ALPS SFT(TD) vs. ALPS ICL	density	15 (15)	$+44.8$	$[+29.2,+61.8]$	$<.0001$	$<$ .001
		coverage	15 (11)	$+20.9$	$[+2.8,+40.3]$	$.0177$	$.1060$
		judge	15 (14)	$+54.1$	$[+37.8,+69.1]$	$<.0001$	$<$ .001
	E2.2 Turn SFT(TD) vs. Turn ICL	density	15 (3) ‡	$-6.7$	$[-17.2,+0.0]$	$<.0001$	$<$ .001
		coverage	15 (5) ‡	$-7.1$	$[-20.3,+0.0]$	$.0043$	.0299

Appendix

Table 19: Exp2: data-scarce test results, PharmaSim . Wilcoxon signed-rank over paired profile

\times

seed cells, exact null. Zero differences are retained in

n

n

gives total pairs and, in parentheses, non-tied pairs.

\Delta

is the mean paired difference in percentage points with a percentile-bootstrap 95% CI (

2{\times}10^{4}

resamples). Holm correction is applied within each (contrast block, metric) family. Bold marks

p_{\text{Holm}}<.05

. A ‡ on the

n

column marks at least 40% tied pairs, and a § on

\Delta

marks

|\Delta|<0.05

. Rows with either flag should be read from

\Delta

and the non-tied count rather than from

p

	Contrast	Metric	$n$ (inf.)	$\Delta$	95% CI	$p$	$p_{\text{Holm}}$
TugLet Exp2: teacher guidance, Qwen3 (8 contrasts)
	E2.1 ALPS SFT(TD) vs. ALPS ICL	density	18 (18)	$+36.1$	$[+20.7,+51.6]$	$.0006$	.0045
		coverage	18 (17)	$+28.4$	$[+11.2,+47.7]$	$.0033$	.0263
		judge	18 (18)	$+33.7$	$[+27.4,+39.6]$	$<.0001$	$<$ .001
	E2.2 Turn SFT(TD) vs. Turn ICL	density	18 (15)	$+8.1$	$[-3.1,+20.2]$	$.8317$	$1.0000$
		coverage	18 (15)	$-3.4$	$[-5.3,-1.6]$	$.0034$	.0263

Appendix

Table 20: Exp2: data-scarce test results, TugLet . Wilcoxon signed-rank over paired profile

\times

seed cells, exact null. Zero differences are retained in

n

n

gives total pairs and, in parentheses, non-tied pairs.

\Delta

is the mean paired difference in percentage points with a percentile-bootstrap 95% CI (

2{\times}10^{4}

resamples). Holm correction is applied within each (contrast block, metric) family. Bold marks

p_{\text{Holm}}<.05

. A ‡ on the

n

column marks at least 40% tied pairs, and a § on

\Delta

marks

|\Delta|<0.05

. Rows with either flag should be read from

\Delta

and the non-tied count rather than from

p

Large Language Model-Based Simulation

See all 39 papers

HACHIMI-II: From Role-Play to Learner Behavior Modeling for Faithful Student Simulation
Oct 14, 2026Yilin Jiang, Fei Tan, Xuanyu Yin +2User Simulation AI in Education
East China Normal University, Shanghai, China · The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · Shanghai Innovation Institute, Shanghai, China
Who Am I? History-Aware Profiles for Student Simulation in Tutoring Dialogues
May 28, 2026Zhangqi Duan, Shuyan Huang, Alexander Scarlatos +3Large Language Model-Based Role-Play Simulation Intelligent Tutoring Systems
University of Massachusetts Amherst · 2Eedi
StudentSim: Training LLM-based Student Simulators
Sep 1, 2026Ke Yang, Chenglong Wang, Michel Galley +4Intelligent Tutoring Systems AI in Education
1Microsoft Research · University of Illinois Urbana-Champaign

Oct 14, 2026J/K move · Enter open · S save

HACHIMI-II: From Role-Play to Learner Behavior Modeling for Faithful Student Simulation

Yilin Jiang, Fei Tan, Xuanyu Yin, Jing Leng+1

East China Normal University, Shanghai, China · The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · Shanghai Innovation Institute, Shanghai, China

User Simulation AI in Education

Open paper AskPDF arXiv

Simulated students let researchers test tutoring systems and teaching strategies before real learners are available, but only if the simulation reproduces where a particular learner succeeds and fails. The common approach asks a large language model to role-play a student, but a role does not model what this learner knows or will do next, so the model's own competence fills the gap. Simulated students perform at a level set by the model, differ little from one another, and seldom make the real learners' errors. We model the learner's behavior and recast student simulation as third-person prediction of the learner's next action. HACHIMI-II builds a student model from the learner's history, infers the learner's state for the current item or tutor turn, and compiles it into a behavior harness that constrains the next action. On three item-response datasets, the Wasserstein-1 distance between simulated and real student ability is 2.4--8.4

\times

smaller than that of the strongest baseline. On two dialogue datasets, HACHIMI-II reproduces real student errors with the same knowledge type, locus, and mechanism 1.8--5.1

\times

more often than the strongest baseline, and ablations trace the gains mainly to the explicit learner state. Predicting the learner makes capable models usable as simulated students.

Abstract

Large Language Model-Based Simulation

Explore similar work

HACHIMI-II: From Role-Play to Learner Behavior Modeling for Faithful Student Simulation

Who Am I? History-Aware Profiles for Student Simulation in Tutoring Dialogues

StudentSim: Training LLM-based Student Simulators