AI evidence extraction

pAI/MSc: ML Theory Research with Humans on the Loop

Mahmoud Abdelmoneum, Pierfrancesco Beneventano, Tomaso Poggio · arXiv (Cornell University) · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

2/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Technical system design and engineering with components: (1) modular multi-agent architecture spanning 23 agents across six pipeline phases; (2) artifact contract framework defining intermediate outputs and validation gates; (3) quality-enhancing mechanisms including persona council debate, adversarial novelty falsification, theory-experiment independence, reviewer hard blockers, multi-model counsel, and tree search over proof strategies; (4) optional modules for reasoning and optimization; (5) operational model with human-on-the-loop oversight and checkpoint/budget tracking..

Primary method

Design science research; engineering-first approach with iterative refinement based on operational failures and lessons learned

Main result

The system achieves a practical compression of human steering burden in AI-assisted academic research. As stated in the paper: "Can we design an AI system that, in at most 10 human steers, pushes a strong Hypothesis to a Written Article of serious academic quality?" The authors report that "moving from a strong human-developed idea to a solid machine-learning-theory manuscript still often takes on the order of 10³ (or 5·10⁴, the baseline defined by [27]) prompts—if possible at all—with frontier reasoning models and agentic systems." Key findings include that single-agent ideation converges too quickly, structured debate improves planning quality, theory-experiment independence produces stronger outputs, and long-horizon runs require explicit stopping criteria.

Research paradigm

Design science / Engineering research

Author conclusions

"The key constraint is high quality, the objective function to minimize is the human steering burden (or oversight) required to reach that." The authors conclude that their system achieves "an artifact contract in which progress is represented by named intermediate outputs and stage-level validation gates rather than by free-form dialogue alone." They note: "Academic research, especially theory, weakens all three assumptions" (bounded domain, short evaluation loops, automatic objectives) found in successful prior agentic systems, requiring their novel approach. Critically: "pAI/MSc is our attempt to do exactly this. System prompts and control flows are designed to make novelty claims explicit, stress-test them against competing explanations, align mathematics with experiments, and demand intervention experiments when causal language is involved."

Risk of bias

Model selection bias: system defaults to Claude Opus 4.6, GPT-5.4, and Gemini 3 Pro Preview, which may embed biases present in those frontier models; Training data bias: LLM-based agents inherit biases from their training corpora; Evaluation bias: automated reviewer agent may have systematic blind spots regarding novel research directions; Confirmation bias: persona council debate structure could converge toward predetermined positions rather than genuine exploration; Cost-driven bias: expensive compute limits exploration breadth, potentially biasing toward efficiency over novelty; No user study validation; system claims are based on engineering design rather than empirical testing with human researchers; Evaluation framework is incomplete - authors acknowledge "Toward stronger evaluation" as an open problem; Potential confirmation bias in design of rigor mechanisms (e.g., adversarial novelty falsification may be calibrated toward author assumptions about research quality); Limited scope: focused specifically on machine learning theory rather than general academic research; No comparison with actual human-conducted research workflows or gold-standard acceptance rates

Limitations

The authors explicitly state: "What the system guarantees and what it does not" (Section 6.1)
Key limitations include: "Parallel debugging remains expensive and only partially solved" - when failures occur "after several hours of execution, the cost of diagnosis can be substantial." The system "can validate artifact existence, parseability, selected internal consistency properties, and review-gate thresholds
It cannot certify scientific correctness, novelty, or acceptance probability." Regarding automation: "Our focus, thus, is developing the best possible system for what we see as a key necessary step toward that goal" (of full automation), suggesting the current system remains limited
"Long-horizon runs need explicit stopping criteria" as "a run can continue to generate text while adding little scientific value." Additionally, "no short-horizon scalar objective faithfully captures novelty, correctness, and long-run scientific value."

Open questions raised

How to design better computable proxies for research quality without reducing research to the proxy
Solving parallel debugging in long, partially parallel research workflows
Developing reliable stopping criteria for long-horizon runs based on conceptual (not just token-based) diminishing returns
Full automation of research ideation and hypothesis generation (explicitly marked as future work: "Yet")
Fully automated path from idea to article (marked as future work: "Yet")
Better evaluation frameworks for manuscript-level research quality rather than capability benchmarks

Code: github.com/PoggioAI/PoggioAI_MSc (primary system code); Claude Code Skill PoggioAI_MSc-claude (alternative implementation, noted as producing lower quality); https://github.com/PoggioAI/PoggioAI_MSc; github.com/PoggioAI/PoggioAI_MSc (main code repository); PoggioAI_MSc-claude (Claude Code Skill, noted as producing lower quality manuscripts)Extracted from: pdfAgreement 63%

Explore related topics

AI Manuscript Writing Academic Writing Design Science