Frontier language-model agents excel at complex tasks, but high inference costs make transferring their capabilities to smaller models essential. Traditional knowledge distillation achieves this through parameter updates, requiring access to the student's weights and additional training compute while encoding the transferred knowledge implicitly in the parameters. We introduce D-Skill, a framework that treats the teacher as a black box and distills its behavioral advantage into an interpretable natural-language skill by analyzing paired rollouts and identifying behavioral divergences and agreements. The resulting lessons are consolidated offline, with any compatible skill-update backend, into a skill document that is provided to the student at inference time. D-Skill constructs the skill entirely offline from fixed trajectories, without labels, further environment interaction, or weight updates after rollout collection. Across SpreadsheetBench-Verified, ALFWorld, and chess puzzles spanning multiple teacher-student pairs, D-Skill closes 75%-95% of the teacher-student gap without touching model weights. After a single offline pass over the paired traces, D-Skill outperforms one-epoch SFT baselines and remains competitive with multi-epoch SFT. Under the same model family, the resulting student also incurs 5.7x-8.3x lower inference cost per task than the teacher. Our results demonstrate that natural-language skill distillation offers a performant, interpretable, and efficient alternative to weight-based capability transfer.
Figures & tables
Figure 1: D-Skill closes the gap at a fraction of its cost. D-Skill closes 84% of the teacher-student gap at 8.3× lower inference cost on SpreadsheetBench-Verified. Teacher: GPT-5.6 Terra; student: GPT-5.6 Luna.
Figure 2: Ground-truth-free skill distillation from paired rollout traces. (1) Rollout: Paired teacher and student traces are obtained from fresh rollouts or saved executions. (2) Difference analysis: A Difference Analyst Agent compares the fixed traces without answer keys, labels, or further environment interaction, extracting divergence and agreement items. (3) Skill update: A Consolidation Agent converts the lessons into patches, merges them, and applies the update to SKILL.md .
Table 1: Main results. Each block reports the bare student (floor), the student with the distilled D-Skill skill, and the teacher; %GC is the fraction of the teacher–student gap closed.
Qwen3.6 35B-A3B
GPT-5.6 Luna
Method
Score
%GC
Score
%GC
Student trace only
47.7± 6.7
38.3%
49.8± 8.8
45.4%
Teacher trace only
56.3± 6.1
55.2%
63.5± 5.4
70.7%
D-Skill
71.7± 3.6
85.1%
73.3± 1.0
89.0%
Table 2: Teacher-student contrast improves performances. Results average 3 independent skills, each evaluated twice.
Consolidation
Analyst
Sonnet
Qwen
Sonnet
39.0± 3.5 (88%)
34.3± 3.4 (59%)
Qwen
31.1± 4.5 (39%)
32.6± 3.9 (48%)
Table 3: A good analyst is needed. Chess accuracy of the Qwen student with the skill produced by each analyst-consolidation pair. Results average 3 independent skills, each evaluated twice.
Method
Score
%GC
Student (Qwen3.6-35B-A3B)
24.8± 0.4
—
SkillOpt
29.6± 1.4
29.7%
D-Skill w/ SkillOpt backend
36.9± 5.7
74.9%
D-Skill
39.0± 3.5
87.7%
Table 4: D-Skill is compatible with an alternative skill-update backend. Results average three independently distilled skills, each evaluated twice. The SkillOpt-backed variant retains SkillOpt’s held-out validation gate.
Qwen3.6-35B-A3B
GPT-5.6 Luna
Method
Score
%GC
Score
%GC
Student
28.0± 1.4
—
25.3± 1.1
—
Teacher (DS-V4-Pro + Skill)
79.3± 1.8
100%
79.3± 1.8
100%
Trace2Skill (student trace, student-authored)
33.0± 4.2
9.7%
31.1± 1.6
10.7%
Trace2Skill (teacher trace, teacher-authored)
34.2± 2.8
12.0%
37.0± 1.6
21.7%
Trace2Skill (student trace, teacher-authored)
36.0± 1.1
15.6%
46.2± 5.9
38.6%
Table 5: SpreadsheetBench-Verified: D-Skill vs. Trace2Skill baselines. All numbers are means over six evaluations, comprising two passes for each of three independently distilled skills.
Figure 3: One-pass D-Skill outperforms one-epoch SFT and remains competitive with its multi-epoch performance. Both methods use Qwen3.6-35B-A3B as the student and the same teacher-trajectory corpus on (a) SpreadsheetBench Verified and (b) ALFWorld. The x-axis reports the number of teacher tasks processed, and the shaded bands show SFT performance over epochs 2-5.
Figure 4: Inference cost per task at model list prices on (a) SpreadsheetBench Verified and (b) ALFWorld. Costs are reported in cents episode.
Tasks
Total
Per task
SSB-V
99
3.28$
3.3 ¢
ALFWorld
57
1.51$
2.6 ¢
Table 6: One-time distillation cost GPT-5.6 Terra → Luna, billed at list prices. Per task is per training task.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Tool
Input and output
<hanging></hanging>
No input. Lists every piece that is attacked and has no defender, for both sides.
<safety>m1, …</safety>
Up to five candidate moves. For each: whether it gives check (checkmate and stalemate named explicitly), the number of legal replies, every forcing reply (checks and captures with the captured piece named), attackers vs. defenders of the destination square, and the agent’s pieces left undefended afterwards.
<threats></threats>
No input. Null-move audit: the opponent’s forcing moves if the agent passed, and any immediate mate threats.
<board>move</board>
One move. The board after the agent’s move only, with material, undefended pieces, and check/checkmate status.
<simulate>move, r1, …</simulate>
One candidate and up to five opponent replies, each simulated independently from the position after the candidate: resulting board, material, undefended pieces, and whether the agent is in check.
<line>m1, r1, m2, …</line>
A complete line of two to six plies supplied by the agent (own moves and opponent replies alternating): the final board, material, undefended pieces, and check/mate status.
Appendix
Table 7: Analysis tools available to the chess agent. Every tool is a static computation on a copy of the board; nothing is searched or evaluated.
Skill-based prompting has become a practical mechanism for improving LLM agents, yet existing methods often treat skills as summaries of the agent's own experience or of successful demonstrations. This creates a mismatch for weaker student agents. A failed trajectory may not reveal the missing knowledge or strategy, while a teacher trajectory may be too implicit to internalize. We propose SKILL-KD, a contrastive skill distillation framework that treats skills as an explicit distillation medium between agents of different capabilities. Given a student failure and the teacher trajectory on the same task, SKILL-KD distills their actionable discrepancy into a textual skill patch, evaluates it by re-running the student, and iteratively refines it when the student still fails. To prevent skill drift from repeated local updates, SKILL-KD maintains trace-linked edit histories and performs Drift-Aware Skill Consolidation to decide whether each patch is added, merged, or skipped. Across five agent benchmarks and two student settings, SKILL-KD consistently improves frozen student agents over fixed-model adaptation baselines.
Qiming Shi, Yibo Dou, Jiawen Zhu +5
State Key Lab of CAD&CG, Zhejiang University · School of Software and Microelectronics, Peking University · School of Software Technology, Zhejiang University +1
Reinforcement learning (RL) has been widely used to train LLM agents for multi-turn interactive tasks, but its sample efficiency is severely limited by sparse rewards and long horizons. On-policy self-distillation (OPSD) alleviates this by providing dense token-level supervision from a privileged teacher that has access to ground-truth answers. However, such fixed privileged information cannot capture the diverse valid strategies in agent tasks, and naively combining OPSD with RL often leads to training collapse. To address these limitations, we introduce Skill-SD, a framework that turns the agent's own trajectories into dynamic training-only supervision. Completed trajectories are summarized into compact natural language skills that describe successful behaviors, mistakes, and workflows. These skills serve as dynamic privileged information conditioning only the teacher, while the student always acts under the plain task prompt and learns to internalize the guidance through distillation. To stabilize the training, we derive an importance-weighted reverse-KL loss to provide gradient-correct token-level distillation, and dynamically synchronize the teacher with the improving student. Experimental results on agentic benchmarks demonstrate that Skill-SD substantially outperforms the standard RL baseline, improving both vanilla GRPO (+14.0%/+10.9% on AppWorld/Sokoban) and vanilla OPD (+42.1%/+40.6%). Project page: https://k1xe.github.io/skill-sd/
Hao Wang, Guozhi Wang, Han Xiao +8
Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences · vivo AI Lab · The Chinese University of Hong Kong +2
Agent skills, which consist of reusable strategies that guide agent reasoning and action, have shown strong potential for improving model capability at inference time. However, current skill construction methods treat the problem as one-shot extraction, overlooking a fundamental tension: a skill tailored to the specific task fails to transfer, while the abstracted skill often provides insufficient guidance. We attribute this fragility to the absence of explicit mechanisms for skill specification and generalization. To address this gap, we introduce SkillComposer, a framework that decomposes skill construction into three learnable operations: create, improve, and merge. Trained via systematic rejection sampling recipe, SkillComposer enables language models to self-evolve skills at inference time and supports three deployment modes: offline for building generalized libraries, online for task-specific refinement, and hybrid for combining both. Comprehensive experiments on τ2-Bench, LiveCodeBench v6, and AppWorld show that SkillComposer consistently outperforms baselines. Our SkillComposer-4B improves a 27B executor by up to +4.5 on agent tasks and +3.4 on code tasks, while generalizing across domains and task types unseen during training. Analysis reveals that merge and improve address orthogonal quality dimensions and that skill composition is a transferable meta-ability, providing a practical recipe for skill-augmented inference.
Qi Zhang, Zhaopeng Feng, Xiaonan Shi +8
1Zhejiang University · 2Tongyi Lab · 3National University of Singapore