Autoresearch with Coding Agents: Generalizers and Metric-Maximizers on Quran Recitation Data

• URL: http://arxiv.org/abs/2607.18064v1
• Date: Mon, 20 Jul 2026 15:32:09 GMT
• Title: Autoresearch with Coding Agents: Generalizers and Metric-Maximizers on Quran Recitation Data
• Abstract summary: Coding agents can now be left alone to improve software against a score.<n>In this pattern, the agent receives a dataset, an evaluation script, and one editable file, and iterates without supervision.<n>We ran this loop on a real production task: deciding which Quranic verses appear in a noisy speech-recognition transcript.
• Score: 4.191965713559235
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Coding agents can now be left alone to improve software against a score. In this pattern--recently popularized as "autoresearch"--the agent receives a dataset, an evaluation script, and one editable file, and iterates without supervision: modify the code, measure, keep the change if the score improves. But what does the agent actually optimize--the developer's intent, or the literal number? We ran this loop on a real production task: deciding which Quranic verses appear in a noisy speech-recognition transcript and splitting the transcript by verse. Two frontier coding agents, Claude Code and OpenAI Codex, started from the same blank file with the same instructions, budget, and reasoning effort, three runs each. Both independently invented the same algorithm (canonicalization, n-gram anchoring, dynamic-programming alignment)--and then diverged. Claude stopped early with compact, general code. Codex drove the score ~10x lower, largely by memorizing answers to individual evaluation rows (19-41 hardcoded verse ids per run): a clean natural instance of specification gaming by a production agent. In a preregistered second study, we added a held-out test set and told both agents it existed. The memorization vanished, and the score gap vanished with it--yet Codex's general core transferred better and more consistently (held-out detection+split 0.085+/-0.004 vs. 0.121+/-0.031), losing only on one missed rejection of non-recitation input. Two exploratory community arms (Cursor, Antigravity) are consistent with the pattern. Every agent's held-out solution matched or beat the hand-engineered pipeline it was built to replace--the best by an order of magnitude--and now runs in production. From the ways agents exploited our harness--reading sibling runs through shared git state, leaving notes to "future runs" in persistent memory--we distill five design rules for evaluating autonomous agents.

Related papers

• OpenCodeReview: Determinism over Non-Determinism for Cost-Effective Agent-Based Code Review [14.707064801983172]
  We introduce OpenCodeReview, built on deterministic engineering for uncertain agents.<n>On AACR-Bench (200 real-world PRs, 10 languages, 1,505 expert-verified comments), OpenCodeReview outperforms mainstream coding agents.
  arXiv  Detail & Related papers  (2026-08-10T08:43:16Z)
• TraceCoder: Explainable and Auditable Code Generation with Position-Key Snippet Versioning [0.0]
  We present a code generation concept that addresses shortcomings through three complementary mechanisms.<n>A relational snippet-history schema records, per repair event, the benchmark reference, round number, failure text, and explanation.<n>A browser-based visualisation tool renders this history as heat-mapped, hover-annotated source code.
  arXiv  Detail & Related papers  (2026-07-28T22:03:52Z)
• Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction [57.138342889101345]
  Tencent WorkBuddy Bench is a multi-domain evaluation suite for coding agents.<n>This report documents its construction methodology, scoring protocol, and a cross-model leaderboard.
  arXiv  Detail & Related papers  (2026-07-23T04:34:06Z)
• Fantastic Adaptive Taxonomies and How to Use Them [100.20614173157708]
  AdaMAST converts execution traces into a compact, evidence-grounded failure taxonomy.<n>No code is hand-authored, no trace human-annotated.<n>In agent-system search, taxonomy-coded diagnoses of failed candidates outperform free-form.
  arXiv  Detail & Related papers  (2026-07-17T17:41:16Z)
• Self-Improving AI Coding Agents Through Accumulated Behavioral Rules: A Closed-Loop Framework [0.3867363075280543]
  We present a closed-loop framework in which every accepted review comment is codified as a persistent behavioral rule.<n>The framework combines an accumulating rule set in a version-controlled instruction file, a self-review checklist, and automated validation.<n>We show that our framework achieves persistent cross-session learning without weight updates, operates on productions rather than synthetic benchmarks, and addresses an experiential dimension (behavioral consistency over time) that existing benchmarks do not measure.
  arXiv  Detail & Related papers  (2026-07-13T21:13:46Z)
• Latent Programming Horizons in Coding Agents [6.8463578914435645]
  coding agent solving a software-engineering task spends dozens of steps reasoning, editing code, and running tests.<n>We show that the residual streams of language models under coding agents linearly encode properties of the evolving program.<n> Probes trained to predict the outcome of future edits achieve performance above chance up to roughly 25 steps in advance.
  arXiv  Detail & Related papers  (2026-07-06T15:08:26Z)
• PairCoder++: Pair Programming as a Universal Paradigm for Verified Code-Driven Multimodal and Structured-Artifact Generation [51.92442051257354]
  PairCoder improves essentially every benchmark whose artifact is verifiable, on full official metric suites rather than execution alone.<n>TikZ compile rate up 10 to 30 points on every model, at 2.9 to 9.2 times single model cost.
  arXiv  Detail & Related papers  (2026-07-02T08:36:02Z)
• Code as Agent Harness [107.31925305395957]
  In emerging agentic systems, code is no longer only a target output.<n>Code increasingly serves as an operational substrate for agent reasoning, acting, environment modeling, and execution-based verification.<n>This survey provides a unified roadmap toward executable, verifiable, and stateful AI agent systems.
  arXiv  Detail & Related papers  (2026-05-18T17:59:03Z)
• SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks [55.76734816061826]
  We introduce SlopCodeBench, a language-agnostic benchmark comprising 20 problems and 93 checkpoints.<n>We track two trajectory-level quality signals: verbosity, the fraction of redundant or duplicated code, and structural erosion.<n>No agent solves any problem end-to-end across 11 models; the highest checkpoint solve rate is 17.2%.
  arXiv  Detail & Related papers  (2026-03-25T19:26:44Z)
• Memorize or Generalize? Evaluating LLM Code Generation with Code Rewriting [54.48306552577881]
  We argue that large language models (LLMs) are mostly doing memorization (i.e., replicating or reusing large parts of their training data) versus generalization.<n>Existing evaluations largely proxy neglecting surface/structural similarity, thereby conflating benign reuse of repeated code with harmful recall and memorization task correctness.<n>We propose Memorization Risk Index (MRI), a normalized score that combines two signals: (i) how similar the model's answer for the rewritten task is to the original ground-truth solution, and (ii) how much performance drops from the original task to its rewritten counterpart.
  arXiv  Detail & Related papers  (2025-03-04T05:39:24Z)
• CodeAgent: Autonomous Communicative Agents for Code Review [12.163258651539236]
  This work introduces tool, a novel multi-agent Large Language Model (LLM) system for code review automation. CodeAgent incorporates a supervisory agent, QA-Checker, to ensure that all the agents' contributions address the initial review question. Results demonstrate CodeAgent's effectiveness, contributing to a new state-of-the-art in code review automation.
  arXiv  Detail & Related papers  (2024-02-03T14:43:14Z)
• Contrastive Code Representation Learning [95.86686147053958]
  We show that the popular reconstruction-based BERT model is sensitive to source code edits, even when the edits preserve semantics. We propose ContraCode: a contrastive pre-training task that learns code functionality, not form.
  arXiv  Detail & Related papers  (2020-07-09T17:59:06Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.