acceptodds
Under review as a conference paper at ICLR 2027

D-Skill: Distilling Teacher Knowledge through Agent Skills

10 views

Abstract

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 82%–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 5.7×–8.3× 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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Agents, Skills, Distillation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_HyhCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposed the idea of distilling the knowledge skill from teacher model to student models. The key motivation is to show that prompting the skill extractor to compare the difference between teacher trajectories and student trajectories can have more performance gain compared to distill knowledge between success and failure trajectories. The D-Skill has three stages: 1/ rollout collection for teacher and student agent 2/ conduct difference analysis to extract divergence lesson and agreement lesson 3/ update skill by skill extractor like Trace2Skill. The experiment results show the performance gain.

Critical strengths

1/ The idea of D-Skill is clear, reasonable and easy to follow. 2/ The experiment is comprehensive and clear. For example, Table 2 conducts the ablation study to compare skills extracted from student trajectories, teachher strrajectories and D-Skill. It shows the the D-skill can improve the performance significantly.

Critical weaknesses

1/ Not a weakness, but interesting to have the generalization analysis. As in the line 52 mentioned "...and sufficient training data is needed for the student to generalize". It would be great to include the out of domain evaluation. I think SFT would hurt the generalizability a lot. I am not sure to what degree D-Skill would hurt generalizability.

Questions for authors

See weakness

1 comment
AuthorsOpenReview ↗

Out-of-domain generalization.

We follow Trace2Skill’s OOD setup (Ni et al., 2026), which tests SpreadsheetBench skills on WikiTQ and HiTab. Each table becomes a spreadsheet task, scored by the benchmark’s official evaluator on 300 random test questions per benchmark. We use the SpreadsheetBench documents of all four Table 1 pairs unchanged. Values are mean ± SD over three documents, each with 2 passes.

WikiTQHiTab
Qwen3.6-35B-A3B (no skill)38.0 ± 6.124.2 ± 1.2
+ DeepSeek→Qwen document84.2 ± 2.377.8 ± 1.8
+ Terra→Qwen document76.3 ± 4.371.8 ± 3.3
GPT-5.6 Luna (no skill)81.3 ± 0.967.6 ± 0.8
+ DeepSeek→Luna document82.6 ± 1.676.1 ± 2.2
+ Terra→Luna document80.6 ± 1.574.0 ± 0.9

D-Skill does not hurt generalization. The documents improve Qwen by 38-54 points and Luna by 6-9 points on HiTab, and leave Luna on WikiTQ unchanged.

Reviewer_1nD3OpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes D-Skill, a capability-distillation framework that does not require model parameter updates. Its core idea is to collect execution trajectories from a teacher model and a student model on the same tasks, and then use an analyst model to compare their behavioral differences in information gathering, planning, tool use, and error recovery. These differences are distilled into natural-language rules and consolidated into a skill document that can be loaded at inference time. Once the trajectories have been collected, the entire skill-construction process operates offline, without requiring access to model weights, teacher logits, human annotations, or additional environment feedback.

Critical strengths
  1. The research problem has a degree of novelty and practical relevance. The paper explores natural-language skills as a medium for capability transfer between models of different scales, enabling black-box distillation without access to model parameters or logits. Although the method is largely heuristic, this problem formulation offers a potentially useful alternative to conventional parameter-based distillation.

  2. The experiments provide reasonably strong evidence across multiple tasks and model pairs. D-Skill consistently improves student performance on SpreadsheetBench-Verified, ALFWorld, and chess puzzles across several teacher–student configurations. These gains suggest that the proposed idea is not limited to a single environment and provide meaningful empirical support for its potential cross-domain applicability.

Critical weaknesses
  1. The method is highly heuristic, and the source of its performance gains is not clearly isolated. D-Skill relies heavily on a strong analyst model and carefully designed prompts to compare trajectories, extract natural-language rules, and consolidate them into a skill document. However, these components lack systematic mechanistic analysis and sensitivity evaluation. It is therefore difficult to determine whether the improvements primarily arise from the proposed teacher–student contrast mechanism or from the particular prompts and strong analyst model, limiting the method’s technical novelty and generalizability.

  2. The empirical comparisons with related methods are incomplete and inconsistent across tasks. Although the paper compares D-Skill with Trace2Skill, SkillOpt, and SFT, these comparisons are conducted on different subsets of benchmarks rather than across the full evaluation suite. Several closely related training-free skill- or agent-distillation approaches are discussed but not evaluated empirically. Consequently, the current evidence is insufficient to establish that D-Skill offers a consistent and general advantage over existing methods.

Questions for authors
  1. Please clarify the core technical contribution of D-Skill relative to Trace2Skill, SkillGen, and AgentDistill. Beyond providing paired teacher and student trajectories to an analyst model for comparison, does D-Skill introduce any mechanism or capability that these methods cannot provide? The answer would directly affect my assessment of the paper’s methodological novelty.
  2. Please clarify the extent to which the performance gains can be attributed to teacher–student contrast rather than to prompt design and the use of a strong analyst model. Based on the existing teacher-only/student-only and analyst-model ablations, what evidence isolates these two sources of improvement? Please also indicate whether the complete task-specific analyst prompts and consolidation prompts are included in the submission and where they can be found.
  3. Please explain why the current baseline comparisons are sufficient to support the paper’s overall conclusions. Trace2Skill, SkillOpt, and SFT are evaluated only on different subsets of tasks, while other closely related methods are not empirically compared. Please clarify how this fragmented evaluation supports the claim that D-Skill provides a consistent advantage over existing approaches, and whether a unified comparison is prevented by incompatibilities in the methods or task interfaces.
6 comments
AuthorsOpenReview ↗

Q1. Core contribution relative to Trace2Skill, SkillGen, and AgentDistill

We address a different dimension from Trace2Skill and SkillGen: capability distillation from a stronger teacher to a weaker student. Our contribution has two parts:

  1. Gap closure through textual distillation. The resulting skills transfer teacher capabilities without updating the student’s weights. Table 1 shows that they close most of the teacher-student performance gap across the evaluated settings. Table 5 further shows that the evaluated Trace2Skill variants, including those using teacher trajectories and stronger skill authors, do not recover D-Skill’s gains.
  2. Contrastive learning signal. D-Skill contrasts teacher and student trajectories on the same tasks to extract guidance targeted to the student’s capability gaps. The teacher supplies a stronger behavioral reference that may be missing from a weak student’s own experience. Table 2 demonstrates the benefit of paired analysis over either trajectory source alone. This distillation setup is complementary to skill-generation backends; our implementation uses a modified Trace2Skill consolidator.

The three methods differ from D-Skill in what they learn from and the medium of the learning:

  • Trace2Skill and SkillGen induce skills from agent trajectories and outcome feedback, without explicitly contrasting matched teacher and target-student behavior. D-Skill introduces this contrast as a distillation signal and claims that the signal closes the capability gap. In our evaluation, Trace2Skill closes 8-13% of the reference teacher-student performance gap; its strongest teacher-authored variants close 32-38%, compared with 88% for D-Skill (Table 5).
  • AgentDistill differs from D-Skill in two respects. First, the distillation signal: AgentDistill extracts procedures from teacher trajectories without teacher-student contrast, whereas D-Skill compares their trajectories on the same tasks to target student-specific gaps. Second, the distillation medium: AgentDistill packages knowledge as executable MCP tools, whereas D-Skill produces free-form textual skills that can express procedures, heuristics, exceptions, and verification strategies.
AuthorsOpenReview ↗

Q2. Teacher-student contrast vs. prompt design and analyst strength

Table 2 separates the effect of the contrast from both prompt design and analyst strength. All three conditions use the same analyst and consolidation models, the same 200 distillation tasks, the same lesson format, and the same consolidation procedure (Section 4.2). The teacher-only and student-only prompts are D-Skill’s paired prompt, edited only where it mentions the missing trace so that the analyst studies one trace instead of comparing two. Every other sentence is identical.

With the prompt and the analyst held fixed, the student trace alone closes 38-45% of the teacher-student gap and the teacher trace alone closes 55-71%, while the pair closes 88% (Qwen3.6-35B-A3B: 47.7 and 56.3 vs. 73.2; GPT-5.6 Luna: 49.8 and 63.5 vs. 72.8). The improvement therefore comes from the paired comparison.

The complete prompts are now in the revised PDF, shown in Appendix B in blue. Appendix B.1 gives the exact slot values that complete the Appendix B template for every domain and teacher-student pair. Appendix B.2 lists all nine passages that differ in the teacher-only and student-only prompts, with the changed words in bold. Appendix B.3 lists the complete prompt for the consolidation model.

AuthorsOpenReview ↗

Q3. Scope of the baseline comparisons

Our central finding is that paired teacher-student analysis can produce textual skills that close most of the teacher-student performance gap. Each experiment addresses a specific aspect of this finding:

  • Gap closure (Table 1): The main results demonstrate capability transfer and gap closure across multiple benchmarks and teacher-student pairs.
  • Paired distillation signal (Table 2): Comparing paired analysis with teacher-only and student-only analysis establishes the additional value of the paired contrastive signal.
  • Backend compatibility (Table 4): D-Skill lessons improve SkillOpt while retaining its downstream update procedure, demonstrating compatibility of gap closure with distillation with another skill-update backend.
  • Existing skill induction (Table 5): The Trace2Skill comparisons examine whether existing trajectory-based skill induction, including variants using teacher traces and stronger authors, achieves comparable transfer.
  • Transfer medium (Figure 3): SFT comparisons evaluate textual transfer relative to parameter updates using the same teacher-trajectory corpus.

The differing benchmark coverage reflects these specific experimental objectives; we do not claim that unified evaluation is technically impossible. Together, these experiments support our central finding that paired teacher-student contrast can distill capabilities into textual skills that close most of the teacher-student performance gap across the evaluated settings.

Reviewer_V5WUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes D-Skill, a framework that distills the teacher's behavioral advantages into interpretable and reusable natural-language Agent Skills. These skills are then incorporated into the student's context during inference to improve its task performance.

Critical strengths
  1. D-Skill presents a simple and interesting idea by comparing paired teacher–student rollouts on the same tasks, offering an alternative to Knowledge 
  2. The method shows large and consistent improvements across different benchmarks and teacher–student model pairs.
  3. Promising results compared to SFT. With the same teacher trajectory corpus, D-Skill performs better than one-epoch SFT and achieves results close to multi-epoch SFT after just one offline pass.
Critical weaknesses
  1. It is unclear how skill performance scales with more training tasks or generalizes to different task distributions.
  2. It would be better to compare D-Skill and SFT under a similar computational budget, since D-Skill additionally uses student rollouts and relies on strong analyst and consolidation models.
Questions for authors

1 How many task instances are needed to generate useful skills? Could the authors show how skill length and test performance change as more distillation traces are added? Is there a saturation point? Also, would the learned skills still help when the test tasks come from a different distribution? 2 Could the authors provide a more complete cost comparison that includes student rollouts, analyst calls, and skill consolidation? It would be helpful to compare D-Skill and SFT under a similar compute or monetary budget. 3 Could the authors provide an ablation comparing Divergence-only, Agreement-only, and Full D-Skill? I'm curious how much each component contributes to the final performance and whether combining them provides additional benefits.

3 comments
AuthorsOpenReview ↗

Q1. Tasks needed, skill length, saturation, and OOD transfer

The tables below extend Figure 3 with the length of the skill document after each consolidation chunk. They use the same document lineage as Figure 3 (DeepSeek-V4-Pro → Qwen3.6-35B-A3B), and each score is the mean of two evaluation runs.

SSB-V tasks04080120160200
Skill lines—150257362462592
Accuracy28.059.570.370.375.874.3
ALFWorld games0102029384757
Skill lines—75115148180209241
Success59.073.179.180.679.580.282.5

Few tasks suffice. 40 SSB-V tasks raise accuracy from 28.0 to 59.5, and 10 ALFWorld games raise success from 59.0 to 73.1.

Most gains appear early. About 80 SSB-V tasks provide 91% of the final performance improvement, and 20 ALFWorld games provide 86%. Additional traces yield modest further gains.

OOD transfer. The skills still help on a different task distribution; see our response to reviewer HyhC.

AuthorsOpenReview ↗

Q2: Complete cost comparison and SFT under a matched budget

Table 6 counts only the analyst and consolidation calls. We therefore re-ran the full pipeline on SpreadsheetBench-Verified for the DeepSeek-V4-Pro → Qwen3.6-35B-A3B pair used in the SFT comparison section and logged the usage and cost of every API call for SpreadsheetBench. Teacher rollouts are excluded from both methods since both SFT and D-Skill use the same teacher rollouts, and therefore the costs are identical for this part.

D-Skill stage (SSB-V, 200 tasks)Cost
Student rollouts (Qwen3.6-35B-A3B)$2.73
Analyst calls (DeepSeek-V4-Pro)$1.84
Consolidation (DeepSeek-V4-Pro)$4.40
Total$8.96

For SFT, we multiply the tokens in one epoch of the SFT corpora used for Figure 3 by Tinker's list training price for Qwen3.6-35B-A3B: 1.68 USD per epoch for actions only and 7.62 USD per epoch for reasoning + actions. Accuracies are from Figure 3 and Table 9.

MethodCostSSB-V
D-Skill, one pass$8.9674.3
SFT, actions, 5 epochs$8.4166.0-68.5 (epochs 2-5)
SFT, reasoning + actions, 1 epoch$7.6261.5
SFT, reasoning + actions, 3 epochs$22.8674.5

At D-Skill's budget, SFT reaches at most 68.5. Taking 3 epochs of SFT with both reasoning and actions allow SFT to match D-Skill performance, which costs about 2.6x more. D-Skill's distinctive advantage is that it needs no gradient updates or access to the student's weights, so the same pipeline applies to closed-weight students such as GPT-5.6 Luna, and it produces a readable, editable document. We will add this breakdown to the appendix.

AuthorsOpenReview ↗

Q3. Divergence-only vs. agreement-only vs. full D-Skill

We ran this ablation on one teacher-student pair per benchmark. We reused the lessons from the full D-Skill run, separating them into divergence and agreement only items. Here are the results evaluating with the skills consolidated with divergence-only/agreement-only items (mean ± SD over three documents, each with 2 passes; %GC in parentheses).

SSB-V, DS→QwenALFWorld, DS→QwenChess, Haiku→Qwen
Student / Teacher28.0 / 79.359.0 / 88.424.8 / 41.0
Divergence-only71.9 ± 6.3 (85.6%)83.8 ± 0.9 (84.5%)35.7 ± 0.3 (67.2%)
Agreement-only43.2 ± 2.2 (29.6%)69.4 ± 2.2 (35.4%)30.7 ± 2.1 (36.4%)
Full D-Skill73.2 ± 3.0 (85.1%)84.2 ± 3.3 (85.7%)39.0 ± 3.5 (87.7%)

The effect depends on the task. On SSB-V and ALFWorld, divergence lessons carry most of the effect: divergence-only matches the full method, while agreement-only closes only 22-35% of the gap. On chess, both classes contribute with divergence items contributing more (67.2% vs. 36.4% of the gap on its own), and combining them closes 87.7%. This is intuitive since divergence lessons correct the student’s mistakes; agreement lessons reinforce behaviors it already shows.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.