Can AI Scientists Change Their Minds? Prior-Evidence Conflict in Synthetic Universes

Kargi Chauhan

cs.AISep 29, 2026

Can AI Scientists Change Their Minds? Prior-Evidence Conflict in Synthetic Universes

Authors: Kargi Chauhan

Organizations: University of California, Santa Cruz

Abstract

Can a scientific agent distinguish a law it inferred from evidence from one it merely recognizes? We introduce Synthetic Universes, a controlled benchmark that pairs canonical famous worlds with matched twisted twins governed by nearby noncanonical mechanisms. We evaluate each reported law twice: by executing it on held-out continuations and transfer settings, and by independently checking whether it recovers the generating mechanism. In the current checkpoint of a pre-specified 60-cell study, 22 trials were graded and one additional run ended in infrastructure failure. Among 20 twin trials, 8 pass predictive verification while 5 recover the generator. The dissociation is bidirectional: six parsable outputs predict successfully while missing the mechanism, whereas three recover the mechanism but fail predictive rollout. Drag exhibits the first pattern (5/5 predictive pass, 1/5 mechanism recovery); Gravity exhibits the second (1/5 predictive pass, 4/5 mechanism recovery). Because matched famous controls, the corrected identifiability sweep, and the Evidence Ladder remain incomplete, we do not claim a confirmatory causal prior-conflict effect. Instead, the completed runs establish a narrower verification result: predictive adequacy and mechanism recovery are distinct scientific claims and require distinct tests.

Figures & tables

Figure 1: Synthetic Universes separates predictive verification from mechanism recovery. The agent sees observations but not the hidden law. Its reported equation is executed on an unseen continuation and a transfer world, then independently checked against the generating mechanism. A predictive pass therefore certifies performance on the tested interventions, not mechanism discovery.

Family	Famous mechanism	Twisted mechanism
Spring	$-k\,\mathrm{sgn}(x)\|x\|^{p}$ , $p=1$	$-k\,\mathrm{sgn}(x)\|x\|^{p}$ , $p=1.3$
Gravity	$F_{r}\propto r^{-p}$ , $p=2$	$F_{r}\propto r^{-p}$ , $p=2.3$
Drag	$-k\\|\mathbf{v}\\|^{p-1}\mathbf{v}$ , $p=1$	$-k\\|\mathbf{v}\\|^{p-1}\mathbf{v}$ , $p=1.6$
Pendulum	$-\omega^{2}\sin\theta$	$-\omega^{2}\sin\theta(1+0.4\|\theta\|)$
Conservation †	audited target invariant class	generator under audit
Coupling †	audited pair-interaction class	generator under audit

Table 1: Six benchmark families. Continuous families are written using a common exponent

p

. The dagger marks structural families retained for transparent run accounting but not used for strong mechanistic claims until their generator and non-degeneracy audit is complete.

	Predictive pass		Mechanism recovery
Family	Famous	Twin	Famous	Twin
Spring	1/1	0/5	1/1	0/5
Gravity	1/1	1/5	1/1	4/5
Drag	–	5/5	–	1/5
Pendulum	–	1/2	–	0/2
Conservation	–	1/2	–	0/2

Table 2: Submission-time outcome accounting. Each entry is successes/graded trial records for that cell. Seeds index matched deterministic world-generation randomness, including initial conditions and configured observation perturbations. Parser failures remain visible in the seed matrix; dashes denote cells with no graded record, not zero performance.

Figure 2: Seed-level submission-time status. Missing, parser, and infrastructure cells remain explicit rather than being converted into scientific failures.

Figure 3: Prediction and mechanism recovery dissociate. Counts sum to the 20 parsable graded trials; the two unparseable records from the 22 graded trials are excluded from this four-cell decomposition and remain visible in Fig. 2 .

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Artifact	Audit question
Task + observations	What evidence was actually available to the agent?
Raw response + tool trace	Which hypotheses and calculations were externally visible?
Predictive grader output	Did the reported law pass continuation and transfer?
Mechanism-checker output	Did the executable law match the generating equivalence class?
Parser regression tests	Could notation handling change a scientific label?
Failure logs	Was a missing result scientific, representational, or infrastructural?

Appendix

Table 3: Release artifacts and what they make auditable. The point of the release is not only rerunning a headline score, but reconstructing why a trial received its label.

Family	Scientific distinction	Audit quantity
Spring	restoring response $p=1$ vs. $p=1.3$	matched power $p$ ; functional residual
Gravity	central-force exponent $p=2$ vs. $p=2.3$	matched exponent $p$ ; functional residual
Drag	velocity exponent $p=1$ vs. $p=1.6$	matched exponent $p$ ; functional residual
Pendulum	sinusoidal response vs. amplitude-dependent correction	fixed-domain functional residual / shape fit
Conservation	invariant equivalence class	normalized structural/equivalence residual
Coupling	pairwise interaction structure	product-vs.-alternative structural check

Appendix

Table 4: Mechanism quantities used by the frozen checker. The functional residual gate is

0.05

where applicable; the released source contains the fixed family-specific shape gates and evaluation domains.

Figure 4: Family-level mechanism recovery in the current checkpoint. The figure is descriptive because famous controls are mostly missing. Its main value is diagnostic: Gravity and Drag exhibit opposite relationships between predictive success and mechanism recovery, which is why the paper reports the two criteria separately rather than averaging them into one discovery score.

Figure 5: Failure localization used by Synthetic Universes . Stage 1 asks whether the available evidence discriminates candidate families; Stage 2 asks which law the agent actually reports; Stage 3 executes that law externally. The four-quadrant outcome is therefore the end of a diagnostic chain, not the whole diagnosis.

Figure 6: Evidence Ladder design: change the evidence, not the law. The intervention is intended to move the same hidden mechanism from weakly discriminative to strongly discriminative observations. A completed positive result would require recovery to increase with evidence strength; the schematic itself is not evidence for that trend.

Figure 7: First-pass outcomes ( $k=1$ per condition). These results motivated the replicated matrix and the prediction–mechanism taxonomy. They are shown for provenance, not pooled with the submission-time replicated counts.

Figure 8: First-pass Gravity visible-prefix transition. The reported exponent moves from the familiar value toward the twisted law as more visible analysis is retained. Because this is a single pilot trajectory and operates on visible text rather than hidden reasoning, it is not used to support the paper’s causal claims.

Claim	Current support	What would weaken or falsify it
Predictive adequacy and mechanism recovery are distinct verification targets	Both off-diagonal quadrants are populated; Drag and Gravity show replicated opposite dissociations.	A corrected evaluator that makes the off-diagonal cases disappear, or evidence that their labels arise from parser/grader artifacts.
Paired famous/twin worlds provide a controlled way to study prior–evidence conflict	The observation interface is shared while the hidden mechanism changes.	A demonstrable cue that reveals world identity, or contamination/leakage that gives access to generator labels.
The present data establish the magnitude of a prior-conflict penalty	Not claimed. Famous controls are incomplete.	Requires completion of the matched grid and the pre-specified paired analysis.
Increasing discriminative evidence causes prior abandonment	Not claimed. The Evidence Ladder is incomplete.	A flat or reversed recovery trend under increasing complexity-adjusted oracle evidence would directly challenge this interpretation.
The observed twin failures are specific to pretrained language-model priors	Not claimed. No prior-free or cross-model baseline is complete.	A symbolic baseline showing comparable failures, or strong dependence on agent architecture, would weaken a pretraining-prior explanation.
The uniform rollout threshold fully characterizes scientific correctness	Not claimed. Rollout and mechanism are deliberately separate.	Large agreement on derivative/geometric diagnostics despite raw rollout failure would show that pointwise long-horizon NRMSE is too coarse as a standalone verifier.

Appendix

Table 5: What the current submission claims, and what would overturn it.

Explore similar work

Position: Correct Answer, Wrong Mechanism -- When AI Scientists Defend General Claims Their Own Data Contradicts
Jun 22, 2026Steven Young EuligArtificial Intelligence Scientists Correct
Department of Physics and Laboratory for Particle Physics and Cosmology (LPPC), Harvard University, Cambridge, MA 02138, USA.
Can AI Agents Synthesize Scientific Conclusions?
Jun 9, 2026Hayoung Jung, Pedro Viana Diniz, José Reinaldo Corrêa Roveda +5Scientific Discovery Artificial Intelligence Agents
Princeton University · Universidade Federal de Minas Gerais · Stony Brook University +1
MechBench: Can AI Scientific Agents Discover Mechanisms Beyond Phenomenal Laws?
Sep 28, 2026Zihan Yu, Jiadong Zhang, Jialin Cheng +2Scientific Agents
Department of Electronic Engineering, BNRist, Tsinghua University, Beijing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · Beijing Zhongguancun Academy, Beijing, China +2

Jun 22, 2026, cs.LGJ/K move · Enter open · S save

Position: Correct Answer, Wrong Mechanism -- When AI Scientists Defend General Claims Their Own Data Contradicts

Steven Young Eulig

Department of Physics and Laboratory for Particle Physics and Cosmology (LPPC), Harvard University, Cambridge, MA 02138, USA.

Artificial Intelligence Scientists Correct Answer

Open paper AskPDF arXiv

AI scientist systems are described as tools, coauthors, or founders, but we evaluate them as if only the final answer matters. This position paper argues that outcome-only evaluation is insufficient, and that task outcome, mechanism fidelity, and epistemic honesty must be measured separately. Our evidence comes from 28 episodes of a coding agent attempting to rediscover a known particle identification observable in a Geant4 simulation, including an 8-episode probe across two additional frontier models. In 4/20 primary-model and 3/8 cross-model episodes, agents reach right-looking results through incorrect reasoning that breaks when conditions change, which we call Correct Answer, Wrong Mechanism (CAWM). Honesty and mechanism fidelity dissociate within a single agent trajectory. When given a partially misleading prior, all five agents reject the false component on evidence, yet one defends its chosen observable with physics inconsistent with its own data. In the simulation-based discovery setting studied here, coding agents prove reliable tools but unreliable scientific co-authors for open-ended claim-making, where co-author trust requires mechanism-fidelity verification they do not reliably self-apply. The failure is detectable, and we propose a lightweight test. A one-step regime-shift check needs only the agent's claim and flags the over-generalized cases. A companion recomputation flags the remaining cases when the correct observable is known. Together, these checks flag every CAWM case in this study.

Abstract

Explore similar work

Position: Correct Answer, Wrong Mechanism -- When AI Scientists Defend General Claims Their Own Data Contradicts

Can AI Agents Synthesize Scientific Conclusions?

MechBench: Can AI Scientific Agents Discover Mechanisms Beyond Phenomenal Laws?