Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Let the Abyss Stare Back Adaptive Falsification for Autonomous Scientific Discovery

Peiran Li, Fangzhou Lin, Shuo Xing, Jiashuo Sun, Dylan Zhang, Siyuan Yang, Chaoqun Ni, Zhengzhong Tu

Abstract

Autonomous scientific discovery is entering a more dangerous regime: once the evaluator is frozen, a sufficiently strong search process can learn to win the exam without learning the mechanism the task was meant to reveal. This is the idea behind our title. To let the abyss stare back is to make evaluation actively push against the candidate through adaptive falsification, rather than passively certify it through static validation. We introduce DASES, a falsification-driven framework in which an Innovator, an Abyss Falsifier, and a Mechanistic Causal Extractor co-evolve executable scientific artifacts and scientifically admissible counterexample environments under a fixed scientific contract. In a controlled loss-discovery problem with a single editable locus, DASES rejects artifacts that static validation would have accepted, identifies the first candidate that survives the admissible falsification frontier, and discovers FNG-CE, a loss that transfers beyond the synthetic discovery environment and consistently outperforms CE and CE+L2 under controlled comparisons across standard benchmarks, including ImageNet.

Let the Abyss Stare Back Adaptive Falsification for Autonomous Scientific Discovery

Abstract

Autonomous scientific discovery is entering a more dangerous regime: once the evaluator is frozen, a sufficiently strong search process can learn to win the exam without learning the mechanism the task was meant to reveal. This is the idea behind our title. To let the abyss stare back is to make evaluation actively push against the candidate through adaptive falsification, rather than passively certify it through static validation. We introduce DASES, a falsification-driven framework in which an Innovator, an Abyss Falsifier, and a Mechanistic Causal Extractor co-evolve executable scientific artifacts and scientifically admissible counterexample environments under a fixed scientific contract. In a controlled loss-discovery problem with a single editable locus, DASES rejects artifacts that static validation would have accepted, identifies the first candidate that survives the admissible falsification frontier, and discovers FNG-CE, a loss that transfers beyond the synthetic discovery environment and consistently outperforms CE and CE+L2 under controlled comparisons across standard benchmarks, including ImageNet.

Paper Structure

This paper contains 23 sections, 19 equations, 1 figure, 4 tables.

Table of Contents

  1. Introduction
  2. DASES - Dynamic Adversarial Scientific Environment Synthesis and Mechanistic Co-Evolution
  3. Scientific Task as a Constrained Falsification Problem
  4. DASES as a Closed Scientific Game
  5. Dynamic Adversarial Environment Synthesis
  6. Mechanistic Co-Evolution and Survival Semantics
  7. Scientific Contract: Integrity, Anti-Cheating, and Auditability
  8. Round-Based Orchestration and Scope
  9. Experiments
  10. DASES instantiation for loss discovery
  11. A deceptive but lawful discovery environment
  12. Falsification-driven co-evolution in the discovery lab
  13. The artifact selected by the falsification trace
  14. Results in the discovery lab
  15. Transfer beyond the discovery environment
  16. ...and 8 more sections

Figures (1)

  • Figure 1: DASES co-evolution trajectory in the discovery lab. The static ID validation accuracy remains comparatively high throughout the search, but each falsifier expansion of the lawful OOD frontier induces a sharp test-score drop that exposes a new hidden failure mode. CE first mitigates background shortcut reliance, CE$+$L2 further improves robustness under geometry-focused stress, and FNG-CE is the first candidate to reach the highest plateau while remaining essentially stable under the final falsifier expansion.