acceptodds
Under review as a conference paper at ICLR 2027

Training Beyond Reach: Recursive Self-Improvement with Verifiable Circuit Tasks

Abstract

Circuit-design agents must coordinate structural edits, repair, and tuning to satisfy coupled specifications, yet complete-design successes can be scarce, limiting self-training. We propose (Verifiable Escalation and Reduction for self-training Agents), which uses verified circuits to supply recovery practice and seed further task generation. Reduction changes the starting state to lower the remaining decision burden, normally retaining the seed's full specifications, and certifies a feasible recovery path. Within successful learner rollouts, action selection identifies verified progress, repair, and completion steps for supervision. After each update, the improved learner searches for stronger-specification seeds that yield a fresh recovery-task bank under a fixed verifier. A bootstrap bank derived from human-constructed seeds supplies 1,200 selected actions. One update raises the 9B learner's held-out complete-design success from 0/180 to 77/180 (42.8%). Success-trajectory SFT on the same trajectories reaches 47/180 (26.1%); also improves performance on two external circuit benchmarks. After three updates, the learner reaches 123/180 (68.3%), exceeding fixed-bank recollection and frozen-search controls. These results support linking accessible recovery practice to model-driven renewal of circuit training tasks. Our source code is at https://anonymous.4open.science/r/anonymous_code1-F4C3/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Data Synthesis for Agents, Long-Horizon Agent, Agent Training Environments, Analog Circuit Design

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_K9B4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

VERA constructs certified recovery tasks from verified circuits, trains on selected actions from successful learner rollouts, and uses the updated learner to find stronger-specification seeds for subsequent rounds. Internal success increases from 0/180 to 77/180 after one update and 123/180 after three, with improvements on two external benchmarks.

The recovery-task formulation is useful, and the action-selection results are encouraging. My main concern is that the recursive experiments leave task quality, data volume, and computational cost partly entangled.

Critical strengths
  1. The paper addresses a real self-training bottleneck: complete-task successes may be too scarce to provide useful supervision. Certified recovery tasks offer a practical starting point, and the private recovery record keeps feasibility checking separate from learner supervision.

  2. The comparison with success-trajectory SFT is informative. Using the same successful episodes, selected-action training reaches 77/180 versus 47/180 when supervising all actions. This supports filtering within successful trajectories, although supervision volume also differs. The equal-data and equal-collection controls further support recovery-based training on easy tasks.

  3. Fixed-bank and frozen-search controls provide useful evidence beyond successive checkpoints. External transfer and the negative result for early expansion also strengthen the empirical study.

Critical weaknesses
  1. The recursive gains need better resource controls.

At the third update, full recursion reaches 123/180, versus 97/180 for fixed-bank recollection and 103/180 for frozen search. Cumulative collection also differs: 1,566, 1,404, and 1,446 rollouts, respectively, with different selected-action pool sizes. Search, construction, and certification costs are not included.

Producing more useful data may itself be a benefit of task renewal. Still, the current comparison does not show whether recursion improves performance at a fixed budget, or how much its advantage depends on additional collection and training. A matched-budget comparison and full cost breakdown would substantially strengthen the central claim.

  1. Evaluation diversity and training stability remain uncertain.

The internal benchmark contains 18 tasks, but all six tasks within a family share one starter netlist. It therefore offers limited evidence about generalization across starting topologies. Each condition also uses a single training seed; repeated inference rollouts do not establish stability across training runs.

The external results help, but their scope matters. On PhyChip-v2, first-update success is 72/84 on execution-based L1 tasks and 11/66 on performance-graded L2/L3 tasks. These demonstrate different capabilities and should remain separate when interpreting transfer. AnalogCoder-Pro evaluation covers only the 20-task executable subset.

  1. The added value over simpler recovery curricula is unclear.

The existing controls test fixed-bank recollection and frozen-model search, but do not cover expanding the perturbation bank without tightening specifications or adapting recovery difficulty alone. Such comparisons would help determine whether a stronger specification search is needed, rather than task diversity or difficulty adjustment more generally.

Questions for authors
  1. Does the recursive advantage persist with matched collection budgets and controlled training tokens or optimization steps? What is the total search and certification cost?

  2. Can the key comparisons be repeated across training seeds, with uncertainty estimates accounting for repeated rollouts within tasks?

  3. How similar are training seeds and held-out starters? Would the gains persist with additional starting topologies?

  4. How does target expansion compare with an adaptive recovery curriculum that retains the original specifications?

2 comments
AuthorsOpenReview ↗

Response to K9B4 (1/2): Budget-Matched Recursion and Statistical Reliability

We thank the reviewer for the feedback. The paper reports collection volume, defines aggregation, separates external evaluation criteria and includes frozen search. The revision adds fixed-budget recursion, independent training runs, new starting topologies and same-specification adaptive renewal to address the remaining questions.

1. Fixed-budget recursion and full costs

Q1. Does recursion help at a fixed budget after including search, construction and certification costs? (W1; Question 1)

The paper explicitly reports resource differences between recursive branches. Section 4.5 and Table 18 give cumulative attempts of 1,566 / 1,404 / 1,446 and action pools of 1,650 / 1,480 / 1,410, excluding construction and certification. The revision extends that resource accounting to fixed-budget comparisons.

We fix collection/SFT quotas or impose a common verifier cap covering search, failed proposals, construction, certification, calibration and collection. Cumulative caps are 3,600, 7,300 and 11,600 calls (Appendix G.2). Including search and certification, full VERA retains a consistent lead of about 13 percentage points over the fixed bank (Table 5, p. 9, 8 seeds).

Budget settingFullFixed bankPaired difference
Fixed collection/SFT quotas68.1 ± 3.1%54.4 ± 2.6%+13.8 pp, 95% CI [12.1, 15.4]
Common verifier cap66.3 ± 3.6%52.9 ± 3.8%+13.4 pp, 95% CI [10.6, 16.2]

Actual cumulative calls at the third update are reported in Table 20 (p. 31).

Settingcalls
Cap Full11,182 ± 218
Cap Fixed11,170 ± 236
Joint Full11,184 ± 216
Joint Fixed11,180 ± 240

Figure 8, right (p. 30), plots results by cumulative cap. The appendix (p. 30) also breaks down full VERA's second-update task-production cost into 1,412 search calls, 336 construction/calibration calls and 424 certification calls.

Revision: Table 5 and Appendix G.2 (Tables 19–20; Figure 8) report joint budget matching, cumulative verifier caps and the full task-production cost breakdown.


2. Training stability and the statistical unit

Q2. Are gains stable across training seeds, with uncertainty computed at the appropriate statistical unit? (W2; Question 2)

The paper defines task-level aggregation, and the revision adds independent training runs. Originally, each condition had one training run (seed 42), with a checkpoint fixed by training length. Success was averaged over 10 rollouts per task, then equally over 18 tasks (Appendix A.3, p. 18). These repeated rollouts characterize inference variability.

We use 8 independent training seeds for the primary comparisons and 5 of them for component and prior-method comparisons, each with independent collection and model updates. Branches are paired within seed. Each checkpoint's 10 rollouts per task are averaged first, followed by equal task weighting. Mean and sample SD are computed across training seeds. Paired intervals and randomization tests also use training seeds as the statistical unit, and the main conclusion holds (Appendix G.1, p. 28).

Item (Table 5, Joint setting)Value
Full68.1 ± 3.1%
Fixed bank54.4 ± 2.6%
Paired gain+13.8 pp, 95% CI [12.1, 15.4]
Raw p for the four primary comparisons0.0078125 (Holm-adjusted 0.03125)

Table 1 (p. 6) retains the single-run success counts, labeled separately from repeated-training statistics.

Revision: Table 5 and Appendices A.3 and G.1 specify task-level aggregation, independent training runs, paired bootstrap intervals and corrected significance tests.

AuthorsOpenReview ↗

Response to K9B4 (2/2): Topology Transfer, External Evaluation, and Adaptive Controls

3. Transfer to independent starting topologies

Q3. Do gains persist with independent starting topologies rather than shared family-level starts? (W2; Question 3)

The original 18 tasks test transfer across specifications from shared starts. All 6 tasks within a family use D01 (Section 4.1, Appendix A.5 and Table 9, pp. 19–20), with possible training–evaluation topological similarity noted in Appendix A.2.

We evaluated existing checkpoints on 6 independent topology groups (18 tasks), checking graph isomorphism, ancestry and local structural similarity. The new starts have no isomorphic or ancestral match in the training archive. Performance continues to improve, with full VERA gaining 11.8 percentage points over the fixed bank (Table 6, p. 9, 8 seeds).

Evaluation setFixed bankFullPaired gain
New topologies (6 groups, 18 tasks)47.9 ± 2.5%59.7 ± 2.2%+11.8 pp, 95% CI [10.5, 13.1]

Table 24 (p. 34) also gives the base model's 0.0% and the first-update model's 36.2 ± 3.6%. Appendix G.6 describes the checks, and Figure 11 (p. 34) reports each topology group's performance.

Revision: Table 6 and Appendix G.6 (Table 24; Figure 11) report independent starting topologies, structural/lineage audits and transfer across 8 training seeds.


4. External evaluation criteria and task coverage

Q4. How should PhyChip execution and performance gains, and the AnalogCoder-Pro subset, be interpreted? (W2)

The paper separates PhyChip execution from performance and specifies the AnalogCoder-Pro subset. Table 12 (p. 25) reports L1 execution and L2/L3 performance separately, with performance gains discussed in Section 5 (p. 10). Appendix D identifies the 20 AnalogCoder-Pro tasks with complete assets (IDs 1–8 and 10–21), shared by all models.

Appendix D (p. 25) now clarifies that Table 1, Figure 6(b) (p. 8) and Table 13 (p. 26) pool PhyChip L1 execution with L2/L3 performance. Performance-transfer claims use the separate criteria. Performance success also improves from 0/66 to 11/66.

PhyChip levelL1 (execution)L2 (performance)L3 (performance)L2+L3
Base8/840/450/210/66
First-update model72/849/452/2111/66

AnalogCoder-Pro retains the same task set and denominator. The new VREF/LDO/FLT tests (Tables 6 and 24) provide repeated-run evidence of cross-family transfer, complemented by results from the external subset.

Revision: §5 and Appendix D (Table 12) distinguish execution from performance and specify the AnalogCoder-Pro subset; Tables 6 and 24 add cross-family transfer results.


5. Adaptive renewal at unchanged specifications

Q5. Does stronger-specification search add value over an adaptive recovery curriculum at unchanged specifications? (W3; Question 4)

The paper already tests the search model and the timing of target expansion. Frozen search holds the initial model fixed for target search, with other settings unchanged, reaching 103/180 versus full VERA's 123/180 at update 3 (Table 17, p. 27). The V/H experiment shows that expansion timing matters (Table 4, p. 9). Same-specification adaptive renewal is the additional control in the revision.

The same-specification branch uses the updated learner to find more verified designs within the original bounds. These seed new exercises, with splitting and perturbation weakening enabled. It therefore controls for design diversity, task renewal and difficulty adaptation, with tightening target specifications as the sole difference. Renewal at the original bounds already helps. Stronger-specification seeds add a further 7.8 percentage points, 95% CI [6.4, 9.2] (Table 5, p. 9, 8 seeds).

Budget settingSame-specification renewal (same-spec)Full
Fixed collection/SFT quotas60.3 ± 3.6%68.1 ± 3.1%
Common verifier cap58.6 ± 3.3%66.3 ± 3.6%

Table 19 (p. 29) and Figure 8 (p. 30) give second-update results and training trajectories. The strong hybrid adds a full-method comparison (Table 21, p. 31).

Disabling splitting and weakening gives 64.0 ± 2.8%, versus 68.8 ± 2.3% for full VERA, confirming the benefit of adaptive practice (Table 7, Appendix G.5). The same-specification branch combines seed renewal and difficulty adjustment, directly testing adaptive renewal within the original specification bounds.

Revision: Table 5 and Appendix G.2 (Table 19; Figure 8) evaluate same-specification renewal; Table 7 and Appendix G.5 test adaptation, and Table 21 adds the strong hybrid.

Reviewer_gKQvOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VERA, a self-training framework for LLM agents that design analog circuits (OTAs, inverter chains, and comparators) in ngspice with the SKY130 PDK. The motivation is that the base model (Qwen3.5-9B) solves none of the held-out complete-design tasks, so there are no successes to learn from. VERA starts from three human-built verified circuits, one per family, and creates easier “recovery” tasks by removing structure, injecting wrong connections, or detuning parameters. Each task is certified to have a valid recovery path within the agent’s edit interface and budget. The learner attempts these tasks, and only selected actions from successful rollouts become LoRA SFT targets: actions that pass the task, repair function, or improve a violated spec without worsening others. After each update, the improved learner searches for circuits meeting stronger specifications, and verified results seed a fresh task bank for the next round. On 18 held-out complete-design tasks, success rises from 0/180 to 77/180 (42.8%) after one update, compared with 47/180 for SFT on all actions from the same successful trajectories. It reaches 123/180 (68.3%) after three updates, above fixed-bank (97/180) and frozen-search (103/180) controls. Gains also appear on PhyChip-v2 and a 20-task AnalogCoder-Pro subset.

Critical strengths
  1. A sensible solution to the zero-success problem. When the base model never succeeds, standard self-training (STaR/ReST-style) has nothing to learn from. Turning verified designs into partially broken versions with certified recovery paths is a clean way to create solvable practice. Keeping the recovery path private, so the learner finds its own solution, avoids simply imitating a reference. The jump from 0% to 42.8% after one update shows this works.

  2. Several useful controls. The paper compares recovery vs. complete-design data at equal data size and at equal collection budget. It compares full action selection against supervising all actions and against removing the endpoint check. It tests fixed-bank and frozen-search variants of the recursive loop and the effect of early target expansion. These go beyond what most papers in this area provide, and they support the main claims at the level of the overall pipeline.

Critical weaknesses
  1. The contribution of each component inside the pipeline isn’t shown. The ablations remove whole stages, but not the individual pieces within them: a. Action selection. It is ablated only as a whole (selection vs. all actions vs. no endpoint check). We don’t know whether Pass, Repair, or Progress actions matter most, or whether, for example, Progress actions alone drive the gain. b. Task types and reduction operators. There are four task types (construction, repair, parameter optimization, multi-step) and several operators, with no ablation of which ones help. Would parameter-detuning tasks alone be enough? c. Certification. The paper emphasizes “certified” recovery paths, but never compares against uncertified perturbations. Since only successful rollouts are used for training anyway, it’s unclear whether certification adds anything beyond filtering out unsolvable tasks earlier. d. Adaptive task splitting after bootstrap (weakening or splitting hard tasks) is also not ablated. For the recursive loop, full recursion collects more rollouts (1,566 vs. 1,404) and larger action pools (1,650 vs. 1,480) than fixed-bank recollection. So part of its advantage may come from more data rather than harder tasks. A data-matched comparison would make the “harder tasks drive further improvement” claim more convincing.

  2. The evaluation is narrow and close to the training distribution. The held-out set has only 18 tasks across 3 circuit families, and all six tasks in a family start from the same weak netlist. The training seeds come from the same three families. The authors acknowledge that training and evaluation differ in parameters and specs while “topological similarity may remain.” The complete-design tasks (start from a weak circuit, fix it to meet specs) are structurally similar to the recovery tasks, so it’s unclear how far the gains generalize beyond these three families.

Questions for authors
  1. Can you ablate the action-selection categories separately (Pass only, Repair only, Progress only)? Which contributes most?
  2. Which task types or reduction operators drive the gains? What happens if you train only on parameter-optimization tasks, or remove multi-step tasks?
  3. How does training on uncertified perturbations, keeping only successful rollouts, compare to certified tasks?
  4. Can you compare full recursion and fixed-bank recollection at matched action-pool size or rollout count?
  5. How similar are the held-out tasks to the training seeds and recovery tasks, e.g., in topology or starting netlist? How does performance change on circuit families not used for seeds?
3 comments
AuthorsOpenReview ↗

Response to gKQv (1/2): Action Selection, Task Mixtures, and Certification

We thank the reviewer. The revision extends the original controls with category, task-mixture and certification ablations, matched budgets, and topology/family transfer tests.

1. Action categories and matched supervision volume

Q1. Which action categories drive the gains, and does selection help at matched supervision volume? (W1a; Question 1)

The original ablation shows better learning from the same successful trajectories: full selection reaches 43/60, all-action supervision 27/60, and removing endpoint verification 31/60 (Table 15; Figure 5(c)). §3.3 defines Pass/Repair/Progress, illustrated by the inverter trajectory in Figure 3.

We added single-category training and a uniform-turn control matched in examples, tokens and optimizer steps, using 5 paired seeds. Full selection beats every single category and gains 6.3 pp over uniform sampling (95% CI [4.1, 8.0]).

Supervision ruleDownstream success rate
Pass only22.3 ± 3.4%
Repair only32.2 ± 3.7%
Progress only29.6 ± 4.4%
Uniform-turn matched36.4 ± 2.4%
Full selection (paired reference)42.8 ± 3.7%

Each single-category condition trains on that category alone:

ItemCount and sampling rule
Original successful trajectories213 trajectories, each with 1 terminal Pass (213 total before deduplication and the training/validation split)
Single-category training list1,080 entries sampled with replacement from each category’s training partition, excluding validation examples
Full / Uniform1,080 distinct training examples each, matched in training presentations, tokens, and optimizer steps
Original full-selection split1,200 total / 1,080 training / 120 validation (90:10 sample-level split)

Repeated runs use lineage-grouped splits (Appendix A.3). Pass remains the task-completing action.

Revision: §3.3 defines the categories; Table 7 and Appendix G.5 (Table 23, Panel A; Figure 10) report category and matched-uniform ablations; Appendix A.3 specifies data splits.


2. Task mixtures and reduction operators

Q2. Are parameter-detuning tasks sufficient, and what do structural and multi-step tasks contribute? (W1b; Question 2)

All four task types yield successful experience: construction/modification 44/108, diagnosis/repair 52/144, parameter optimization 46/108, and multi-step design 71/108 in the original run (Table 2).

The added ablations use equal collection and training quotas across 5 paired seeds. The full mixture outperforms parameter detuning alone and removal of multi-step tasks, demonstrating the value of combined structural, repair and parameter practice.

SettingDownstream success rateCollection successes (representative run)
Parameter detuning only33.1 ± 3.9%199/468
Without multi-step tasks29.4 ± 4.0%184/468
Full task mixture (full-method reference)42.8 ± 3.7%216/468

Collection counts are from a representative repeated run, separate from Table 2.

Revision: Table 7 and Appendix G.5 (Table 23, Panel B; Figure 10) report task-mixture ablations; Appendix C defines and illustrates the four task types.


3. Certification alongside successful-rollout filtering

Q3. Does certification add value when training already retains only successful rollouts? (W1c; Question 3)

Certification establishes a feasible path within the edit interface and budget; the learner solves the task independently with the path hidden. The original 45.5% collection success measures learner performance (§§3.2, 4.4).

We compare admission with and without legal replay before collection, fixing the proposal source, public interface and successful-rollout filter; paths stay hidden in both groups. Under the same 4,400-call cap, certification adds 4.6 pp (95% CI [3.2, 5.9]), with fewer admitted tasks and selected actions per thousand calls.

MetricUncertifiedCertified
Success rate55.9 ± 2.7%60.5 ± 4.2%
Independent replay validity61.3%96.9%
Tasks admitted out of 96 proposals7948
Selected actions per 1,000 operational calls216 ± 27188 ± 24

Actual SFT quotas appear below Table 22.

Revision: §3.2 specifies certification; Table 22 and Figure 9 report the common-budget comparison, replay validity, task admission and SFT quotas.

AuthorsOpenReview ↗

Response to gKQv (2/2): Adaptation, Budget Controls, and Train-Test Separation

4. Task splitting and perturbation weakening

Q4. What do adaptive task splitting and perturbation weakening contribute? (W1d)

After bootstrap, learner outcomes guide adaptation through splitting hard combined tasks into certified subtasks or weakening individual perturbations (§3.2; Appendix B.3, Algorithm 1).

We disabled both operations at the third update while retaining updated-model search and identical joint collection/SFT settings. Success falls by 4.8 percentage points, establishing their contribution to the recursive loop (Table 7, p. 10, 5 paired seeds).

SettingThird-update success rate
No adaptation (splitting and weakening disabled)64.0 ± 2.8%
Full (reference)68.8 ± 2.3%
Difference (no adaptation − full)−4.8 pp, 95% CI [−5.4, −4.1] (Appendix G.5, pp. 32–33)

Revision: §3.2 and Appendix B.3 (Algorithm 1) specify adaptation; Table 7 and Appendix G.5 report the control with splitting and weakening disabled.


5. Recursion with matched rollouts and training

Q5. Does full recursion still outperform a fixed bank when rollouts and training resources are matched? (W1d; Question 4)

The paper reports branch-specific resource use and bootstrap budget controls. Section 4.5 (p. 8) describes the three branches' different collection volumes. Table 18 (p. 28) lists cumulative rollouts of 1,566 / 1,404 / 1,446 and action pools of 1,650 / 1,480 / 1,410. Table 14 (p. 26) matches data volume and collection attempts separately at bootstrap.

The primary recursive comparison now uses common quotas. Joint matching fixes 468 collection attempts per update, actions, tokens and optimizer steps. The common verifier cap also includes search and certification. The original results remain for reference. Full VERA leads at updates 2 and 3 under both settings (Table 5 and Table 19, 8 seeds).

SettingFixed bankSame-spec renewalFull
Update 2 (Table 19, p. 29)50.9 ± 3.1%55.8 ± 3.9%60.8 ± 3.3%
Update 3, fixed collection/SFT quotas (Table 5)54.4 ± 2.6%60.3 ± 3.6%68.1 ± 3.1%
Update 3, common verifier cap (Table 5)52.9 ± 3.8%58.6 ± 3.3%66.3 ± 3.6%

Same-specification renewal updates designs and exercises while retaining the original specification bounds. Figure 8 (p. 30) shows the training trajectory.

Revision: Table 5 and Appendix G.2 (Tables 19–20; Figure 8) report fixed collection/SFT quotas, common verifier caps and results across updates.


6. Train–test similarity and transfer

Q6. How similar are test tasks to training circuits, and do gains transfer to unseen topologies and families? (W2; Question 5)

The 18 evaluation tasks are excluded from training, task construction and checkpoint selection. They use different instances, parameters and specifications within the training families, with possible topological similarity (Appendix A.2, p. 18). All 6 tasks per family share one D01 start (Appendix A.5 and Table 9, pp. 19–20). Tables 1 and 12 also test transfer on two external benchmarks.

The revision identifies the original 18 tasks as specification tests from shared starts and adds independent weak starts for new-topology and new-family tests. All evaluation starts undergo graph and lineage checks. None of the 6 new topology groups matches a training circuit by isomorphism or ancestry, and full VERA still beats the fixed bank on both suites (Table 6, p. 9, 8 seeds).

Evaluation setFixed bankFull
New topologies47.9 ± 2.5%59.7 ± 2.2%
New families42.5 ± 2.9%53.9 ± 3.5%

Table 24 (p. 34) includes base and first-update results. The audit covers initial and renewed seeds, exercise starts, accepted endpoints and intermediate netlists in SFT contexts (Appendix G.6). We remove device names, net names and numerical sizes, retain device types, terminal roles and connectivity, then check isomorphism and ancestry. The 6 test groups have the following maximum local similarities.

Topology group123456
Maximum local similarity0.40.50.30.50.30.6

Figure 11 (p. 34) compares each group's success and similarity. Local similarity describes structural proximity. Training overlap is determined through family/topology splits and lineage checks.

Revision: Appendices A.2 and A.5 clarify the original shared starts; Table 6 and Appendix G.6 (Table 24; Figure 11) add independent transfer tests and graph/lineage audits.

Reviewer_gKQvOpenReview ↗

Post-rebuttal assessment

Thank you for the detailed rebuttal and additional experiments. The new component ablations, matched-budget comparisons, and evaluations on unseen topologies and circuit families address most of my concerns. These additions strengthen the empirical support for the method, and I am increasing my rating from 2 (weak rejection) to 3 (weak acceptance).

Reviewer_c8UJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VERA self-trains circuit-design agents by perturbing verified seed circuits into recovery tasks with certified paths, fine-tuning on verified actions from successful trajectories, and using the updated model to find stronger-spec circuits that seed new tasks.

Critical strengths
  1. Tackles a real bootstrapping problem. Certified recovery tasks give a learner with zero complete-design successes usable training signal.
  2. Supports selective supervision. On identical successful trajectories, selected-action SFT beats all-action SFT.
Critical weaknesses
  1. The methodological novelty is not clearly established. VERA draws on reverse curricula, verification-based self-training, action filtering, and adaptive task generation. Certified recovery paths, the Pass/Repair/Progress selection rules, and seed renewal under stronger specifications are concrete design choices, but the paper does not sufficiently clarify the methodological advance beyond adapting and combining these existing ideas for circuit design. The ablations support the usefulness of individual components within VERA, but comparisons with closely related curriculum or self-training methods are needed to better establish its contribution.

  2. The additional benefit of recursive task renewal remains confounded, and evidence for broad generalization is limited. All control conditions improve beyond the first update: fixed-bank recollection rises from 42.8% to 53.9%, frozen search reaches 57.2%, and full recursion reaches 68.3%. The fixed-bank result shows that continued collection and SFT already yield substantial gains without task renewal. Although full recursion achieves the highest success rate, it also uses more collection rollouts and produces larger action pools, leaving additional data and computation as possible contributors to its advantage. Budget-matched comparisons and repeated training runs are needed to establish the robustness and specific contribution of recursive task renewal. Generalization evidence is also limited: bootstrap data originate from only three seed circuits, while internal evaluation covers 18 tasks within the same three circuit families. External evaluations provide encouraging evidence, but testing on unseen seeds and topologies would better substantiate broader transfer.

Questions for authors
  1. Novelty. What mechanism separates VERA from a combination of reverse curriculum construction, verification-based self-training, and adaptive task generation? How do recovery-path certification and Pass/Repair/Progress selection substantively extend the closest existing methods?

  2. Added benefit of recursion. Fixed-bank recollection already raises success from 42.8% to 53.9%, against 68.3% for full recursion. Were collection attempts, SFT tokens, optimizer updates, and total verifier calls (including task construction and certification) matched across conditions? If not, what evidence separates the effect of task renewal from that of extra data and compute? Was variance across independent training runs measured?

  3. Generalization to other circuit types. Bootstrap data come from three seed circuits, and internal evaluation uses the same three families. What evidence shows generalization beyond them? Do the external evaluations show gains on circuit families or topologies unseen in training, and can those results be reported separately?

4.SAFTAC has demonstrated that simulation-grounded fine-tuning can substantially improve analog circuit generation and correction. How does VERA compare with SAFTAC under a comparable evaluation setting?

2 comments
AuthorsOpenReview ↗

Response to c8UJ (1/2): Method Contributions and Budget-Matched Gains

We thank the reviewer for the feedback on novelty, recursive gains, transfer and SAFTAC. The paper addresses the first three through full-method comparisons, the V/H experiment, frozen search and external benchmarks. The revision adds repeated budget-matched experiments, individual mechanism controls, and new-topology and new-family tests.

1. Method contribution and component interactions

Q1. What methodological advance does VERA provide beyond combining existing curriculum and self-training ideas? (W1; Question 1)

VERA connects verifiable recovery tasks, local action supervision and seed renewal at tighter specifications to keep producing training signal when complete designs rarely succeed. The paper demonstrates their interaction in three ways. On the same successful trajectories, action selection raises success from 47/180 to 77/180 (Table 1). The measured inverter example supervises only the first improving edit and final completing edit (Figure 3, p. 5). Early target expansion lowers first-update success from 42.8% to 35.0%, supporting recovery practice before expansion (V/H experiment, Table 4, p. 9). Freezing the initial model for target search gives 103/180 at update 3 versus 123/180 for full VERA, supporting search by the updated learner for stronger-specification seeds (Table 17, p. 27).

We extended these mostly single-run controls with repeated mechanism ablations and direct comparisons against ReST-EM and a strong reverse-start hybrid. All three mechanisms contribute positively in their controls, and full VERA beats the strong combination of existing mechanisms (Table 21, p. 31, third update, 5 paired seeds).

MethodSuccess rate
ReST-EM47.4 ± 3.2%
Strong hybrid (certification, selection, adaptation, same-spec renewal)64.0 ± 2.2%
Full68.8 ± 2.3%
MechanismControl comparisonResult
Target expansion (seeds meeting tighter specifications)Full vs. same-spec (Tables 5 and 19)68.1 ± 3.1% vs. 60.3 ± 3.6%, paired +7.8 pp
Local action selectionFull vs. uniform (Tables 7 and 23)+6.3 pp, 95% CI [4.1, 8.0]
CertificationCertified vs. uncertified (Table 22)60.5 ± 4.2% vs 55.9 ± 2.7%

Revision: §§3.2–3.3 specify certification and action selection; Tables 5, 7 and 21–23 provide repeated mechanism controls and comparisons with related self-training methods.


2. Recursive gains and training variance at matched budgets

Q2. Do recursive gains persist under matched data and compute budgets and across training seeds? (W2; Question 2)

Continued collection and training help even on a fixed bank, which rises from 77/180 to 97/180 (Table 17). The paper reports 1,566 cumulative rollouts for full VERA versus 1,404 for the fixed bank, excluding construction and certification (Section 4.5, Table 18).

To compare recursive task supply at matched resources, we added two budget controls, each repeated across 8 independent training seeds. One fixes joint collection/SFT quotas. The other caps cumulative verifier calls, including search, failed proposals, construction, certification, calibration and collection. At matched resources, full VERA remains about 13 percentage points ahead of the fixed bank, consistently across seeds (Table 5, p. 9, Appendix G.1, p. 28).

Task supplyFixed collection/SFT quotasCommon verifier cap
Fixed bank54.4 ± 2.6%52.9 ± 3.8%
Same-spec renewal60.3 ± 3.6%58.6 ± 3.3%
Full recursion68.1 ± 3.1%66.3 ± 3.6%
ComparisonPaired gain95% CI
Full vs. fixed bank (Joint)+13.8 pp[12.1, 15.4]
Full vs. same-spec renewal (Joint)+7.8 pp[6.4, 9.2]

Actual calls are also similar under both settings (Table 20, p. 31, third update).

Settingcalls
Joint Full11,184 ± 216
Joint Fixed11,180 ± 240

The fixed bank also recollects with the updated learner and continues training. Appendix G.2 and Table 19 (p. 29) give quotas, and Figure 8 (p. 30) shows performance by update and cumulative cap.

Revision: Table 5 and Appendices G.1–G.2 (Tables 19–20; Figure 8) report matched budgets, independent training seeds, paired statistics and full verifier costs.

AuthorsOpenReview ↗

Response to c8UJ (2/2): Cross-Family Transfer and Comparison with SAFTAC

3. Transfer and external results by training overlap

Q3. Do gains extend to unseen circuit families and topologies, and how does external transfer depend on training overlap? (W2; Question 3)

The paper reports external transfer under separate execution and performance criteria. After one update, PhyChip L1 execution rises from 8/84 to 72/84 and L2/L3 performance from 0/66 to 11/66 (Table 12, p. 25, Section 5, p. 10). AnalogCoder-Pro rises from 30/60 to 47/60.

We directly evaluated existing checkpoints on 6 new topology groups with independent weak starts (18 tasks) and three held-out families, VREF, LDO and FLT (9 tasks). Full VERA beats the fixed bank by about 11–12 percentage points on both suites (Table 6, p. 9, 8 seeds).

Evaluation setFixed bankFull
New topologies (18 tasks)47.9 ± 2.5%59.7 ± 2.2%
New families (9 tasks)42.5 ± 2.9%53.9 ± 3.5%

Table 24 (p. 34) includes base and first-update results, and Figure 11 (p. 34) shows topology groups and similarities. The external unseen-family subsets contain 6/3/0 PhyChip L1/L2/L3 tasks and 4 AnalogCoder-Pro tasks. For a dedicated cross-family evaluation, we use the new 9-task suite, evaluated across 8 independent training seeds and checked for isomorphism and shared ancestry with training lineages.

Revision: Table 6 and Appendix G.6 (Table 24; Figure 11) add dedicated transfer tests and the overlap audit; Appendix D (Table 12) separates external evaluation criteria.


4. Comparison with SAFTAC

Q4. How does VERA compare with SAFTAC under a comparable evaluation setting? (Question 4)

VERA and SAFTAC differ in how training data are produced and how the learner is updated. Section 2 discusses simulation-based circuit agents and training. Our original controls use common tasks, verifier and interaction budgets (Tables 1, 15 and 17), with ReST-EM and the reverse-start hybrid added in Table 21.

SAFTAC uses curated simulation-task data, three-stage SFT and simulation-feedback reinforcement fine-tuning, reporting single-pass and feedback-based two-pass generation. The main distinction is data production. VERA generates further training signal from a few verified seeds when complete designs rarely succeed. SAFTAC, Sections 4–5

ItemSAFTACVERA
Training dataCurated simulation-task dataRecovery experience generated by the learner from a small set of verified seeds
Training procedureThree-stage SFT and reinforcement fine-tuning with simulation feedbackA loop of recovery-task construction, local action supervision, and seed and difficulty updates
Main backboneQwen3-8BQwen3.5-9B
Evaluation protocolSingle-pass generation and two-pass generation with feedbackRepeated rollouts under a fixed verifier-call budget

Revision: §2 provides the related-work positioning; Tables 1, 15 and 17 give controlled within-framework comparisons, with ReST-EM and the reverse-start hybrid added in Table 21.

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

The authors introduce VERA (Verifiable Escalation and Reduction for Self-Training Agents), a framework for improving LLM-based circuit design agents by addressing the scarcity of successful training trajectories. VERA generates progressively easier recovery tasks from verified circuit designs through controlled modifications to circuit topology, connectivity, and parameters, with certified recovery paths ensuring task feasibility. It also employs action-level supervision to fine-tune models using verified progress, repair, and completion actions, iteratively building more challenging training curricula from successful designs. The authors evaluate VERA using Qwen3.5-9B, ngspice, and SKY130 across operational transconductance amplifiers (OTA), inverter chains (INV), and comparators (CMP). On 18 held-out complete-design tasks, success improves from 0% for the baseline to 42.8% after one update and 68.3% after three updates, outperforming success-trajectory SFT and fixed-curriculum baselines.

Critical strengths
  • The paper introduces a useful way to construct recovery tasks from verified circuit designs, reducing the amount of work the learner must complete while preserving the original specifications.
  • Action-level supervision is a strong contribution because it selects verified progress, repair, and completion actions rather than treating an entire successful trajectory as equally useful.
  • The recursive training loop is well motivated: the improved learner discovers stronger specifications, which are then used to generate new recovery tasks.
  • The results are encouraging, with held-out complete-design success showing significant jumps.
Critical weaknesses
  • The evaluation is limited to a set of seed circuits spanning three circuit families and simulation conditions. So, it is unclear how well the approach generalizes to substantially different topologies and design tasks.
  • The paper does not report variance across independent training runs, or statistical significance tests for its key results, which makes the the reliability of the reported improvements stronger.
  • Though the paper includes component ablations, it doesn't directly compare against existing self-improvement/curriculum-learning methods. Also, the experiments don't isolate the contribution of certified recovery paths from the benefits of constructing more accessible tasks.
Questions for authors
  1. Can VERA generalize to unseen circuit topologies, rather than relying on one seed design per circuit family?
  2. Does full recursion still outperform fixed-bank training when both use the same number of rollouts, selected actions, and training compute?
2 comments
AuthorsOpenReview ↗

Response to PMZq (1/2): Generalization and Training Stability

We thank the reviewer for the feedback. The paper provides initial evidence on circuit coverage, training stability and prior methods through external transfer, per-task results and controls. The revision adds new topologies and families, 8 independent training seeds, and comparisons with ReST-EM and a strong reverse-start hybrid.

1. Unseen topologies and circuit families

Q1. Does VERA generalize to unseen starting topologies and circuit families beyond the three seeds? (W1; Question 1)

The paper already shows transfer beyond its training seeds. Alongside 18 internal tasks, we evaluated PhyChip-v2 (50 tasks) and AnalogCoder-Pro (20 tasks), whose circuits and tasks do not come from our seeds. After one update, successes rose from 8/150 to 83/150 and from 30/60 to 47/60, respectively (Table 1, p. 6). PhyChip execution and performance are reported separately (Table 12, p. 25). Appendix A.2 (p. 18) also notes possible topological similarity within the shared internal circuit families.

To directly test changes in starting topology and family, we added 6 new OTA/INV/CMP topology groups (18 tasks) and three families outside the seed bank, VREF, LDO and FLT (9 tasks). We evaluated existing checkpoints, with identical test tasks across training branches. Full VERA beats the fixed bank on both suites, extending its gains beyond the original shared starts (8 independent training seeds, Table 6, p. 9, and Table 24, p. 34).

Evaluation setBaseUpdate 1Fixed bankFull
New topologies (6 groups, 18 tasks)0.0%36.2 ± 3.6%47.9 ± 2.5%59.7 ± 2.2%
New families (VREF / LDO / FLT, 9 tasks)0.0%30.8 ± 3.9%42.5 ± 2.9%53.9 ± 3.5%

The paired new-topology gain is 11.8 percentage points, 95% CI [10.5, 13.1] (Appendix G.1, p. 28). Figure 11 (p. 34) shows each group's performance and local similarity to the training archive. All tests use the paper's ngspice simulator, SKY130 process and test conditions.

Revision: Table 6 and Appendix G.6 (Table 24; Figure 11) add topology/family transfer tests and the training-overlap audit; Appendix G.1 reports paired uncertainty.


2. Training stability and significance

Q2. Are the gains stable across independent training runs and statistically significant? (W2)

The original 77/180 versus 47/180 gain over successful-trajectory SFT was measured under a prespecified evaluation protocol. Appendix A.3 (p. 18) specifies one training run per condition (seed 42), a checkpoint at the fixed training length without validation-based selection, and 10 rollouts on each of 18 tasks. Results are averaged within tasks, then equally across tasks, with per-task results in Table 10 (p. 21).

We repeated the primary comparisons with 8 training seeds and the component and prior-method comparisons with 5 of them. Differences are paired within seed, with 10,000 bootstrap resamples of the training seeds. The four primary comparisons were prespecified and Holm-corrected (Appendix G.1, p. 28). Full VERA beats both the fixed bank and same-specification renewal in all 8 runs, with paired intervals excluding 0 (Table 5, p. 9, third update with joint collection/SFT matching).

SettingSuccess rate
Fixed bank54.4 ± 2.6%
Same-specification renewal (same-spec)60.3 ± 3.6%
Full recursion68.1 ± 3.1%
ComparisonPaired gain95% CI
Full vs. fixed bank+13.8 pp[12.1, 15.4]
Full vs. same-spec renewal+7.8 pp[6.4, 9.2]

All four primary comparisons have a two-sided exact sign-randomization p of 0.0078125, the minimum possible with 8 seeds, and Holm-adjusted p of 0.03125. The original 123/180 (68.3%) remains separately labeled as a single-run result.

Revision: Table 5 and Appendix G.1 report repeated training, paired confidence intervals and corrected significance tests; Appendix A.3 specifies aggregation and data splits.

AuthorsOpenReview ↗

Response to PMZq (2/2): Curriculum Baselines, Certification, and Budget Controls

3. Curriculum baselines and certification

Q3. Does VERA outperform related curriculum methods, and does recovery-path certification add value? (W3)

The paper already includes a self-training control and supervision ablations. Successful-trajectory SFT follows the STaR/ReST-EM approach of retaining successful trajectories, reaching 47/180 versus VERA's first-update 77/180 on the same trajectories (Table 1). Table 15 (p. 27) and Figure 5(c) compare all-action supervision and removal of endpoint verification. Section 2 discusses reverse curricula, Backplay and LADDER. Sections 3.2 (p. 4) and 4.4 (p. 7) explain that certification guarantees a legal recovery path, while the learner must still solve the task.

We added direct comparisons with ReST-EM and a strong reverse-start hybrid in the same training framework. The hybrid combines legal replay, action selection, adaptive difficulty and same-specification seed renewal. Full VERA outperforms both, gaining 4.8 percentage points over the hybrid, 95% CI [4.1, 5.4] (Table 21, p. 31, third update, 5 paired seeds).

MethodSuccess rate
ReST-EM47.4 ± 3.2%
Reverse-start hybrid (strong control)64.0 ± 2.2%
Full (reference)68.8 ± 2.3%

With the same proposal source and successful-rollout filter, we compared uncertified admission with full legal replay. Certification improves downstream success under a common 4,400-call cap, while admitting fewer tasks (Table 22 and Figure 9, p. 32).

MetricUncertifiedCertified (Full legal replay)
Success rate55.9 ± 2.7%60.5 ± 4.2%
Independent replay validity61.3%96.9%

Revision: Tables 21–22 and Figure 9 add curriculum baselines and the certification control; §3.2 specifies the feasible-path guarantee and private recovery record.


4. Matched collection, training and verifier budgets

Q4. Does full recursion still outperform fixed-bank training at matched collection, SFT and verifier budgets? (Question 2)

The paper reports resource use and bootstrap comparisons with matched data or collection budgets. Section 4.5 (p. 8) explains why the three recursive task supplies produce different collection volumes. Table 18 (p. 28) lists cumulative attempts of 1,566 / 1,404 / 1,446 and action pools of 1,650 / 1,480 / 1,410. At bootstrap, recovery tasks beat complete tasks with equal data volume (9/60 vs. 0/60) and equal collection attempts (12/60 vs. 0/60, Table 14, p. 26).

The recursive comparisons now use two budget controls. Joint matching gives branches within each seed the same first-update model, 468 collection attempts per update and identical SFT quotas. A common verifier cap also charges search, construction, certification and collection to the cumulative budget. Full VERA retains its lead under both controls, so extra collection or training volume does not explain the gain (Table 5, p. 9, third update, 8 seeds).

Task supplyFixed collection/SFT quotasCommon verifier cap
Fixed bank54.4 ± 2.6%52.9 ± 3.8%
Same-specification seed renewal60.3 ± 3.6%58.6 ± 3.3%
Full recursion68.1 ± 3.1%66.3 ± 3.6%

Each update uses 1,080 actions, 265k target tokens, 2.32M context tokens and 203 optimizer steps (Appendix G.2, p. 29). Figure 8 (p. 30) shows performance by update and cumulative cap, with costs in Table 20 (p. 31). Same-specification renewal also finds new designs, renews seeds and adjusts difficulty. Full VERA's further 7.8-point gain supports tightening the target specifications.

Revision: Table 5 and Appendix G.2 (Tables 19–20; Figure 8) report matched collection/SFT quotas, common verifier caps and stage-by-stage cost accounting.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Rebuttal and Revision Summary: Core Contribution and Further Validation

We thank all reviewers for their helpful feedback. The original submission already supports the paper's central contribution. The additional experiments strengthen and broaden that evidence while preserving the same method and conclusions. VERA learns from recovery tasks built from a few verified circuit seeds, then uses the improved learner to discover seeds meeting tighter specifications for further training.

1. The original paper already provides evidence relevant to the main concerns

The original experiments show that action selection improves over successful-trajectory SFT on the same trajectories (77/180 versus 47/180, Table 1), and that searching with the updated learner improves over frozen-model search (123/180 versus 103/180, Table 17). They also demonstrate external transfer: PhyChip performance successes rise from 0/66 to 11/66, and AnalogCoder-Pro successes from 30/60 to 47/60 (Tables 1 and 12).

The paper further includes curriculum-timing and supervision controls (Tables 4 and 15), bootstrap comparisons with matched data volume or collection attempts (Table 14), and explicit resource accounting (Table 18). The learning mechanism, recursive gains and transfer therefore already have empirical support in the original submission.


2. Additional experiments confirm these findings under stricter tests

  • Matched resources and independent replication. Across 8 training seeds, VERA reaches 68.1% versus 54.4% for fixed-bank training with matched collection and SFT quotas: +13.8 percentage points, 95% CI [12.1, 15.4]. The gain remains +13.4 points under a common simulator-call cap including search and certification. VERA beats both fixed-bank training and same-specification renewal in all 8 runs (Table 5; Appendix G.1).

  • Broader generalization. Evaluating existing checkpoints on 6 unseen topology groups (18 tasks) and 3 held-out families (9 tasks) gives 59.7% versus 47.9% and 53.9% versus 42.5%, respectively, against fixed-bank training across 8 seeds. The new topology groups have no isomorphic or ancestral match in the training archive (Tables 6 and 24).

  • Strong baselines and mechanism controls. Tightening specifications adds 7.8 points over same-specification renewal (Table 5). In 5-seed comparisons, VERA reaches 68.8%, versus 64.0% for the strong reverse-start hybrid and 47.4% for ReST-EM (Table 21). Action selection adds 6.3 points at matched training volume, and certification adds 4.6 points under a common call budget; task-mixture and adaptation ablations provide further support (Tables 7, 22 and 23).

These results consistently reinforce the original findings. We believe the original methodological contribution and empirical evidence, now supported by stricter controls, independent replication and broader evaluation, warrant acceptance.


3. Main changes in the revised manuscript

The main new results appear in Tables 5–7, with experimental protocols and detailed results in Appendix G (Tables 19–24 and Figures 8–11).

  • Budget matching and independent training runs. Table 5 reports results across eight training seeds under two controls. One matches collection and SFT quotas, and the other applies a common verifier-call cap that includes search, task construction, certification, and collection. Appendix G.1 reports paired confidence intervals and significance tests across training seeds. Appendix G.2 details the budgets and costs across training rounds.

  • Curriculum baselines. Table 5 adds a control that renews tasks and adapts their difficulty at the original specifications. This comparison measures what tightening the specifications adds beyond adaptive task renewal. Table 21 compares VERA with ReST-EM and a reverse-start hybrid in the same training framework.

  • Transfer to unseen circuits and overlap checks. Table 6 and Appendix G.6 add evaluations on 18 tasks from six new topology groups and nine tasks from three held-out circuit families. Table 24 and Figure 11 provide detailed results, and Appendix G.6 documents the checks for graph isomorphism and shared ancestry with training circuits.

  • Component ablations. Table 7 and Appendix G.5 compare action selection with uniform sampling at matched training volume, vary the task mixture, and test the effect of disabling task splitting and perturbation weakening. Table 22 and Figure 9 compare certified and uncertified task admission under the same verifier-call cap and successful-rollout filter.

  • Clarifications to the original evaluations. Appendices A.2 and A.5 explain the shared starting circuits used to test different specifications in the original internal suite. Appendix D clarifies how the pooled PhyChip-v2 results relate to the separate execution and performance results in Table 12 and identifies the 20 AnalogCoder-Pro tasks included in the evaluation.

AuthorsOpenReview ↗

Looking Forward to Your valuable Feedback — A Brief Follow-Up

Thank you again for taking the time to review our paper. If time permits, we’d be grateful for any feedback on our rebuttal.

With less than one day remaining in the discussion period, we’d be happy to address any remaining questions or concerns before it closes.

AuthorsOpenReview ↗

Response to Reviewer: Looking Forward to Your Feedback

Thank you again for taking the time to consider our responses and the revised manuscript. If time permits, we would welcome any feedback on points that still need clarification. We would be happy to address them or provide further analysis before the discussion period closes.

If you feel that our responses and revisions have addressed your concerns, we would be grateful if you would consider updating your overall score accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.