acceptodds
Under review as a conference paper at ICLR 2027

TAME: Transfer-Aware Model-Harness Co-Evolution through Learned Harness Revisions

Abstract

Agentic post-training relies on an execution harness that both assists the model and shapes the trajectories from which it learns. A harness that improves immediate success, however, need not produce capability that transfers when that support changes. We examine this distinction through crossed training- and deployment-harness evaluation, separating same-harness performance, capability retained after support removal, and transfer across a common evaluation family. This evaluation tests for harness rank reversal, in which a harness with stronger initial performance yields a weaker trained model under common deployment conditions. This motivates TAME (Transfer-Aware Model–Harness Co-Evolution), which trains a Feedback-Agent language model in the self-evolution framework with reinforcement learning to generate persistent, executable revisions to the training harness. Each revision precedes a fixed-budget inner training stage for the Task-Specific Agent and receives a reward based on the resulting performance change under reference harnesses. We employ token-level proximal policy optimization to train the Feedback-Agent on complete revision-and-training trajectories, assigning credit through transfer gains across subsequent model updates. This makes teaching value an explicit objective for harness adaptation: the capability the model acquires through training, rather than only the success it achieves while supported by the harness.

Primary area:
foundation or frontier models, including LLMs
Keywords:
reinforcement learning, language model agents, agentic post-training, harness optimization, model-harness co-evolution, cross-harness transfer, meta-learning, executable program revision

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Pk1iOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the distinction between a harness that improves immediate, supported execution and one that induces model capabilities that persist and transfer after training. It introduces a crossed training/deployment-harness evaluation protocol and trains a Feedback-Agent with token-level PPO to produce persistent executable harness revisions before each fixed-budget inner-training stage. The optimization signal is based on post-update transfer across a fixed family of reference harnesses, thereby assigning credit through subsequent model improvement rather than immediate execution success alone.

Critical strengths
  1. The paper isolates an important evaluation confound: evaluating a training harness only while it remains attached conflates learned model capability with runtime assistance. The crossed-harness protocol directly measures this distinction across multiple models, domains, and deployment harnesses, making the evaluation conceptually valuable.

  2. The empirical design is comparatively rigorous for harness optimization. Three independent Feedback-Agent fits, paired deployments, disjoint data splits, and targeted ablations of the reward, revision persistence, and credit-assignment mechanism provide meaningful support for the central claims.

Critical weaknesses
  1. The reported gains are heterogeneous. In particular, TAME underperforms both the selected fixed teacher and the frozen reviser on Telecom, indicating that it is not a consistently dominant optimizer. Moreover, the transfer objective depends on a designer-specified reference family and assigns substantial weight to the minimal harness, which limits the generality of the transfer claim.

  2. The method is computationally expensive and operationally complex. In the 100M-token ALFWorld setting, it more than doubles the reported GPU hours and model-call tokens relative to the selected fixed protocol. Reproducibility will also depend on releasing implementation details that are not currently specified or explicitly committed to in sufficient depth.

Questions for authors
  1. Could the authors provide a consolidated table reporting common-harness, minimal-harness, held-out-harness, and attached-harness scores for TAME and every baseline across all model-domain settings, together with an analysis of the Telecom failure case?

  2. How sensitive are the conclusions to the composition and weighting of the reference-harness family, particularly the one-half weight assigned to the minimal harness?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Pk1i

Thank you for the constructive review. We added a new analysis for each point.

W1/Q1: one consolidated table, and differences across settings. Tables 14–17 list common, minimal, HheldH_{\rm held}, HnewH_{\rm new}, and own-attached scores for every method in all 12 model–domain settings: selected fixed, information–executive mixture, one-shot, frozen reviser, execution search, evolved fixed, native-system reward, transfer-based search (where run), and TAME. Excluding search, TAME has the highest common transfer in 11 of 12 settings. These include Telecom with Llama-3.2-3B (+4.6 / +2.3 over fixed / frozen) and with Qwen2.5-3B (+3.6 / +2.4). The one exception is Qwen3-4B/Telecom (−1.7 / −1.9), discussed next. Native-system reward has the highest own-attached score in most settings, as expected, since it optimizes the attached system rather than the shared harnesses.

Telecom (App. D.7, Tables 11–12). On Qwen3-4B/Telecom, the validation-selected fixed teacher is HexecH_{\rm exec}. To test whether the starting harness explains the gap, we started TAME, the frozen reviser, and the fixed teacher from HexecH_{\rm exec}. The learned reviser stayed frozen as before, and nothing else changed (3 blocks × 5 deployments each):

Start = HexecH_{\rm exec}CommonMinimalHheldH_{\rm held}HnewH_{\rm new}Own-attached
Fixed HexecH_{\rm exec} (re-run)45.441.846.847.957.5
Frozen reviser46.743.248.049.058.9
TAME49.045.450.351.360.9

From this start, TAME beats the fixed teacher by 3.6 points and the frozen reviser by 2.3, similar to its margins elsewhere, and it is ahead on every evaluator. Its revisions add 0.79 points on average over KEEP, against 0.20 for the frozen reviser (Fig. 6b). Across all four Qwen3-4B domains, TAME's revisions add 0.85–1.02 points over KEEP on average (Fig. 6c). The original gap therefore comes from the starting harness, not from the learned revision rule. The starting teacher is chosen on the selection split, not on test results.

W1/Q2: the reference family and the minimal-harness weight. Table 10 rescores all fixed endpoints with Jα=αJmin⁡+1−α4(Jdocs+Jinfo+Jexec+Jnull)J_\alpha=\alpha J_{\min}+\tfrac{1-\alpha}{4}(J_{\rm docs}+J_{\rm info}+J_{\rm exec}+J_{\rm null}) for every α∈[0,1]\alpha\in[0,1], and with each reference left out in turn. Against the fixed and frozen controls, the sign never changes in any of the 12 settings (TAME is ahead in 11). Against transfer-based search, the sign is stable in 6 of 7 settings; it flips only for Llama/ScienceWorld at small α\alpha. Within this family of weightings, the ½ minimal weight therefore does not drive the conclusions.

Rescoring keeps the trained models fixed, so we also retrained the reviser with each single reference removed from both reward and selection (Fig. 21, App. G.6). Common transfer drops by only 0.4–1.3 points, less than TAME's margin over both controls in each of the 11 settings, and HheldH_{\rm held} drops by at most 0.9. Removing the minimal harness hurts most: the worst score across all eight harnesses falls by 1.5–2.6 points, which supports giving it the largest weight.

W2: cost and complexity. Table 29 now reports cost next to quality. We see transfer-based search as the relevant alternative, since it optimizes the same objective by selection instead of learning. It costs 148.2 GPU-h per deployment, against 106.5 for TAME with construction amortized over 20 deployments, and it still has lower common transfer in 5 of 7 settings. Because search pays its full cost every time, TAME is cheaper from the 13th deployment on. TAME's recurring cost per deployment (39.5 GPU-h) is close to that of the fixed (36.6) and frozen (40.9) arms. Its overhead over the fixed protocol therefore shrinks as the one-time cost is shared: 2.3× at 20 deployments, 1.4× at 100, and approaching 1.08× (App. D.4).

Reproducibility. The revision adds the HnewH_{\rm new} specification with its hash, the partition procedure and counts, and the search protocol. We will release all code and implementation details, including Feedback-Agent training and deployment, the revision validator and editable harness surface, the reference and held-out harnesses, and the task-partition procedure.

We welcome any further questions.

Dear Reviewer Pk1i,

We would welcome your thoughts on the additional analyses in our November 13 response and revised paper.

Following your suggestions, we consolidated the endpoint results in Tables 14–17 and added reference-weight sensitivity and leave-one-reference-out retraining analyses (Table 10; Figure 21). The Telecom follow-up compares the methods from the same executive-harness initialization (Appendix D.7; Table 12), while Appendix D.4 separates teacher-construction costs from recurring deployment costs.

Do these additions help resolve your concerns about the scope of the gains and the practical tradeoff? If any concern remains, please let us know so that we can address it during the remaining discussion period.

Thank you again for your careful feedback.

Dear Reviewer Pk1i,

Thank you for your constructive suggestions.

  • Our consolidated results show competitive, considerable gains with TAME over the fixed and frozen controls, including the selected fixed teacher and the frozen reviser, in most (11 of 12) settings throughout the tested reference-weight family, supported by leave-one-reference-out retraining.
  • In the matched-initialization Qwen3-4B/Telecom follow-up, TAME competitively reaches 49.0 common transfer versus 45.4 and 46.7 for the fixed and frozen controls (Table 12).
  • The revised cost accounting also makes the benefits of reviser reuse explicit.

Together, we believe that these findings strengthen the evidence for transferable improvements and address your questions about their scope and practical value. Would you kindly consider updating your recommendation to 3 (weak acceptance)?

Thank you so much for your consideration.

Best regards,

The authors

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

This paper studies the teaching value of execution harnesses during agentic post-training with the proposed method called TAME. It distinguishes performance achieved with a model’s training harness attached from capability retained under a common minimal harness and transfer across a fixed family of deployment harnesses. Through crossed training–evaluation experiments, it shows that a harness with stronger initial execution performance can produce a weaker trained model under common evaluation conditions. In essence, TAME trains a Feedback-Agent with token-level PPO to generate persistent, executable harness revisions. Each revision precedes a fixed-budget GRPO update of the task model. The outer reward measures the resulting change in performance under independently fixed reference harnesses, with credit propagated across complete sequences of revisions and model updates.

Critical strengths
  1. From the training and evaluation perspective, the authors make an important evaluation distinction and tests it with an appropriately crossed design such that readers can know the contribution/benefit of each component, like the harness-related parts.
  2. In the design, the authors conduct extensive experiments like the reviser learning-related setup includes meaningful controls and unusually careful accounting.
Critical weaknesses
  1. The authors may consider adopting more advanced RL algorithms like DAPO rather than PPO/GRPO to catch the trend
  2. When running the experiments to show superiority, readers can be more easy to understand the improvement when it is trained with a fixed training budget rather than a fixed end-to-end resource budget.
Questions for authors
  1. As mentioned in the weakness point #1, have the authors tested the more advanced algorithm like DAPO?
  2. Like the mentioned point #2, have the authors tested the fixed training budget-related training?
4 comments
AuthorsOpenReview ↗

Response to Reviewer UbnU

Thank you for the positive assessment and the two concrete suggestions. We followed both.

W1/Q1: DAPO. We reran the inner task-model updates with DAPO for TAME and for the fixed and frozen-reviser controls. Everything else stayed the same: the Feedback-Agent, its outer token-level PPO, the 4×25M schedule, the partitions, and frozen deployment (App. D.8, Table 13). Common transfer at 100M tokens on Qwen3-4B:

SettingInnerFixedFrozenTAMETAME − fixed / frozen
ALFWorldDAPO66.268.471.2+5.0 / +2.9
ALFWorldGRPO66.368.771.8+5.5 / +3.2
ScienceWorldDAPO41.041.543.8*+2.8 / +2.3
ScienceWorldGRPO41.142.045.5+4.4 / +3.5

*One block of five deployments; every other cell has 15.

The controls change by at most 0.5 points between optimizers, and TAME stays ahead of both under DAPO (Fig. 6d). On WebShop, the DAPO-trained controls (53.0 / 54.7) also stay below GRPO-trained TAME (58.1). In these settings, the benefit of learning the reviser does not depend on the inner optimizer.

We keep token-level PPO for the outer learner on purpose. The outer return is a delayed transfer gain spread over four revision stages. A learned critic with GAE across stages is what lets early revisions receive credit for gains that appear only after later updates (Sec. 4; Table 4 in App. A). Critic-free group methods such as GRPO and DAPO would need many complete four-stage trajectories from the same start for each group, and each trajectory costs 100M inner tokens. Our ablations also show that removing cross-stage credit lowers transfer (Fig. 5).

W2/Q2: fixed training budget. All quality comparisons fix the training budget, not total resources. Every arm keeps four stages of 25M counted model-call tokens (100M in total), and every call a training harness adds, including retries and auxiliary calls, counts toward that budget. Construction, discarded branches, proposals, and evaluation fall outside the budget and are reported separately (Table 29); App. D.9 now says this explicitly. Quality differences therefore cannot come from extra training on the retained trajectory. Budget sensitivity from 25M to 800M tokens is in the appendix figures. Transfer-based search shows the resource side: with about 1.4× TAME's amortized GPU-hours (400M branch tokens, 100M retained), it still has lower common transfer than TAME in 5 of 7 settings.

Thank you again for these suggestions.

Reviewer_UbnUOpenReview ↗

Response to the author comments

Thank you authors for putting these details. To be honest, it is a little bit out of my expectation that DAPO is weaker than GRPO given that DAPO is an advanced version of GRPO.

Reviewer_UbnUOpenReview ↗

Follow-up on the degraded performance of DAPO

Any follow up on the previous question?

Dear Reviewer UbnU,

Thank you for following up, and apologies for the delayed response. DAPO introduces valuable improvements to GRPO, whose benefits can vary with the task and training configuration. Our experiment evaluates cross-harness transfer in interactive environments under a fixed 100M-token budget, extending beyond the mathematical-reasoning setting of DAPO’s original evaluation. The ordering we observed is specific to this configuration.

The encouraging result is that TAME’s gains carry over to DAPO: on ALFWorld, TAME achieves 71.2 common transfer, compared with 66.2 for the fixed teacher and 68.4 for the frozen reviser, all using DAPO. This strengthens the evidence that learning harness revisions provides useful improvements with both evaluated inner optimizers.

Best regards,

The authors

Reviewer_fqUVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper examines whether an execution harness that improves an agent’s immediate task performance also provides an effective environment for training the underlying model. Its cross-evaluation compares models trained under different harnesses using common deployment conditions, including a minimal harness and a weighted family of reference harnesses. This allows initial execution rankings to be compared with the rankings of the resulting trained models. TAME trains a Feedback-Agent to generate executable harness revisions before each fixed-budget GRPO training stage. Accepted changes persist, and PPO assigns credit across complete revision-and-training trajectories using the task model’s performance improvements under fixed reference harnesses. During fresh training runs, the reviser’s parameters are frozen while harness revisions and task-model updates continue. Experiments cover three task models and four domains, with a primary retained training budget of 100M model-call tokens. They show that initial harness rankings can reverse after training. In the primary Qwen3-4B comparisons, TAME improves common-reference performance over fixed and frozen-reviser controls on ALFWorld and ScienceWorld, but underperforms both on Telecom.

Critical strengths

S1. Useful problem formulation and evaluation. The paper addresses a practical evaluation problem for models reused under different execution harnesses. Strong performance with a supportive harness does not establish that the same harness produces the most useful training experience. Evaluating trained endpoints under shared reference harnesses makes the comparison less dependent on the assistance supplied by each method’s own final program. The minimal-harness evaluation and the broader crossed results provide complementary views of performance after support is reduced or changed. Rank reversals on ALFWorld and WebShop, alongside settings where executive assistance retains an advantage, give concrete evidence that immediate execution quality is insufficient on its own for selecting a training harness.

S2. Relevant controls and informative ablations. The frozen-reviser control uses the same generator initialization, editing interface, revision schedule, and task-model training budget, providing a relevant test of learning the revision policy. Comparing common-reference and native-system rewards after task-model updates also examines how the choice of objective changes the resulting teacher. The ablations on persistent source changes and credit across successive stages connect the sequential design to downstream performance. Improvements over fixed and frozen-reviser controls on ALFWorld and ScienceWorld support the method’s usefulness in those settings. Reporting the negative Qwen3-4B/Telecom result, harmful individual revisions, and teacher-construction costs also gives readers evidence about where the approach can fail and what it requires.

Critical weaknesses

W1. The held-out evaluation and task partitions are insufficiently specified. The paper reports eight fixed evaluation harnesses, but Table 7 describes only seven and does not specify the held-out harness. Its observations, prompts, control flow, and recovery behavior are therefore unclear, making it difficult to judge what change in deployment conditions the held-out results test. This directly affects the claimed extension beyond the harnesses included in the training reward. The data protocol also states that four task partitions are disjoint without reporting their per-domain sizes or a concrete procedure for identifying and grouping related task variants. These details determine the evaluation’s scale and allow its intended separation to be assessed. An explicit held-out harness specification and a reproducible description of the four partitions are needed to interpret the scope of the reported generalization.

W2. The transfer-based search comparison is incomplete. Table 9 lists “Transfer-based search” separately from one-shot optimization and execution-based search, but Figures 18 and 22 do not identify its corresponding performance. The manuscript also does not clearly connect this label to a candidate-selection procedure or another named baseline. The frozen-reviser control tests the effect of training the generator, while WHALE selects candidates using the current model’s execution performance; these comparisons leave the results of the separately costed search alternative unresolved. With shared construction cost amortized over 20 deployments, Table 9 reports 106.5 GPU-hours per deployment for TAME and 147.8 for transfer-based search. Without the latter’s performance, this cost difference cannot establish their quality–cost tradeoff. Identifying or reporting its common-reference and held-out scores, together with its search procedure, would clarify the practical benefit of learning the reviser.

Questions for authors

Q1 — Held-out evaluation and task splits. Please specify the eighth held-out evaluation harness and its differences from the five reward harnesses. What are the per-domain task counts in the four splits, and how were related task variants grouped when constructing them?

Q2 — Transfer-based search comparison. How does the “Transfer-based search” baseline in Table 9 select harnesses, and which performance results correspond to it? Please identify or provide its common-transfer and held-out scores at the reported 100M retained training-token budget, alongside its reported cost, for comparison with TAME.

3 comments
AuthorsOpenReview ↗

Response to Reviewer fqUV

Thank you for the precise questions. Both point to information missing from the submission, and the revision now provides it.

W1/Q1(a): the held-out evaluator HheldH_{\rm held}. It is now in the harness table (Table 20 in the revision) and specified in App. D.1. It was frozen before any training or selection and was never used for outer rewards, candidate ranking, checkpoint selection, or reviser feedback.

  • What it gives the model: static environment and command documentation, the public tool schema, and one deterministic fix of the first malformed action.
  • What it does not give: subgoals, lists of admissible actions, state or inventory summaries, multi-step retries, hidden state, solutions, or recommended actions. The model must plan, choose, and format every action itself. After a schema or recoverable tool error, the environment is left unchanged and the harness returns the error category with the relevant schema. The model writes its own retry, and later failures get no help.
  • Protocol: the task, verifier, model, environment, action space, and limits are the same as for the other seven harnesses. All calls and invalid decisions count, and each score averages four rollouts per task.

In short, HheldH_{\rm held} is HdocsH_{\rm docs}-style documentation plus the schema and one deterministic recovery. It has none of HinfoH_{\rm info}'s state support and none of HexecH_{\rm exec}'s subgoals or multi-step repair. No reward harness offers this level of recovery support, so it tests transfer to a new condition.

A second held-out evaluator, HnewH_{\rm new}. We specified it independently and froze it on 2026-10-18, before any outcome-dependent analysis; its SHA-256 digest is in App. D.2. It shows no documentation, state summaries, or subgoals. The history is a ledger of at most four recent events (decision, raw action, schema status, executed action, public return) within 1,024 tokens. After a schema failure the model writes every retry itself, with no deterministic repair, and one certified read-only observation refresh is allowed per episode. HnewH_{\rm new} is reported for all methods in all 12 settings (Tables 14–17) and adds to, rather than replaces, HheldH_{\rm held}. In every setting, the two evaluators agree on the sign of TAME's difference from the fixed and frozen controls.

W1/Q1(b): task partitions. We group related variants by source-parent and variant-family metadata and by exact normalized duplicates. Groups are ordered by SHA256("tame-split-v1|20261001|"‖domain‖"|"‖group key) and assigned 60/15/10/15% to the inner/outer/selection/final splits; official benchmark partition rules take precedence. Every method, model, and fit uses the same realized partition. Counts are tasks (groups):

DomainInnerOuterSelectionFinal
ALFWorld162 (54)41 (14)27 (9)40 (13)
ScienceWorld198 (66)50 (17)33 (11)49 (16)
WebShop312 (104)78 (26)52 (17)78 (26)
Telecom96 (32)24 (8)16 (5)24 (8)

W2/Q2: transfer-based search. The procedure is in Sec. 6 and App. D.4. At each of four stages, search copies the full parent learning state and trains four branches for 25M tokens each: one that keeps the current harness (KEEP) and three with edits from the frozen generator. It continues from the branch with the highest post-update common transfer on the outer-reward split. Search thus optimizes TAME's reward by picking among trained branches instead of learning a policy. Each deployment trains 400M branch tokens, of which 100M are kept. The 15 search deployments are paired with TAME's 3×5 blocks. Results on Qwen3-4B/ALFWorld at 100M retained tokens (Table 3):

CommonMinimalHnewH_{\rm new}Own-attachedGPU-h/deploy (N=20)
Transfer-based search70.9±1.667.4±1.673.1±1.979.5±1.4148.2
TAME71.8±1.268.2±1.274.4±1.281.0±1.3106.5

We did not run HheldH_{\rm held} for the new search arm, so the table shows HnewH_{\rm new} for both.

Search pays 148.2 GPU-h on every deployment, while a deployment with the frozen learned reviser costs 39.5 GPU-h. TAME's one-time construction therefore pays for itself after 13 deployments in GPU-hours (14 in tokens). Across all seven search settings (Table 9), TAME has higher common transfer in 5, by 0.5 to 1.2 points (both exceptions are Telecom), and higher HnewH_{\rm new} in 6. Its individual edits are also more useful than the frozen generator's: +0.99 vs +0.48 points over KEEP on ALFWorld (Fig. 6a). Table 29 now includes the search row, so cost and quality can be compared side by side.

We would be glad to clarify anything further.

Dear Reviewer fqUV,

Thank you again for your helpful review. As the discussion period approaches its close, we wanted to follow up on our November 13 response.

The revision now specifies the original held-out harness and the task-partition procedure and counts (Appendices D.1–D.3; Table 8). It also provides the transfer-based search procedure, performance, and cost comparison (Appendix D.4; Table 3), including results for both methods under the additional held-out evaluator HnewH_{\mathrm{new}}.

We would appreciate hearing whether these additions address your two main concerns, or whether any further clarification would help your assessment. We would be happy to respond before the discussion closes.

Thank you for your time.

Dear Reviewer fqUV,

Thank you again for your time and thoughtful feedback. We would like to follow up on our response and ask whether you have any remaining questions or concerns. We welcome further discussion and would be glad to provide any additional clarification.

  • The revision specifies the held-out harnesses and task partitions and adds the requested transfer-based search comparison.
  • On Qwen3-4B/ALFWorld, TAME competitively achieves 71.8 versus 70.9 common transfer and 74.4 versus 73.1 under the additional held-out evaluator, with 28% fewer GPU-hours per deployment when construction is amortized over 20 deployments (Table 3).

These results strengthen the practical and effective case for learning harness revisions and directly address your evaluation and baseline questions. We would be very grateful if you could consider raising your recommendation to 3 (weak acceptance).

Thank you so much for your consideration.

Best regards,

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their careful reading. The reviews asked for four things: (i) a full specification of the held-out evaluation and the task partitions, (ii) a transfer-based search baseline with its cost reported separately, (iii) all endpoints in one place, with an analysis of Telecom, and (iv) evidence that the conclusions do not hinge on the reference weights, the inner optimizer, or how the budget is defined. The revision adds all four.

Main updates

  1. Transfer-based search (fqUV, Pk1i). This baseline uses TAME's reward to select revisions instead of learning a reviser. At each of the four stages, it trains four branches from the same parent for 25M tokens each (one keeps the harness unchanged, three apply edits from the frozen generator) and continues from the branch with the best post-update common transfer. On Qwen3-4B/ALFWorld at 100M retained tokens, TAME scores higher on every evaluator we ran for both methods:

    • common transfer 71.8 vs 70.9, minimal 68.2 vs 67.4, HnewH_{\rm new} 74.4 vs 73.1, own attached harness 81.0 vs 79.5;
    • cost: 106.5 vs 148.2 GPU-hours per deployment, with TAME's one-time construction spread over 20 deployments (Table 3).

    Across the seven settings where search was run, TAME has higher common transfer in 5 and higher HnewH_{\rm new} in 6 (Table 9). Individual edits also help more: on ALFWorld, a TAME revision adds 0.99 points on average over leaving the harness unchanged from the same starting point, against 0.48 for a frozen-generator edit (Fig. 6a).

  2. Held-out evaluation (fqUV). The eighth harness, HheldH_{\rm held}, is now in the harness table (Table 20 in the revision; Table 7 in the submission) and fully specified in App. D.1. We also add a second held-out evaluator, HnewH_{\rm new}. It was frozen with a SHA-256 digest before any outcome-dependent analysis and is reported for every method in all 12 settings (App. D.2). When TAME is compared with the fixed and frozen controls, the two held-out evaluators give the same sign in every setting. App. D.3 and Table 8 give the partition procedure and the realized split sizes.

  3. All endpoints in one place (Pk1i). Tables 14–17 list common, minimal, HheldH_{\rm held}, HnewH_{\rm new}, and own-attached scores for every method in all 12 settings. Excluding search, TAME has the highest common transfer in 11 of 12 settings, 2.3–5.7 points above the fixed and frozen controls. In the twelfth, Qwen3-4B/Telecom, starting TAME from the validation-selected executive teacher gives +3.6 points over that teacher and +2.3 over the frozen reviser (App. D.7, Table 12).

  4. Reference weights (Pk1i). Against the fixed and frozen controls, TAME's difference keeps its sign for every minimal-harness weight α∈[0,1]\alpha\in[0,1] and when any one reference is dropped, in all 12 settings (Table 10). We also retrained the reviser with each reference removed, and TAME stays ahead of both controls in all 11 settings where it leads (Fig. 21, App. G.6).

  5. DAPO and the budget (UbnU). With DAPO as the inner optimizer for TAME and both controls, TAME leads by 5.0 / 2.9 points on ALFWorld and 2.8 / 2.3 on ScienceWorld (App. D.8). All quality comparisons use the same training budget of 4×25M counted tokens; end-to-end resources are reported separately (App. D.9).

  6. Code (Pk1i). We will release all code and implementation details (Reproducibility statement).

We also replaced two vague phrases ("in some settings", "across most experiments") with exact counts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.