mlsanityData to 7 Oct 2026

The Verifier is the Curriculum: Precision Sets the Return on Search in Code Self-Distillation

arXiv 2607.09709 ↗First posted 23 Jun 2026cs.AI cs.SEcited by 2

Not on the map. Why?

Abstract (arXiv)

Post-training a code generator against a learned judge can optimize proxy features that raise the score without improving the artifact. We study the opposite signal: a deterministic, judge-free filter that asks only whether a generated project launches cleanly under a headless engine (strict-launch). Under this gate, rejection-sampling self-distillation compounds out-of-family generalization: on GameCraft-Bench a 14B model raises the per-candidate clean-launch rate on four held-out families from 8.8% to 42.2% and coverage at 32 candidates from 84% to 100%, the gold references' own ceiling, beating the supervised model on every one of the 25 held-out tasks. The gate costs one engine invocation per candidate: no reward model, no judge. At a fixed admitted count, what governs the loop is verifier precision. Swapping in a lenient build check alone erases the gain (p=0.0012); a matched gold-duplication control regresses below the supervised model. Under a semantic gate on APPS, dialing fuel precision from 1.0 to 0.25 at fixed candidate count prices that fuel linearly: over 23 training seeds, half-clean fuel returns +3.59 percentage points against the +3.69 a linear rate predicts. Under count-matched rejection-SFT only one direction of verifier error carries a measurable cost. Search obeys the same gate: quadrupling the harvest budget is worth +1.62pp behind a strict gate and nothing distinguishable from zero behind a partial-credit one. Recall is nearly free; search pays only through a precise gate: the verifier is the curriculum.

Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Builds on 25 of its 37 references that are in the map. First posted on arXiv 23 Jun 2026.

Builds on

25 of 37 references.

References · 25

  1. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  2. 2021Evaluating Large Language Models Trained on Codecited by 8,053
  3. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
  4. 2016Concrete Problems in AI Safetycited by 1,582
  5. 2025Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?cited by 948
  6. 2019Improved Precision and Recall Metric for Assessing Generative Modelscited by 925
  7. 2021Measuring Coding Challenge Competence With APPScited by 879
  8. 2023Reinforced Self-Training (ReST) for Language Modelingcited by 486
  9. 2023Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Modelscited by 325
  10. 2024V-STaR: Training Verifiers for Self-Taught Reasonerscited by 214
  11. 2025ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Casescited by 51
  12. 2024Cheating Automatic LLM Benchmarks: Null Models Achieve High Win Ratescited by 42
13 more
  1. 2025Reinforcement Learning with Verifiable yet Noisy Rewards under Imperfect Verifierscited by 40
  2. 2026Embarrassingly Simple Self-Distillation Improves Code Generationcited by 36
  3. 2026GATES: Self-Distillation under Privileged Context with Consensus Gatingcited by 30
  4. 2026Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Usecited by 25
  5. 2026LLMs Gaming Verifiers: RLVR can Lead to Reward Hackingcited by 24
  6. 2026GameDevBench: Evaluating Agentic Capabilities Through Game Developmentcited by 21
  7. 2026OpenGame: Open Agentic Coding for Gamescited by 12
  8. 2025When Does Verification Pay Off? A Closer Look at LLMs as Solution Verifierscited by 11
  9. 2026An Imperfect Verifier is Good Enough: Learning with Noisy Rewardscited by 11
  10. 2026GameGen-Verifier: Parallel Keypoint-Based Verification for LLM-Generated Games via Runtime State Injectioncited by 9
  11. 2026Delay, Plateau, or Collapse: Evaluating the Impact of Systematic Verification Error on RLVRcited by 6
  12. 2026JAMER: Project-Level Code Framework Dataset and Benchmark on Professional Game Enginescited by 5
  13. 2026Teacher-Free Self-Training Amplifies but Does Not Compound: A Pass@K Crossover on a Free-Verifier Domaincited by 4

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026GameASG-Bench: Benchmarking Autonomous Software Generation for Game Developmentcited by 0
  2. 2026RSIGame: Autonomous Agentic Game Development with Recursive Self-improvementcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Primal Generation, Dual Judgment: Self-Training from Test-Time Scaling
  2. 2025SELF-REDRAFT: Eliciting Intrinsic Exploration-Exploitation Balance in Test-Time Scaling for Code Generation
  3. 2026G-Zero: Self-Play for Open-Ended Generation from Zero Data
  4. 2026The Imitation Game: When LLMs Learn to Reason Like Programs via Code-Centric Reasoning Data Synthesis
  5. 2026SpecBench: Measuring Reward Hacking in Long-Horizon Coding Agents
  6. 2026Self-Distillation Policy Optimization via Visual Feedback: Bridging Code and Visual Artifacts
  7. 2026More Convincing, Not More Correct: Self-Play Reward Hacking of Reference-Free LLM Judges
  8. 2026Information-Gain Rewards over Diversity-Pruned Tests: GT-Anchored Verifier Co-Training for Reliable Code Generation