Anthropic Reveals AI Research Bottleneck: Agents Excel at Engineering, Not Discovery

Anthropic Evaluates Frontier Agents on Scientific Research Tasks

Anthropic's paper "Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?" (arXiv:2610.28763) introduces Agent4Sci , a testbed where agents receive a paper, code, data, and a target metric, then attempt to improve the metric starting from a published SOTA baseline. Experiments run on 8 H200 GPUs for 12 hours with zero human intervention across 20 benchmarks, 25 tasks, and 11 scientific domains . To ensure clean evaluation, the pipeline separates three isolated roles: a climbing agent that modifies code, a scoring agent that evaluates on held-out test sets, and an audit agent that checks for cheating—eliminating test-set contamination and metric manipulation.

Benchmark Results: All Tasks Surpassed, But Gains Are Engineering-Driven

Every benchmark was beaten; no exceptions.

Average relative improvements: Claude Opus 10.4% , Sonnet 9.6%, Gemini 6.8%, Codex 4.4%.

Analysis of thousands of operation traces categorized agent actions into seven strategy types: hyperparameter tuning, model architecture changes, training scheduling, feature engineering, data augmentation, loss design, and model ensembling. The high-frequency strategies form a standard ML engineering checklist (learning-rate schedules, widening/deepening networks, backbone swaps, weight decay, data augmentation). Each model shows distinct preferences: Opus favors cosine learning rates and larger models; Codex leans toward pooling and data augmentation; Gemini prefers feature engineering. Problem-specific innovations are extremely rare across all three agents.

Experimental workflows also differ: Opus iteratively monitors logs and corrects; Codex submits a single long script and waits; Gemini oscillates between the two modes. The authors conclude: "This is auto-engineering, not auto-research."

Three Pathologies and Corresponding Skills

1. Novelty Misjudgment (新颖性误判)

Manifestation: Treats well-known tricks (e.g., cosine LR) as "creative breakthroughs".

Skill: Research Ideation ("mad scientist" persona + literature mining).

2. Resource Waste (资源浪费)

Manifestation: GPUs frequently idle; parallelizable experiments run serially.

Skill: Experiment Management (real-time monitoring + resource prioritization).

3. Exploration Collapse (探索坍缩)

Manifestation: After initial success, dives into local hyperparameter sweeps and ensembling.

Skill: Periodic Self-Checklist (hourly prompt: "Is this innovation or just tuning?").

Adding Research-Process Scaffolding Doubles Gains

Equipping agents with the above skills yields dramatic improvements:

Opus average improvement jumps from 10.4% to 20.1% .

Codex rises from 4.4% to 8.1% .

The standout case is PFDelta , where improvement reaches 91% . Forced to "think like a researcher," the agent recognized that the evaluation metric PBL (power balance loss) can be derived from Kirchhoff's laws without labels. It reformulated the training objective from supervised MSE to a primary PBL loss, and at inference used the GNN prediction as a "warm start" for the power-flow equations, iteratively refining voltage variables until physical residuals approached zero. This constitutes physics-informed reasoning , not mere hyperparameter tuning.

Conclusion: Bottleneck Is Research Taste, Not Compute

The paper argues that the primary barrier to AI-driven scientific discovery is not compute or parameter count, but research taste and methodology . Providing agents with explicit research-process prompts and scaffolding delivers immediate, measurable gains—far more effective than waiting for the next model generation. The next frontier for AI in science is increasing the proportion of genuine scientific insight in the improvements, not just chasing higher benchmark scores.

https://arxiv.org/pdf/2610.28763
Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?
https://github.com/maszhongming/agent4sci