Skip to content

evaluation-red-teaming.md

Latest commit

 

History

History
37 lines (32 loc) · 7.89 KB

evaluation-red-teaming.md

File metadata and controls

37 lines (32 loc) · 7.89 KB
Raw

Evaluation / Red Teaming

Use this category for examples where autoresearch is used to attack, evaluate, harden, or iteratively improve systems based on measurable test results.

Submission format

- [Name](URL) - Industry: one-sentence description of the autoresearch use case.

Entries

  • Claudini - AI safety research: uses an autoresearch-style loop to invent and benchmark new LLM attack algorithms, keeping only methods that outperform baselines.
  • autovoiceevals - Voice AI evaluation: attacks voice agents with adversarial callers, proposes prompt changes one at a time, and keeps or reverts them based on eval results.
  • We Used Autoresearch on Our AI Skill, It Taught Us to Write Better Tests - AI skill evaluation: runs a prompt-migration skill against six fixed codebase test cases, scores each change on correctness, completeness, and efficiency, and keeps only improvements while cherry-picking around harness overfit.
  • AutoPrompter - Prompt evaluation: combines promptfoo-style metrics with autoresearch-style closed-loop iteration, generating datasets, testing target models, and refining prompts through a persistent experiment ledger.
  • AutonomousTester - UI testing evaluation: adapts autoresearch to Playwright test generation by editing only tests/test_suite.py, measuring coverage_score, and auto-fixing or discarding test changes until coverage improves.
  • Autoresearch for Agents from Scratch - Support-agent prompt evaluation: applies Karpathy's keep/revert loop to system_prompt.md, scoring frozen adversarial support cases by tool-call accuracy and lifting the prompt from 0.05 to 0.80 over 15 experiments.
  • LLM Privacy + Cost Router — Classifier Experiment - Privacy classification evaluation: runs a Karpathy-style autoresearch experiment across regex and prompt variants for a hybrid LLM privacy classifier, validating the best configuration at 96.7% holdout accuracy with 4.6% false negatives.
  • AutoMemory - Agent memory evaluation: lets an agent rewrite its own memory system against LongMemEval, using an immutable evaluator over random question samples and iterating on code plus strategy notes in response to scored failures.
  • How to stop your autoresearch loop from cheating - Autoresearch evaluation hardening: reports 71 experiments across nanochat training and MoE compression, showing loops drift quickly unless experiments are isolated and evaluator gates block shortcut gains.
  • Autoreason - Output evaluation: extends Karpathy-style autoresearch to subjective writing and coding tasks by running incumbent-versus-revision-versus-synthesis tournaments under blind multi-judge Borda scoring and stopping only when the unchanged version wins twice, outperforming standard self-refinement baselines on writing tasks and 150 CodeContests problems.
  • AutoResearchBench - Evaluation benchmark: a dedicated benchmark for autonomous scientific literature discovery with complementary deep and wide research tasks, showing that even the most powerful LLMs achieve only ~9% accuracy on literature-focused open-ended search.
  • Jailbreak Autoresearch - Jailbreak red-teaming: applies Karpathy's autoresearch loop to prompt-harness experiments, testing different header and footer harnesses against a fixed body and scoring responses against a rubric to find more effective jailbreak strategies.
  • JustAsk - System prompt extraction: uses autoresearch-style curiosity-driven code agents to iteratively probe and reveal system prompts from frontier LLMs, accepted at ICML 2026.
  • DSBench AutoResearch - Benchmark evaluation: applies Karpathy-style hill-climbing to each of DSBench's 74 Kaggle modeling and 38 Modeloff analysis tasks, running per-task autoresearch repos with a forensic audit that beats 82/112 DSBench baselines.
  • Trace2Evolve - Customer-support agent evaluation: applies Karpathy-style autoresearch to improve a tool-using support agent by generating hard cases, scoring traces, classifying failures, and keeping candidate improvements only when benchmark metrics and reliability gates pass across retail and airline task splits.
  • SciTriage - Autoresearch evaluation: evidence gates and claim discipline for AutoResearch agents, providing a diagnostic layer that determines whether experiment results are meaningful before deciding to continue or stop a run.
  • claude-haiku-harness - Prompt evaluation harness: uses an AutoResearch-style eval-driven loop to iteratively benchmark and improve prompts against production-quality targets, aiming to match Claude Haiku 4.5 quality with Sonnet/Opus benchmarks.
  • AutoMedBench - Medical evaluation benchmark: provides a medical-domain benchmark (MedAutoBench) for evaluating autoresearch agents on MedQA, MedMCQA, and PubMedQA clinical tasks with sandboxed execution and automated scoring.
  • NanoGPT-Bench - ML research benchmark: evaluates AI systems on open-ended GPT-2 speedrun research tasks, finding current coding agents recover under 10% of human-achievable speedup and primarily tune hyperparameters rather than perform algorithmic innovation.
  • ResearchArena - Research loop evaluation: provides a scaffold for off-the-shelf agents to carry out full research loops from ideation through experimentation to paper writing, finding that feasibility does not yet equal quality in autonomous research.
  • Anti-Autoresearch - Autoresearch integrity forensics: derived from the ARIS research-agent platform, applies a three-layer audit stack (experiment-audit, result-to-claim, paper-claim+citation-audit) with cross-model adversarial review and deterministic verdict, detecting 39 hack-patterns across 7 families to flag manipulated or fabricated autoresearch results.
  • AHA: Agent Hacks Agent - Production-agent red-teaming: applies an autoresearch loop to red-team trajectories, surfacing reusable failure mechanisms and building a frozen Vulnerability Concept Graph so agents discover their own attack surfaces before adversaries do, with published paper, casebook, and interactive website.
  • Autoresearch Novelty Bench - Novelty benchmark: evaluates autonomous research agents on whether their proposed hypotheses are novel, anticipate future breakthroughs, or merely rediscover known techniques, built on Prime Intellect's speedrunning archive.
  • ResearchClawBench - Scientific research evaluation: benchmarks 40 real-science tasks across 10 disciplines, scoring AI agents on whether they can re-discover or surpass published research findings through autonomous pipelines.
  • How Do Agents Fail on AutoResearch - Autoresearch evaluation: an end-to-end diagnostic evaluation across 100 real-world frontier research tasks that systematically characterizes where and how agents fail, providing a failure taxonomy beyond aggregate success metrics.