acceptodds
Under review as a conference paper at ICLR 2027

SkillBackprop: Two-Stage Textual Backpropagation for Skill Optimization

Abstract

Optimizing skills for frozen language models requires more than identifying failures. It also requires assigning task loss appropriately and determining how the skill should be revised. Unlike differentiable model parameters, skill instructions do not admit analytic derivatives that can directly guide these decisions. To provide a backward learning pathway for skills, we introduce SkillBackprop, a two-stage textual backpropagation framework for structured skills. SkillBackprop separates backward learning into failure localization and loss-reduction estimation. The first stage aggregates execution evidence to trace failed outcomes to the responsible parts of the skill. The second measures how candidate revisions affect validation loss and uses the resulting loss-reduction signals to guide skill updates. Together, the two stages provide a backward pathway from task-level loss to targeted modification. We evaluate SkillBackprop on SpreadsheetBench, OfficeQA, SearchQA, and DeepPlanning-Shopping, where it achieves scores of , , , and , respectively. Across all benchmark and model settings, SkillBackprop matches or outperforms the strongest baseline, with gains of up to percentage points. Ablations further show complementary benefits from failure localization and loss-reduction estimation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, agent learning, agent skills, prompt optimization, skill optimization, credit assignment, execution-aligned learning, text-space optimization, structured language policies

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KpynOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper optimizes reusable textual skills for frozen LLM executors on multi-stage agentic tasks. It argues that natural-language "textual gradients" implicitly fuse two decisions — where in a skill to apply an update and whether a proposed revision actually reduces task loss. SkillBackprop separates them: a skill is compiled into a fixed operator DAG; Stage 1 (failure localization) uses an LLM Analyzer to attribute each failed execution to exactly one responsible operator and aggregates loss-weighted blame across a mini-batch; Stage 2 (loss-reduction estimation) has an LLM Optimizer propose typed candidate edits at the selected operator, each measured on a validation subset, accepting only candidates with positive measured loss reduction (ε = 0), with an epoch-level consolidation ledger. Evaluation: SpreadsheetBench, OfficeQA, DeepPlanning-Shopping, SearchQA × three frozen backbones (DeepSeek-v4-Flash, MiniMax-M3, GLM-5.3) vs No Skill / Compiled Skill / GEPA / SkillOpt, plus ablations, a blinded human re-audit of 103 failure attributions (93.2% supported), and a cross-model transfer study. Reported average: 74.24 vs 65.46 for the strongest baseline (SkillOpt).

Critical strengths
  1. A principled decomposition with a measured (not LLM-judged) acceptance rule. The where/whether separation is a real conceptual delta over GEPA/SkillOpt, candidate promotion is deterministic code rather than LLM preference, and the ablations (random/first-failure attribution, accept-all, monolithic skill, no clustering, no consolidation) support that the components are complementary. The blinded human audit of failure attributions is a welcome validation step that most textual-gradient work skips.
  2. Unusually transparent engineering. Identical task-exposure and update schedules vs SkillOpt, benchmark-native metrics (no LLM-as-judge), validation Q/A exclusion from the policy, full initial+final skill artifacts, prompt contracts, split manifests, and qualitative traces (including a rejected candidate) are all provided. This substantially raises confidence that the pipeline does what is described.
Critical weaknesses
  1. The central comparative claim is statistically unsupported. Every model×benchmark cell is a single archived run — no variance, CIs, or significance tests anywhere. On DeepPlanning the margin over the strongest baseline equals exactly one test case (1.19 pts of 84; 75.00 vs 73.81); on SearchQA it is ≈6/840 cases. The transfer study itself provides internal evidence of large run-to-run variance: a GLM-5.3-learned skill transferred to DeepSeek-v4-Flash beats the skill trained on DeepSeek itself by +22.9 pts on SpreadsheetBench (84.29 vs 61.43, Table 3), an unexplained inversion. The abstract's headline numbers are the best model's (GLM-5.3) row, not cross-model averages. Multi-seed results could resolve this, but as submitted the claim "matches or outperforms the strongest baseline across all settings" is not established.
  2. Thin baseline coverage and unverified budget parity. The paper's named foils — TextGrad and ProTeGi — are never evaluated, which matters because ProTeGi already couples textual "gradients" with validation-based beam selection, undercutting the novelty framing of Stage 2; the closest work (SkillGrad) is cited but not benchmarked. GEPA is capped at 100 metric calls while SkillBackprop consumes ~160 training rollouts plus per-candidate validation evaluations whose total rollout/token cost is never reported, so compute parity is unestablished. Additionally, the ε = 0 acceptance threshold on 8-task validation sets (one task = 12.5 pts, ~20 sequential accept decisions) is fragile against adaptive overfitting, and the emphasized relative gains are computed against Compiled Skill, which is worse than No Skill in 8 of 12 cells.
Questions for authors
  1. Can you report ≥3 independent optimization runs per model–benchmark cell with standard deviations, and explain the Table 3 transfer inversion (transferred skill beating the target-trained skill by 22.9 pts)? This single question most affects my assessment.
  2. What is SkillBackprop's total rollout/token cost per benchmark, including candidate evaluations on V, versus GEPA's 100 metric calls? Can you add a compute-matched GEPA configuration and TextGrad/ProTeGi baselines?
  3. The human audit uses a single annotator with no inter-rater reliability, and OfficeQA agreement is only 79.3% (11/29 ambiguous). How sensitive are the downstream gains to localization errors?
  4. (Clarification) Algorithm 1's EC(v) = f_contract(v) + α f_lineage^p(v) + β f_downstream(v) and U(v) use undefined f-functions, α, β, p, T. Please define them and provide sensitivity analysis.
Reviewer_jcCvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SkillBackprop optimizes textual skills for a frozen LLM executor. A skill is compiled into a short chain of operators, and each update has two stages. First, failures are localized to the operator at the earliest divergence that matters to the scorer. Second, candidate edits to that operator are accepted only if they strictly improve accuracy on a fixed validation subset. The method is evaluated on SpreadsheetBench, OfficeQA, DeepPlanning-Shopping and SearchQA with three executors. It reports averages of 71.5–78.4, against 65.5–66.1 for SkillOpt and GEPA. The paper further includes ablations, a human audit of localization, and cross-model transfer.

Critical strengths

S1. The decomposition of the update step is clear and well motivated. Deciding where to edit (failure localization) is kept separate from deciding whether an edit helps (measured validation gain). The accept/reject decision is made in code rather than by LLM judgment. This is a sensible response to the criticism that "textual gradients" are not real gradients.

S2. The improvements are consistent across settings. SkillBackprop matches or outperforms the strongest baseline in all 12 model–benchmark combinations. The paper complements these main results with component ablations, a blinded human audit and cross-executor transfer.

S3. The paper is exceptionally transparent. It provides the full initial and learned skills (App. I), pseudocode and prompt contracts. This openness is commendable and made the detailed inspection in W2 possible.

Critical weaknesses

W1. The results rely on single runs and very small selection sets. No seeds, SDs or confidence intervals are reported. Candidate edits are accepted on validation subsets of only 8, 8, 12 and 60 tasks, with ε = 0 (Table A1). On SpreadsheetBench and OfficeQA, a single task flip therefore changes validation accuracy by 12.5 pp, which makes acceptance highly susceptible to noise. The transfer results point the same way: on SpreadsheetBench, the GLM-5.3 skill transferred to DeepSeek scores 84.29, whereas DeepSeek's own optimized skill scores only 61.43 (Table 3). Mean ± SD over at least three runs is needed.

W2. The size of the gains is hard to reconcile with the learned skills. According to App. I, the final OfficeQA skill for GLM-5.3 differs from the compiled initial skill by a single passage appended to retrieve_evidence.rule_005, yet test EM rises from 29.76 to 70.24. Likewise, the SpreadsheetBench skill changes only rule_009 and rule_010 and adds one write_file recovery sub-skill, while accuracy rises from 35.71 to 87.86. Three additions would help explain these jumps: failure-category breakdowns before and after optimization, per-clause ablations, and repeated evaluations of both skills in the same environment.

W3. Part of the improvement may reflect evaluator or tool quirks rather than transferable knowledge. The revised rule_010 tells the agent to write literal values even when the user asks for a formula, because uncached formulas read as empty at scoring time. The new sub-skill targets the overwrite flag of write_file. Moreover, No Skill outperforms Compiled Skill in 8 of 12 settings. The relative gains over Compiled Skill (up to +197%) are therefore inflated by a weak starting point, and gains over No Skill and the strongest baseline would be more informative.

W4. The compute budgets of the compared methods are not matched. Table A6 equalizes the number of update opportunities, but not rollouts, LLM calls or tokens. Some baselines are also "reconstructed from archived manifests" (App. D). A fair comparison requires reporting rollouts, optimizer tokens and wall-clock time for every method.

Questions for authors

Please see the Weaknesses section.

Reviewer_gYEtOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SkillBackprop, a framework for optimizing structured agent skills in multi-stage tasks. The method organizes each skill as an execution graph of operators and uses execution traces to identify which stage is responsible for task failure. It then evaluates candidate skill revisions on a validation set to select updates that effectively reduce task loss. Experiments across 4 benchmarks and 3 LLMs show that SkillBackprop outperforms existing baselines, with larger improvements on multi-stage tasks.

Critical strengths

S1. The paper addresses an important problem of attributing task-level failures to specific skill components in multi-stage tasks, which is relevant to improving agent reliability and skill reuse.

S2. The proposed framework clearly separates failure localization from evaluating the effectiveness of candidate skill updates.

S3. SkillBackprop achieves outperforming results across all 12 model and benchmark settings, with larger improvements on multi-stage tasks compared to existing baselines.

Critical weaknesses

C1. The module of failure localization, targeted skill edits, and validation-based update selection have already been explored in existing works such as SkillOpt, SkillAdaptor, and TRUCE. The paper does not clearly clarify the novelty of combining these mechanisms in operator DAGs. It also lacks direct comparisons with relevant existing methods.

C2. The paper attributes each failure to the earliest operator where a relevant error occurs. However, this may not be the best component to update, as multiple operators may contribute to the same failure.

C3. The paper focuses on credit assignment in multi-stage skills. But only 2 of 4 benchmarks involve multiple stages, with just 2 and 3 sequential operators. The lack of experiments on branching dependencies and failures involving multiple stages limits the generalizability of the findings.

Questions for authors

Q1. Can the authors provide direct comparisons with existing methods to better clarify the novelty of SkillBackprop?

Q2. Please extend the evaluation on more complex workflows beyond the current two- and three-stage sequential chains.

Q3. Why is the earliest operator associated with a failure considered the best target for an update? Is there empirical evidence that updating this operator is more effective than updating alternative operators?

Q4. The loss-weighted aggregation in Eq. (2) appears equivalent to counting failures attributed to each operator. What additional benefits does this formulation provide over simple failure counting?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.