acceptodds
Under review as a conference paper at ICLR 2027

Decomposed Prompt Optimization For Reliable LLM-as-Judge

Abstract

Generative LLM-as-judge systems for safety guardrails, reward models, automatic evaluators, and niche domains are typically configured by a natural-language prompt - a flexible lever for making a weak judge stronger without modifying its weights. We study how far this lever goes for small frozen judges, and find that unconstrained prompt optimization i.e. rewriting the prompt as a single free-text block, buys surprisingly little. The missing ingredient in here is *decomposition*: a judge prompt splits into two functionally distinct components: *criterion* (what condition to detect), and *scoring schema* (how to map a judgment to a decision), and making this decomposition *explicit* to the optimizer is what makes automatic prompt optimization pay off. Across four judge models (B-B) and six tasks we find that (i) reflective optimization improves every judge, most where it is weakest (up to F1); (ii) treating the prompt as two explicit components beats optimizing it as one monolithic block on every judge (up to F1); and (iii) *which* component to search and in what order, is the most effective way to *spend* a compute budget: phasing converts extra search into the largest gains on the hardest structured tasks (up to F1) once initial search saturates. Applying this recipe to a real-world financial-compliance policy that fixed-taxonomy guards cannot express, a frozen B judge, optimized only through its prompt, improves F1 from to while baseline guards stall at -.

Primary area:
foundation or frontier models, including LLMs
Keywords:
prompt optimization, llm-as-judge, large language model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_thkpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper splits an LLM-judge prompt into a criterion (what to detect) and a scoring schema (how to decide), then uses GEPA to compare optimization regimes that edit these parts separately, jointly, or in phases. Across four frozen 8B–31B judges and twelve binary judging tasks, it reports that this decomposition beats optimizing the prompt as one block, and that it lifts an 8B judge on a financial-compliance benchmark from 0.70 to 0.92 F1.

Critical strengths
  1. Splitting the judge prompt into a criterion that defines what to judge and a scoring schema that defines how to decide is reasonable and intuitive, and the optimized result is human-readable, auditable rule text.
  2. Experiments show that GEPA-based prompt optimization clearly improves over zero-shot prompting, especially for weak judges and in the FinProof financial-compliance setting, where F1 rises from 0.70 to 0.92.
Critical weaknesses
  1. The method applies GEPA without modification, and its joint mode, which edits the criterion and the scoring schema in turn, is essentially GEPA's existing round-robin optimization over multiple prompt modules.
  2. Compared with plain GEPA on the whole prompt, decomposition adds only 0.003 and 0.004 F1 on gemma-4-31b and granite-guardian-4.1-8b, which is below the paper's own run-to-run noise, and at matched compute plain GEPA is even slightly better.
  3. The paper never tests the teacher models directly as judges, even though qwen3-next-80b-a3b activates only 3B parameters per token and could be a stronger judge at similar cost; the headline FinProof and ATBench results also use three times the search budget with no plain GEPA control at that budget.
  4. Since only GEPA is tested and the criterion-plus-scoring-schema format is the native prompt template of Granite Guardian, it is unclear whether the benefit of decomposition holds for other optimizers or judge prompts.
Questions for authors

See Weaknesses.

2 comments

Thank you for the review. Please find below the responses to the concerns raised in the review::

W1 - Existing GEPA mechanism : That is correct. Joint uses GEPA’s existing round-robin component updates [R1]. The blanket single-string characterization in §2 was not precise for GEPA and MIPRO and the revised related-work section corrects it. Our contribution is applying criterion–schema decomposition to prompt optimization for more reliable judgments from frozen, single-call models, with comparisons of component compatibility and optimization schedules.

W2 - Small gains and matched budgets :

  • Small differences - The +0.003/+0.004 Gemma-31B/Guardian differences do not establish an advantage. The larger Table 1 differences, +0.128/+0.043 on the two general Granite judges, compare joint against whole-prompt GEPA under a fixed teacher and nominal evaluation target. They are descriptive contrasts.
  • Matched budgets - We agree that the Guardian comparison at matched nominal targets (submitted Table 12) shows no overall joint advantage: results favor different modes on different tasks, and combined is higher on ATBench (0.809 versus 0.678) and AsseBench (0.684 versus 0.580). That table selects the teacher separately for each task and mode (on validation in the revision); it does not hold the teacher fixed or measure total compute cost.
  • New same-teacher runs - For Guardian, the revision compares joint and phased runs at nominal totals near 6,400 while holding the teacher fixed. Across three teachers, the structured-task aggregates for two- and three-phase schedules are within 0.01 F1 of joint, so we do not claim that phasing is better at equal compute.

W3 - Teachers and high-budget controls.

  • Teachers as judges : Granite-4.1-30B is already both a reflection teacher (§3.1) and a judge (Tables 1, 3, 6). The Qwen teacher's 80B total/3B activated parameters [R3] alone do not establish comparative serving cost.
  • ATBench headline : Guardian's 0.955 F1 (submitted Table 8) uses three-phase search with a nominal target of 1,600 judge evaluations per phase, totaling 3 × 1,600 = 4,800; actual calls may overshoot.
  • FinProof : On FinProof, at the same nominal 3,200-evaluation target, validation-selected joint reaches 0.83 versus 0.78 for whole-prompt GEPA. The two-phase result (0.870) uses a 3,200 target per phase.
  • What this shows. : The larger phased runs show attained performance at the stated targets, without establishing optimal compute allocation.

W4 - Template and optimizer generality : The criterion/schema headings are native to Guardian, whose released helper already accepts both components [R5]. Joint and combined start from the same rendered seed prompt and are instructed to keep the same section headers, so neither regime gets a template the other lacks. Table 1's general Granite and Gemma judges extend model coverage. Our conclusions are limited to the tested models, this template and GEPA.


Please let us know if any further clarification would be helpful.

Thanks for the authors' response. Since my concerns are not fully addressed, I decide to maintain my original rating.

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

This paper studies prompt optimization for frozen LLM-as-a-Judge models. The authors decompose the judge prompt into two components, a criterion and a scoring schema, and use GEPA to optimize them separately or jointly. Experiments across multiple judge models and evaluation tasks show that prompt optimization can substantially improve F1, especially for weaker models. The paper also analyzes component contributions, optimization budgets, and performance on a financial-compliance benchmark.

Critical strengths
  1. Improving the reliability of small LLM judges without fine-tuning is useful, especially for applications where models need to adapt to different evaluation criteria.

  2. The authors evaluate multiple judge models across 12 tasks and provide ablation studies on prompt components and optimization budgets, which help understand when the method works.

Critical weaknesses

(1) The proposed approach mainly applies GEPA to two predefined prompt components. It is unclear whether the improvements come from the criterion/schema decomposition itself or simply from optimizing smaller text spans. Comparisons with alternative decompositions or standard multi-component prompt optimization are needed.

(2) The proposed method is optimized using task-specific data, while the guardrail baselines are evaluated zero-shot. It would be more convincing to compare against a standard GEPA baseline with the same training data and compute budget.

(3) Table 3 reports the best results across different optimization modes, teachers, and budgets. It is unclear how these configurations were selected. Using test performance for this selection could lead to overly optimistic results, especially given the small test sets.

(4) Figure 1 is difficult to read. The handwritten text is messy and unclear, making it hard to follow the workflow. A cleaner figure with standard fonts would improve readability.

Questions for authors

See Weaknesses. If the authors can address my concerns, I would be willing to raise my score.

1 comment

Thank you for the recognizing the strengths of the paper and for the questions.

W1 - Semantic versus arbitrary decomposition : Joint uses GEPA’s existing round-robin component updates [R1]. The revised related-work section states that GEPA, MIPRO and TextGrad already expose structure to the optimizer. Our research question is narrower: how criterion and schema coordinates affect optimization of a single-call judge.

Table 4's test-selected exploratory summaries for Guardian motivate studying how the two components combine. On ATBench and tool-call faithfulness, the criterion transplant (optimized criterion with seed schema) has lower mean test F1 than the seed prompt, while using both optimized components has higher mean test F1 than either transplant. Arbitrary partitions could also interact, however. These results do not isolate semantic boundaries from smaller-span optimization, and we do not claim a uniquely effective partition.

W2 - Standard GEPA control :

  • Cross-judge study : This control already exists. "Combined" is whole-prompt GEPA with the same labeled splits, fixed teacher, and nominal judge-evaluation target as joint (§4; Tables 1–2). Joint–combined differences are +0.128/+0.043 on the two general Granite judges and +0.003/+0.004 on Gemma-31B and Guardian. These are descriptive contrasts.
  • FinProof : The revision adds a whole-prompt comparison on five label-stratified test splits (152 examples each) on which every system, guards included, is scored. On FinProof, at the same nominal 3,200-evaluation target, validation-selected joint reaches 0.83 versus 0.78 for whole-prompt GEPA. The two-phase result (0.870) uses a 3,200 target per phase.
  • Guard baselines : These stay zero-shot. ShieldGemma now receives a summary of the BFSI policy as its guideline (0.584); Llama Guard 3 (0.659) and WildGuard (0.624) use default taxonomies.

W3 Selection of "best" : The revision uses validation selection for the cross-judge best-regime summary, twelve-task headline results, matched-budget comparison, and FinProof. The guardian's twelve-task mean is 0.818, versus 0.827 for the submission's test-selected grid maxima (Table 3). Component transplants, the six-configuration ablation, and budget curves remain test-selected exploratory analyses. We withdraw the ≈0.008 selection-bias claim and report the two means descriptively. For Table 3, the teacher, mode and budget are now selected on validation for each scorer–task pair. Table 1's joint–combined contrast uses one fixed teacher and nominal evaluation target, and its best-minus-seed column now uses the validation-selected regime.

W4 - Figure readability. The revised PDF replaces Figure 1 with a revised diagram of the three-phase schedule. Each phase refines one component while the other is held fixed, and the selected components form one prompt for one judge call. Figure 2 now plots the seed, combined and joint values from Table 1, with both optimized bars at the same teacher and nominal target


Please let us know if you have any other questions. We would be happy to clarify any remaining points.

Reviewer_EMeYOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper argues that a judge prompt is not an undifferentiated string but decomposes into a criterion (what condition to detect) and a scoring schema (how to map a judgment onto a label), and that making this decomposition explicit to an automatic optimizer is what makes optimization pay off. Six modes are defined over the two coordinates: criterion-only, schema-only, combined (one free-text block, the monolithic baseline), joint (round-robin), and two- and three-phase block-coordinate schedules. All searches use GEPA with a separate reflection teacher swept over four models, optimize balanced accuracy on validation, and report test F1. A cross-judge study covers four frozen judges from 8B to 31B on six tasks with five seeds; a full suite extends to twelve datasets and six scorers. Reported findings: optimization lifts every judge and most where it is weakest (granite-4.1-8b from 0.144 to 0.583); the joint regime beats combined for every judge (+0.128, +0.043, +0.003, +0.004); phased schedules are indistinguishable from joint search at matched total compute but keep improving past the budget where joint saturates; a transplant table and an AUROC decomposition suggest the criterion carries discrimination while the schema moves the decision threshold. A case study on FinProof, an adversarial banking, finance and insurance compliance benchmark, reports a frozen 8B guardian improving from 0.70 to a five-seed mean of 0.92 F1 while ShieldGemma-9B, LlamaGuard-3-8B and WildGuard-7B score 0.67 to 0.71 zero-shot.

Critical strengths
  1. The compute-matched analysis of phasing, which argues against the paper's own tidy story. A per-budget table makes the modes line up crit, schema, joint, two-phase, three-phase, and the authors point out that a k-phase run at budget B spends kB evaluations. At matched total evaluations, joint at 3200 and two-phase at 1600 reach 0.732 and 0.735, indistinguishable. Reporting this rather than the flattering table is the right call.
Critical weaknesses
  1. The headline decomposition advantage is smaller than the paper's own measure of its instrument's noise, on half the judges. Section 4.1 declares effects significant against a "noise floor" of 0.0126 to 0.0144 F1, while the conclusion states that "re-running the same prompts shifts the guardian's F1 by 0.0268, and selection compounds it". Two different noise estimates differing by roughly a factor of two appear in one paper, and the smaller one is used to support the claims. Against the larger figure, the +0.003 and +0.004 gaps on gemma-4-31b and the guardian are an order of magnitude below noise, and the +0.043 on granite-4.1-30b is 1.6 times it before any adjustment for selection. No paired significance test appears anywhere. Test partitions hold at most 140 examples, so a 0.043 difference also sits inside the sampling error of the test set, and Table 3's "best" columns are maxima over mode, teacher and budget grids whose per-cell standard deviations reach 0.16 against a claimed selection bias of 0.008 whose derivation is never given.

  2. The one large effect is confounded with output-format repair, and the paper reports no parse rates. The +0.128 comes from granite-4.1-8b, whose zero-shot F1 in Table 3 is 0.000 on ragtruth, 0.000 on toolcall, 0.016 on off-topic and 0.043 on atbench. An F1 of exactly zero across whole datasets is the signature of a judge that never emits a parseable verdict, a failure the introduction itself raises by citing JSONSchemaBench. The scoring schema is the component that specifies output format, so the most economical explanation of the decomposition advantage on this judge is that an explicitly addressable schema coordinate gets it to emit a label at all. No parse rate is reported for any judge or regime, so this cannot be excluded, and if it is correct the paper measures a formatting fix rather than better judging.

  3. The abstract contradicts the results sections, and the applied case study misdescribes its baselines. The abstract claims explicit decomposition "beats optimizing it as one monolithic block on every judge" when two of four gaps are nulls the paper itself calls "a predicted null", and claims that "phasing converts extra search into the largest gains" when Section 5.2 is titled "the phasing improvement is mostly a compute artifact". Section 6 then states that the deployed guards "cannot be told the custom BFSI policy" and that "the gap is expressiveness, not scale". ShieldGemma conditions each decision on a written harm policy supplied in its prompt and can be given a custom policy; it was instead run zero-shot with its defaults, which is why all three guards report precision of exactly 1.000 with recall near 0.50. The comparison is between a judge optimized on the benchmark's own training split and three guards never shown the policy, on 368 examples, which establishes only that telling a model the policy helps.

Questions for authors
  1. What is the parseable-verdict rate for every judge under the seed prompt and under each regime? Restricted to examples where both the combined and joint prompts parse, does the granite-4.1-8b advantage survive and at what size?

  2. Which noise estimate is correct, 0.0126 to 0.0144 or 0.0268? Please report a paired significance test for joint against combined per judge, with test-set sizes stated, and derive the 0.008 selection-bias figure.

  3. On FinProof, please run ShieldGemma with the BFSI policy in its policy field, add a general judge given the policy zero-shot and few-shot, and add a classifier fine-tuned on the same training split. Which part of the gain is the decomposition rather than access to labelled data?

  4. Are the five seeds five partitions or five optimizer seeds on one partition? If the latter, test-set sampling error is unaccounted for in every reported standard deviation.

  5. The optimizer maximizes balanced accuracy but you report F1. Given that Table 5 attributes +0.069 F1 to re-thresholding alone, how much of the reported gain is threshold movement under a metric mismatch?

1 comment

Thank you again for the detailed review.

Q1/W2 - output-format repair : Granite-8B's highest Table 1 mean is criterion-only: F1 rises from 0.144 to 0.583 with schema text frozen; combined and joint reach 0.388 and 0.516. Its largest mean improvement in Table 1 therefore did not require editing the schema. Zero F1 alone does not diagnose parse failure: valid all-negative predictions also produce it. Criterion edits or component interactions could still repair formatting. The +0.128 joint–combined difference is a descriptive result.

Q2/W1 - noise & selection. :

  • Noise - The submission describes 0.0126–0.0144 as a run-to-run noise floor (section 4.1) and 0.0268 as the shift in Guardian's F1 from re-running the same prompts (Section 7). We withdraw the noise-floor significance argument and the "predicted null".
  • Test size - Cross-judge test partitions hold at most 140 examples per task.
  • Selection - The revision uses validation selection for the cross-judge best-regime summary, twelve-task headline results, matched-budget comparison, and FinProof. The guardian's twelve-task mean is 0.818, versus 0.827 for the submission's test-selected grid maxima (Table 3). Component transplants, the six-configuration ablation and budget curves remain test-selected exploratory analyses. We withdraw the ≈0.008 selection-bias claim and report the two means descriptively.

Q4 - seeds : The revised FinProof protocol reports a five-seed mean over five label-stratified test splits of 152 examples each (109 attacks, 43 benign). Within each cross-judge task and seed, scorers and regimes share the same partition; we do not read the reported standard deviations as test-sampling uncertainty.

W3 - abstract contradictions : The revised abstract summarizes the observed differences across judges; the introduction reports the four joint–combined contrasts. For Guardian, no regime has the highest mean F1 on every task at matched nominal targets (submitted Table 12). We present phased results as attained performance at the stated nominal budgets, without claiming optimal compute allocation.

Q3 / W3 - FinProof controls. :

  • New protocol - The revision re-runs FinProof on these five splits, scoring every system, guards included, on them. These results replace the submission's single-split numbers (0.700 → 0.915 on 368 examples).
  • Policy access for the guards - ShieldGemma takes a written policy [R2] and now receives a summary of the BFSI attack policy as its guideline (F1 0.584). Llama Guard 3's prompt utilities accept supplied category definitions [R4], but we ran it and WildGuard with default taxonomies (0.659 and 0.624). We have corrected the blanket claim that deployed guards cannot accept custom policies.
  • Results - The frozen guardian improves from 0.666 to 0.870 (two-phase, validation-selected).
  • Decomposition comparison - On FinProof, at the same nominal 3,200-evaluation target, validation-selected joint reaches 0.83 versus 0.78 for whole-prompt GEPA. The two-phase result (0.870) uses a 3,200 target per phase.
  • Guard comparison -The guards stay zero-shot while optimized rows use the labeled training split, so that comparison still mixes policy access with labeled-data adaptation. The tested operating points are not threshold-matched.

Q5 - Balanced accuracy versus F1. :

  • Objective - In our experiments, GEPA uses inverse-frequency-weighted correctness as its per-example reward, with class weights computed from the training split; its mean on that split equals balanced accuracy. Candidate selection within each run uses validation, and we report test F1; optimizing this reward need not optimize F1.
  • Threshold - Table 5's +0.069 is the mean test-F1 gain from re-thresholding Guardian's seed scores using a validation-tuned threshold, with the prompt unchanged. With both labels in the top-5 log-probabilities at the verdict position, scores use renormalized P(yes); with only yes, P(yes); with only no, 1−P(no).
  • Ranking - On Guardian, applying only the optimized criterion or only the optimized schema, with the other component at its seed, changes mean AUROC by +0.033 or +0.009, respectively. These separate interventions do not quantify how much of the total F1 gain comes from ranking versus threshold changes. An F1 improvement alone does not establish better ranking. Schema edits also raise ATBench AUROC by +0.067, so their effects are not limited to threshold shifts.

We would be glad to address any additional questions you may have.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common Response

We sincerely thank the reviewers for their reviews. In response to the reviews, we have uploaded a revised paper with the following changes:

  • Validation-selected headline results.: The revision uses validation selection for the cross-judge best-regime summary, twelve-task headline results, matched-budget comparison and FinProof. The guardian's 12-task mean is 0.818, versus 0.827 for the submission's test-selected grid maxima (Table 3). Component transplants, the six-configuration ablation and budget curves remain test-selected exploratory analyses. We withdraw the ≈0.008 selection-bias claim and report the two means descriptively.
  • A re-run FinProof protocol. FinProof is re-run on five label-stratified test splits of the public benchmark (152 examples each: 109 attacks, 43 benign). Every system, guards included, is scored on the same splits, and ShieldGemma receives a written summary of the BFSI attack policy as its guideline. The frozen guardian improves from 0.666 to 0.870. On FinProof, at the same nominal 3200-evaluation target, validation-selected joint reaches 0.83 versus 0.78 for whole-prompt GEPA. The two-phase result (0.870) uses a 3200 target per phase. These results replace the submission's single-split FinProof numbers.
  • Higher-budget joint runs for Guardian : at nominal totals near 6400 under three teachers. The structured-task aggregate differences between phased schedules and joint are within 0.01 F1 in these same-teacher comparisons.
  • Robustness and scope.
    • Group-aware re-splits for the two tasks whose example-level splits shared scenarios or prompts across partitions.
    • We have added a second custom-policy case study (multilingual tax advice) that uses synthetic optimization data and synthetic-test configuration selection. Pooled tax F1 increases from 0.791 to 0.932 ± 0.030 (mean ± SD across five prompt deployments). The generator design was iterated using benchmark evaluations, so we present observed transfer under that development process.

These revisions strengthen the paper by separating validation-selected headline results from test-selected exploration, improving the comparability of the FinProof systems through shared splits and explicit policy access, and clarifying schedule comparisons at matched nominal targets. The group-aware re-splits test sensitivity to data overlap, while the multilingual tax case study extends the evidence for the complete prompt-optimization procedure to a second custom-policy setting.

AuthorsOpenReview ↗

Follow-up for all reviewers: additional controls and details

Due to unavoidable circumstances, we did not have access to our compute resources and full experiment records when we wrote our previous response. Having regained access, we are addressing some of the additional reviewers' concerns. We address them below, with new controls run on the same five-split protocol used in our previous response and revised paper.

  • Seeds (EMeY Q4) - Within a seed all scorers and regimes share one partition, as we stated. Across various seeds, each seed draws a different stratified train/val/test partition and also sets GEPA's random state. The reported SDs therefore reflect repartitioning, optimization and inference variation. For the fixed seed prompt, the per-task SD averaged over the six tasks is 0.022 to 0.044 F1 across judges.

  • Formatting (EMeY Q1/W2) - In the cross-judge study, Guardian parses 100% of verdicts in every regime, and Gemma-31B at least 99.9% on average. Granite-8B's zero-shot zero-F1 cells come from valid "no" answers: they contain no parsed "yes" at all, and on RAGTruth every output parses. Granite-8B's joint prompt parses less often than its combined prompt (94.9% vs 97.8% on average). Its joint − combined gap grows from +0.128 to +0.150 F1 when each prompt's own unparsed outputs are excluded. On the examples both prompts parse, the aggregate counts bound the gap between +0.124 and +0.195, so it is not format repair.

  • Ranking (EMeY Q5) - Across the 12 Guardian tasks (validation-selected configurations), mean Δ\Delta AUROC from seed to optimized prompt is +0.043. The gains are concentrated on ATBench (+0.35) and AsseBench (+0.11) where optimization improves discrimination itself, not only the decision threshold.

  • FinProof controls (EMeY Q3, kJcJ W2, thkp W3). These use the same five splits (152 examples each, 72% attacks). Always answering "attack" scores 0.835 F1 but 0.5 balanced accuracy (BA), so we report both. The hand-written policy describes all 13 attack categories. We wrote it from the dataset's category codes and example items, so it carries more task knowledge than the one-line seed.

systemF1BArecallFPR
Guardian 8B, one-line seed0.6660.7480.5010.005
Guardian 8B, hand-written 13-category policy0.7780.8130.6400.014
Guardian 8B, same policy + 16 labeled examples0.7930.8170.6660.033
Guardian 8B, GEPA from the one-line seed (two-phase, 3,200 per phase, val-selected)0.8700.8690.7890.051
Llama-Guard-3-8B, the 13 categories as custom categories0.7490.8000.6000.000
ShieldGemma-9B, the 13-category policy as guideline0.6900.7600.5300.009
Qwen3-Next-80B (teacher) as judge, zero-shot0.7140.7780.5560.000
Llama-3.3-70B (teacher) as judge, zero-shot0.7760.8120.6390.014

These guard rows complement the runs in our previous response and the revised paper, which used Llama Guard's default taxonomy (0.659) and a ShieldGemma policy summary covering 7 of the 13 categories (0.584).

Starting from a one-sentence seed, GEPA lets the frozen Guardian outperform several alternatives on both F1 and BA:

  • the same model given the hand-written full policy, with or without 16 examples;
  • both policy-conditioned guards;
  • the teachers used directly as judges. The Guardian's margin is +0.156 F1 over Qwen3-Next and +0.094 over Llama-3.3-70B (95% paired t intervals across the five splits: [+0.101, +0.211] and [+0.042, +0.146]); every split favors the Guardian. Granite-4.1-30B, also a teacher, scores 0.615 zero-shot on FinProof. Nemotron was not available as a judge endpoint.

For the Guardian, automatic optimization of a one-line description thus yields a better operating point than writing the policy out by hand, with no weight updates and a readable final criterion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.