PaperlayerView exact revision on arXiv

Source-linked AI summary

AI Research Preference Models

Thomas Simon Foster, Bassel Al Omari, Tingchen Fu, Thomas Mann, Carl Domond, Lucia Cipolina-Kun, Bhavul Gauri, Muna Aghamelu, Alexander D. Goldie, Eryk Helenowski, Jean-Christophe Gagnon-Audet, Alberto Pepe, Saba Nazir, Daniel Izcovich, Noam Levi, Rishi Hazra, Karen Hambardzumyan, Nicolas Baldwin, Xian Li, Martin Josifoski, Paris Giampouras, Masoud Jalili Sabet, Anya Sims, Hela Momand, Tatiana Shavrina, Despoina Magka, Jason Weston, Yulin Wang, Anirudh Goyal, João Henriques, Yoram Bachrach, Emily McMilin, Jakob Nicolaus Foerster

arXiv:2608.13940v1cs.AI

TL;DR

AI research agents can generate candidates faster than they can evaluate them, making compute allocation a central research bottleneck. This paper introduces AI Research Preference Models that rank candidates for execution, raising AIRS-Bench’s average normalized score from 0.684 to 0.711 and 0.729 while reaching baseline performance in roughly 15 hours.

  • Problem

    Frontier AI research agents lack cheap feedback, so selecting which candidate research directions merit limited execution compute remains a central challenge.

  • Method

    AI Research Preference Models use frozen pretrained models to rank candidate solutions, either by inference alone or with small-scale pilot experiments.

  • Results

    RPM-guided AIRA-dojo improves average normalized AIRS-Bench score from 0.684 to 0.711 and 0.729, while matching the unguided agent’s 24-hour performance in roughly 15 hours.

  • Takeaways & Limitations

    Ranking candidates with RPMs enables more compute-efficient research search and produces new state-of-the-art results on two AIRS-Bench tasks.

  • Takeaways & Limitations

    The evaluation assumes negligible future LLM inference cost, while current inference adds latency and infrastructure-dependent monetary overhead.

Abstract

from arXiv · show

AI research agents (AIRA) can now propose, implement, and evaluate their own machine learning experiments, but progress on frontier tasks is throttled by cost: a candidate solution can be written in minutes, whereas evaluating it can take hours to days of GPU time. An agent can therefore propose far more candidates than it can afford to run, and its progress depends on its research preference: how it allocates a fixed execution budget across many candidates. We introduce AI Research Preference Models (RPMs) that predict which of multiple candidate solutions are most worth executing, without paying the cost of executing them all. We build RPMs from frozen pretrained language models (with no task-specific training), in two forms: an inference-only model that reasons over candidate plans, code, and prior executed solutions, and an agentic model that additionally runs small-scale pilot experiments before deciding. We integrate both into the AIRA-dojo search agent and evaluate on AIRS-Bench, a recent benchmark of machine learning research tasks for AI research agents. The two variants raise the average normalized score from 0.684 to 0.711 and 0.729 respectively, and reach the unguided agent's 24-hour performance in roughly 15 hours, using less than two-thirds of its execution budget. Our best RPMs also yield new state-of-the-art results on two AIRS-Bench tasks.

1 Introduction

AI research agents are bottlenecked by the hours-to-days of GPU compute required to execute experiments, not by generating candidate solutions. The paper introduces frozen-weight Research Preference Models that select promising candidates, raising AIRS-Bench performance and improving compute efficiency.

  • Motivation: AI research agents face slow progress because executing experiments can consume hours to days of GPU time, despite quick candidate generation.Frontier machine learning research lacks cheap feedback for training models and measuring performance.
  • Approach: Research Preference Models allocate experimental compute by choosing which of N newly generated candidates is most valuable to execute next.The RPM uses the full context of previously executed solutions and is integrated into the AIRA-dojo search harness.
  • Results: 0.711: Inference-only RPMs raise AIRA-dojo’s AIRS-Bench performance from 0.684 to 0.711.These frozen-weight language models reason over candidate plans, code, and previous solutions.
  • Results: 0.729: Agentic RPMs further raise AIRS-Bench performance to 0.729, approaching the validation oracle ceiling of 0.748.Agentic RPMs extend Inference-only RPMs with the ability to run small-scale pilot experiments.
  • Results: Roughly 15 hours: agents with the best RPMs match the unguided agent’s 24-hour performance using less than two-thirds of its execution budget.The best RPMs also produce new state-of-the-art results on two AIRS-Bench tasks.

2 Background

An AIRA autonomously generates and executes code to produce artifacts optimized by task-specific reward functions. The paper evaluates such agents on AIRS-Bench and frames their search as scaffold-controlled exploration of candidate-solution graphs, where preference models can guide execution choices.

  • Agent definition: An AIRA is an autonomous computer system that generates and executes code to produce artifacts evaluated by task-specific reward functions.Artifacts include model weights, optimized code snippets, and answers to questions.
  • AIRS-Bench: AIRS-Bench contains 20 machine learning research tasks spanning language modeling, mathematics, bioinformatics, and time series forecasting.It evaluates idea generation, experiment analysis, and iterative refinement without baseline code, with each task specifying a problem, dataset, target metric, and published state-of-the-art value.
  • AIRS-Bench: Agents receive a training dataset, test inputs, and 24 hours of H200 GPU access to train models and submit predictions, while the true test score remains hidden.Agents may use validation scores, including cross-validation results, to guide their search.
  • Search scaffolds: An AIRA consists of an LLM backbone and an algorithmic scaffold that governs candidate generation, evaluation, and iterative refinement across an evolving solution graph.The graph stores code scripts, execution logs, and metric scores; preference models can decide which accumulated candidates are worth executing.

3 AI Research Preference Models

AI Research Preference Models (RPMs) address costly experimental evaluation by ranking candidate research solutions relative to one another before allocating compute. They include inference-only and agentic variants and are integrated into AIRA-dojo to select promising mutations before execution.

  • Motivation and formulation: RPMs reformulate experimental allocation as relative ranking because language models forecast absolute metrics and execution outcomes unreliably.The ranking selects promising research paths before dedicating compute to pursuing them.
  • RPM variants: Inference-only RPMs rank candidates using lightweight reasoning over search history and code diffs, while agentic RPMs first run rapid sandbox pilots.The two variants use different amounts of test-time compute before ranking candidate solutions.
  • AIRA-dojo integration: RPMs are scaffold-agnostic and are integrated into AIRA-dojo’s evolutionary child-creation phase to select among multiple candidate modifications before GPU execution.The implementation replaces single-child creation with parallel candidate generation and an RPM-guided tournament.
  • Agentic RPMs: Agentic RPM pilots use multi-turn reasoning, tool calls, and environment feedback in an exact AIRA environment with access to a single H200 GPU.The pilot agent can access the training and unlabeled test datasets and use Python, Bash, and submit_solution tools.
  • Search context: The RPM grounds comparisons in earlier search results by receiving up to K non-buggy tree nodes with their validation scores.These context nodes are collected by breadth-first traversal from the selected parent.

4 Experimental Setup

The study integrates RPMs into AIRA-dojo and evaluates them on 20 public AIRS-Bench text and tabular tasks under a standardized 24-hour, single-H200 protocol. It also uses oracle and No-RPM comparisons plus an offline dataset from 40 separate unreleased multimodal tasks to support RPM development.

  • End-to-end evaluation: RPMs are integrated into AIRA-dojo’s child-creation phase for evaluation against AIRS-Bench tasks.The augmented scaffold is evaluated on 20 publicly released tasks spanning text and tabular modalities.
  • End-to-end evaluation: 20 public AIRS-Bench text and tabular tasks receive 24 hours on one H200 each and are repeated across 10 seeds.Evaluations follow the original AIRS-Bench protocol and use Hidden Consistent Evaluation from AIRA2.
  • Controlled comparison: Identical models generate and select child candidates, ensuring improvements are attributed to the framework rather than a stronger selection backbone.Both the AIRA-dojo operator and RPMs share a Qwen-based model, as specified in the setup passage.
  • Baselines and oracles: The No-RPM baseline selects generated candidates uniformly at random, while Test and Validation Oracles execute all candidates and choose the highest scorer.Only the selected candidate’s compute time counts toward the 24-hour limit; the oracles are not viable in practice.
  • Offline evaluation: An offline dataset from 40 unreleased AIRS-Bench image, video, and audio tasks enables rapid RPM development before expensive end-to-end evaluations.The development and evaluation sets are split by modality to avoid task contamination; full evaluations require 200 H200 GPUs for 24 hours.

5 Results

RPM-guided AIRA-dojo improves search decisions through complementary compute strategies: inference-only selection delivers immediate gains, while agentic pilots trade early speed for stronger later performance. These gains are statistically significant, improve research efficiency, and produce benchmark breakthroughs, with offline analyses identifying scaling and behavioral factors behind selection quality.

  • End-to-End Performance: Inference-only RPM provides immediate, steady gains without candidate-execution cost, whereas Agentic RPM’s per-step proxy experiments slow early progress before surpassing other methods.The two RPMs therefore navigate decision quality versus compute time differently.
  • Significance Testing: 0.5923 and 0.5913 average improvement probabilities show statistically significant advantages for Inference-only and Agentic RPMs over No RPM, with 95% confidence lower bounds of 0.5066 and 0.5018.Both intervals strictly exclude the 0.5 no-difference threshold.
  • State-of-the-Art Breakthroughs: 94.1% accuracy on WinoGrande with Agentic RPM surpasses the previous SOTA of 88.1%, while Inference-only RPM reaches 95.7% accuracy on SVAMP.The WinoGrande run fine-tunes Qwen2.5-14B-Instruct with LoRA on shuffled-label data and averages logits across label orderings to reduce position bias.
  • Research Efficiency: 14.88 hours and 15.50 hours are the times Inference-only and Agentic RPMs take to match No RPM’s final score of 0.684, yielding 1.61× and 1.55× speedups.The No-RPM baseline requires the full 24-hour allocation to reach that threshold.
  • Offline Evaluation: Offline evaluation finds Agentic RPMs more accurate than the best Inference-only configuration, with both outperforming random selection, matching end-to-end trends.The offline framework supports rapid tuning and analysis without running full 24-hour evaluations.
  • Compute and Behavioral Scaling: 78.52%, 82.78%, and 84.02% selection accuracy result from agentic compute limits of 5 minutes, 30 minutes, and 4 hours, respectively.Behavioral analyses further find that full-dataset evaluation is more reliable than subsampling, while timeouts and removing schedulers or augmentations reduce accuracy.

6 Related Work

This work relates AI research agents to methods that allocate experimental compute without evaluating every candidate, and to preference, judge, and reward models for comparative evaluation. Its closest distinctions are grounding comparisons in the live search tree and executing pilot ML experiments without model training.

  • AI research agents and automated ML research: AI research agents formalize machine learning research as agentic search involving idea generation, coding, experimentation, result analysis, and paper drafting.
  • Model-based methods for expensive search: Model-based methods use world models, surrogate models, or learned policies to plan, generate synthetic rollouts, or allocate experimental compute efficiently.
  • Preference, judge, and reward models: RPMs connect to preference, judge, and reward models, including pairwise-comparison and ranking approaches used in reinforcement learning from human feedback.
  • Preference, judge, and reward models: Unlike Zheng et al. (2026), RPMs ground pairwise preferences in the live search tree and executed-solution validation scores rather than a static data report.
  • Preference, judge, and reward models: RPMs implement Goldie et al. (2026)'s proposal to select promising research-agent leaves without model training and extend agentic verification with pilot ML experiments.

7 Limitations

The study’s claims are limited by inference latency and off-policy, biased offline data, while RPM integration and portability remain demonstrated only in a narrow setting. Parent-selection integration and broader scaffold or backbone transfer are left for future work.

  • Inference cost: 0.660 hours of inference latency per 24-hour run reduces the Inference-only RPM’s normalized score from 0.711 to 0.708 at 23.34 hours.The latency uses a self-hosted Qwen3.6-27B; exact latency and monetary overhead vary by hosting infrastructure.
  • Data bias: Offline data from prior greedy AIRA-dojo runs are off-policy and biased because they use different backbones and task modalities, including subtree-max label bias.Consequently, the paper’s main claims rest on end-to-end results.
  • Integration scope: RPM integration is scoped to child creation; final-node selection shows no significant improvement over validation-based selection, while parent-selection integration remains future work.The final-node result is reported under the Hidden Consistent Evaluation protocol’s strong test-validation generalization.
  • Portability: Portability is demonstrated only with AIRA-dojo, Qwen3.6-27B, and one benchmark, while the agentic variant additionally requires a sandboxed execution-environment clone.Transfer to other scaffolds and backbones is left for future work.

8 Conclusion … A.1 Prompt Optimization

RPMs rank unexecuted candidates to allocate scarce evaluation compute, using either inference-only reasoning or pilot experiments, and improve AIRA-dojo performance on AIRS-Bench. The inference-only RPM’s prompt is optimized without weight updates through an APO/MIPROv2-style search that converges on a structured PI evaluation rubric and an optimized candidate-selection template.

  • 8 Conclusion: RPMs direct execution budgets toward promising unexecuted candidates instead of forecasting absolute candidate outcomes.They address the gap between fast candidate generation and expensive evaluation by selecting which candidates are worth executing.
  • 8 Conclusion: The inference-only RPM uses candidate solutions and prior evaluations, while the agentic RPM additionally runs pilot experiments in a sandbox.Both variants are built from frozen pretrained models and designed to be portable across search scaffolds.
  • 8 Conclusion: 0.684 to 0.711 and 0.729: RPM-guided selection raises AIRA-dojo’s average normalized AIRS-Bench score for the inference-only and agentic variants, respectively.The comparison uses the same backbone and mutation operators.
  • 8 Conclusion: Candidate selection is presented as an explicit use of test-time compute, especially when evaluating solutions costs much more than generating them.Agents can invest computation both in proposing candidates and in deciding which candidates to execute.
  • A.1 Prompt Optimization: APO optimizes the inference-only ranking prompt at inference time without underlying weight updates, using GPT-5 as ranker and meta-proposer on pairwise AIRS-Bench trajectory data.The optimization is applied to a ranking system prompt rather than model parameters.
  • A.1.1 Optimization Configuration: 10 candidate prompt variations and 40 Bayesian minibatch trials drive a MIPROv2-style search using Thompson sampling over Beta(1, 1) accuracy posteriors.Underperforming templates are pruned, while top survivors receive deeper validation to mitigate optimization-to-holdout shrinkage.
  • A.1.1 Optimization Configuration: The optimized prompt replaces the baseline strict-judge framing with a PI persona and an ordered five-criterion rubric covering fit, extensibility, learning curves, novelty, and risk-adjusted potential.The rubric also shifts from binary bug intolerance toward distinguishing fixable roughness from fundamental conceptual flaws and uses context to identify gaps and penalize redundancy.
  • A.1.3 Prompt Templates Baseline Template (Before): The baseline and optimized templates choose between two candidates by emphasizing eventual test-score potential, with the optimized version explicitly assessing alignment, extensibility, learning-curve gains, novelty, and fixability.Both templates require a single A/B decision after reasoning over the task, context, plans, and code; the optimized template asks evaluators to separate trivially fixable issues from fundamental flaws.

A.1.2 Results … A.2.2 Ensembling Configuration

Offline prompt optimization improves candidate-selection accuracy and produces the ranking prompt used by the core agent. The ensembling setup evaluates three frontier LLMs individually and jointly, using repeated rollouts with either majority voting or an LLM arbiter.

  • A.1.2 Results: Offline prompt optimization improves mean single-sample accuracy by +0.7 pp and majority-vote accuracy by +1.2 pp.The optimized PI rubric is evaluated on a matched dataset with n = 990, where accuracy measures correct higher-potential candidate selections.
  • A.1.2 Results: The optimized prompt layout is hand-carried directly into the core ranking agent.
  • A.2 Model Ensembling: GPT-5, Claude Opus 4.8, and Gemini 3.1 Pro are evaluated individually and ensembled as offline-ranking baselines.
  • A.2.1 Single-Model Configuration: Single-model evaluation uses the optimized ranking prompt, maximum reasoning effort, and the maximum number of context nodes that fit within each model.
  • A.2.2 Ensembling Configuration: Both ensembling configurations use 3 independent rollouts for each baseline model before aggregating predictions.
  • A.2.2 Ensembling Configuration: Majority Vote extracts each rollout’s model response and selects the final prediction by simple majority, breaking ties uniformly at random.
  • A.2.2 Ensembling Configuration: The LLM Arbiter uses Claude Opus 4.8 to assess anonymized reasoning traces alongside the original ranking payload and select the final candidate.The arbiter is instructed to evaluate argument quality rather than blindly follow the majority choice.

A.2.3 Results

Offline ranking results show that frontier models perform similarly, while ensembling improves selection accuracy; the LLM-Arbiter achieves the best result by weighting argument quality rather than vote counts.

  • Offline ranking results: 68.04% accuracy is achieved by standard majority voting, improving over individual frontier-model performance.Claude Opus 4.8 and Gemini 3.1 Pro perform around 67.4%, while GPT-5 reaches 64.66%.
  • Offline ranking results: 69.35% accuracy is achieved by the LLM-Arbiter, outperforming standard majority voting through argument-quality weighting.Majority voting reaches 68.04%.

A.3 Reasoning Analysis · B Agentic RPM · B.1 Prompt Templates

The reasoning analysis shows that inference-only RPM accuracy improves when selections cite correctness, pretrained-backbone, and prior-evidence signals, while abstract architecture rationales hurt performance and deeper historical-context citation helps. The agentic RPM workflow uses staged prompts for pilot experiments, iterative experiment selection, and final prediction, with selection advantage correlated with final performance.

  • A.3 Reasoning Analysis: 8,715 pairwise rollouts, with 3 independent rollouts per pair, were analyzed across gpt-5, claude-4.8-opus, and gemini-3.1-pro.The traces came from evaluations described in Section A.2.1.
  • A.3 Reasoning Analysis: 68.3% vs. 62.4% selection accuracy resulted when reasoning grounded choices in Prior-Evidence, compared with when that justification was omitted.Correctness and Pretrained Backbone grounding also improved accuracy: 66.9% vs. 63.9% and 67.0% vs. 63.9%, respectively.
  • A.3 Reasoning Analysis: Selection accuracy scales monotonically with the number of distinct historical context-node metric values cited in each reasoning trace.The analysis extracts, deduplicates, and grounds explicit node identifiers and validation metrics from prompt history.
  • A.3 Reasoning Analysis: 67.3% to 64.7% selection accuracy occurred when models relied on abstract Architecture justifications.The analysis categorized traces into Correctness, Pretrained Backbone, Prior-Evidence, Hyperparameters, and Architecture.
  • B.1 Prompt Templates: The agentic RPM prompt templates cover initial pilot experimentation, iterative feedback, and final prediction.These stages are represented in Figures 8, 9, and 10.
  • B.1 Prompt Templates: Pilot prompts require a fast Python script under {time_limit} to generate evidence distinguishing two candidate solutions without evaluating on the test set.Suggested strategies include partial cross-validation, subsampling, targeted ablations, or a single train/validation split.
  • B.1 Prompt Templates: The feedback prompt decides whether to stop or propose exactly one next experiment targeting the highest decision value and most important unresolved ranking uncertainty.It instructs a separate coding agent to execute the experiment and a separate judging agent to make the final prediction.
  • B Agentic RPM: Pearson r = 0.55, p = 0.0007; Spearman ρ = 0.56, p = 0.0004 for the correlation between selection advantage and final normalized score.Agentic RPM achieved both higher selection advantage and final score than Inference-only RPM and Random selection.

C End-to-End Evaluations … D.3.2 Why there is little to win: validation ≈test on AIRS-BENCH

The evaluations show that better local candidate selection predicts stronger end-to-end search performance, while post-hoc code-based final-node selection adds little because validation already tracks test closely on AIRS-Bench. The RPM’s useful advantage is therefore concentrated in online child selection, where candidates are unexecuted and lack validation scores.

  • C.1 Impact of Selection Quality: Pearson r = 0.55 and Spearman ρ = 0.56 show a significant positive correlation between local selection advantage and final end-to-end performance.The corresponding p-values are 0.0007 and 0.0004; Inference-only RPM beats random selection, while Agentic RPM achieves the highest selection advantage and final score.
  • D.1 The question, and why final-node selection is different: RPM intervention can target parent, child, or final-node selection, but only child selection lacks a free validation score because its candidates are unexecuted.Parent and final-node selection operate on already-executed candidates with validation scores.
  • D.1 The question, and why final-node selection is different: The Pick Your Poison study tests whether an RPM can select a better final submission than the highest-validation node from a completed search tree.It motivates reserving RPM computation for child selection, where no comparable shortcut exists.
  • D.2 Method: PYP presents top-validation candidates at N ∈ {2, 4, 6, 8, 10} and uses tournament selection with match_size = 5 and nmatches = 30.The validation oracle is guaranteed to appear in each presented pool; reported cells have a failed-match rate below 0.5%.
  • D.3 Results on the test set (public text-and-tabular tasks): At N=2, the RPM is at or above the validation oracle for all four variants, but its RPM −oracle gap declines and becomes negative by N=4–6 for three variants.The gpt5_x_qwen variant stays near zero as pool size grows; oracle and actual-best means remain flat across N.
  • D.3.1 The “beats-rate” is tie-dominated; the override-win-rate is near chance: 78–96% beats-the-oracle rates are tie-dominated: 69–85% of trees have identical test scores, while strict wins are ∼8–9% and strict losses are ∼5–22%.Override-win-rate is therefore the more informative metric because it conditions on cases where the RPM changes the validation-based choice.
  • D.3.2 Why there is little to win: validation ≈test on AIRS-BENCH: Median per-tree Spearman correlation is 0.92–0.96, the best-validation node is also best-test on 68–71% of trees, and perfect-selector headroom is only ≈0.024–0.031 normalized.These results explain why final-node code judgment has little room to improve over validation on AIRS-Bench.
  • D.3.2 Why there is little to win: validation ≈test on AIRS-BENCH: +0.0233 test is recovered by gpt5_x_gpt5 on the ∼30% of N=10 trees where the best-validation node is not the best-test node.This recovers ∼44% of available headroom, indicating that code judgment helps mainly when validation is unreliable or unavailable, including online child selection.

D.3.3 Robustness: more matches reduce aggregation noise, they do not add skill … D.5 The PYP judge prompt (best-of-N, no plans, validation score shown)

More tournament matches reduce aggregation noise but do not add code-judgment skill, while matched Qwen scoring confirms weaker selection for Qwen variants. On validation, inference-only RPM selection does not reliably improve the val-oracle; PYP selects among five code candidates using shown validation scores and tournament voting.

  • D.3.3 Robustness: more matches reduce aggregation noise, they do not add skill: 3.3× more matches at N=10 increased variants beating the oracle from 1/4 to 2/4, while override-win-rate remained essentially unchanged below 50%.At N=4, increasing matches from 30 to 50 raised oracle-beating variants from 2/4 to 3/4.
  • D.3.3 Robustness: more matches reduce aggregation noise, they do not add skill: Match_size= 5 outperformed full-pool matching because diverse sub-comparisons drive tournament value.Changing match_size from 2 to 5 lifted per-tree beats-rate to 78–96% and reduced the mean oracle gap by ∼0.01–0.015.
  • D.3.4 Using the matched judge: Qwen-scored Qwen variants: Qwen-scored Qwen-search variants reproduced the GPT-5-judge pattern, with near-zero gap at N=2 and increasingly negative gaps as N grew.Qwen is the weaker selector for gpt5_x_qwen and qwen_x_qwen under the matched Qwen3.6-27B judge.
  • D.3.4 Using the matched judge: Qwen-scored Qwen variants: 69–85% of N=10 oracle comparisons were ties, while strict wins were ∼9% and Qwen-operator strict losses outweighed strict wins.The high beats-rate therefore primarily reflects ties rather than strict improvements.
  • D.4 Results on the validation set (unreleased multimodal tasks): +0.004 to +0.011 was the GPT-OSS-120b and CWM gap over the val-oracle across every N, whereas O3-Mini was ∼−0.005 below it.The validation set reproduced the test-set picture: RPM ≈val-oracle with no measurable net advantage.
  • D.4 Results on the validation set (unreleased multimodal tasks): Inference-only RPM judging does not reliably improve AIRA-dojo’s highest-validation default because validation closely tracks test under Hidden Consistent Evaluation.The informative selector test is instead where validation is unreliable or unavailable, especially online child selection.
  • D.5 The PYP judge prompt (best-of-N, no plans, validation score shown): Each PYP match asks a strict judge to choose the best among 5 candidate solutions with full executed code, no plans, and per-candidate normalized validation scores shown.The example concerns molecular-property prediction for internal energy at 0K (U_0), evaluated by MeanAbsoluteError.
  • D.5 The PYP judge prompt (best-of-N, no plans, validation score shown): One point per match determines the tournament winner, with random tie-breaking; when N < match_size, the match size and boxed-letter menu shrink to N.The judge outputs a boxed candidate letter that maps back to a candidate selected for the tree.
Loading 2608.13940v1…