Research map / Multi-agent systems and games / LLM agents

What Fits (Into Few Tokens) Doesn't Overfit: Compression and Generalization in ML Research Agents

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Tests whether adaptive ML research agents overfit a validation set by compressing their strategy into short prompts for a fresh reproducer and by restricting feedback to one-bit ladder signals, finding little overfitting and a description-length explanation.

Problem

Adaptive reuse of held-out validation/benchmark data should invite overfitting, yet benchmark-driven ML shows little of it in practice; the paper asks whether successful ML strategies are highly compressible and whether that explains the lack of overfitting.

Approach

Use LLM research agents as resettable explorers and test compressibility via two bottlenecks: output compression (a fresh reproducer rebuilds the strategy from a ≤32/64-token prompt plus training data) and input compression (explorer sees only one-bit 'improved on running best' feedback, a ladder mechanism), with description-length generalization bounds for both.

Key results

• Across 41 improvement checkpoints, 32-token reproducers match or exceed the explorer on 38 (92.7%) under a one-sided 5% relative-gap criterion.
• Ladder (one-bit) explorer matches or slightly exceeds score-based explorer on holdout across all 8 datasets.
• Under aggressive validation prompting, 38 of 102 checkpoints overfit by >10% relative gap; 128-token reproducer failure flags all 38 with 100% sensitivity and 91% specificity.

Surprising findings

• One-bit ladder feedback matches or slightly exceeds full numerical score feedback on holdout performance.
• A 16-token prompt suffices to reproduce a language-modeling strategy with no loss in holdout performance.

Limitations

• Possible pre-training contamination of LLM with validation data is a side channel not fully ruled out.
• Output compression certifies only the reproducer's hypothesis; explorer CIs rely on the ladder bound, which covers only classification tasks.
• Overfitting stress tests use aggressive prompting and small or noisy validation sets, which may not reflect ordinary benchmark reuse.

Experimental setup

Datasets: Folktables (ACSIncome) Gene Expression (synthetic) SST-2 CIFAR-10 CIFAR-100 ImageNet-1K HH-RLHF WikiText-103

Benchmarks: Folktables (ACSIncome) Gene Expression SST-2 CIFAR-10 CIFAR-100 diffusion ImageNet-1K HH-RLHF reward modeling WikiText-103

Metrics: accuracy MSE bits per byte log-loss validation-holdout gap sensitivity specificity

Baselines: score-based feedback explorer empty-prompt reproducer

Builds on: Ladder mechanism (Blum & Hardt) Adaptive data analysis PAC-Bayes prompt bounds MLAgentBench MLE-bench

Closest papers by meaning

• Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression (ICML 2025 · similarity 0.87)
• MLR-Bench: Evaluating AI Agents on Open-Ended Machine Learning Research (NeurIPS 2025 · similarity 0.86)
• A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective (NeurIPS 2025 · similarity 0.86)
• When Less is More: The LLM Scaling Paradox in Context Compression (NeurIPS 2026 · similarity 0.85)
• Fix the Structural Bottleneck: Context Compression via Explicit Information Transmission (NeurIPS 2026 · similarity 0.85)
• Slower Generalization, Faster Memorization: A Sweet Spot in Algorithmic Learning (NeurIPS 2026 · similarity 0.85)