Treating the Symptom, Missing the Cause: Anatomy of a Scientific Reasoning Failure

Anna Gogleva; Thomas Charman; Martin Balla; Karl Gemayel

cs.AINov 3, 2026

Treating the Symptom, Missing the Cause: Anatomy of a Scientific Reasoning Failure

Authors: Anna Gogleva, Thomas Charman, Martin Balla, Karl Gemayel

Organizations: Relation Therapeutics

Abstract

Scientific agents can find interventions that work without understanding why they do. We study this failure mode in controlled synthetic regulatory systems, where agents must experimentally identify a hidden fault and rescue the resulting phenotype. Across seven LLMs, every model restores the healthy phenotypes more often than it identifies the fault. Claude Opus 4.6 and Sonnet 4.6 rescue every organism while diagnosing only 55% and 35% of them. This failure is not simply due to insufficient exploration or unavailable evidence. In most cases, the cause is experimentally identifiable, and agents encounter evidence that distinguishes competing causes in 91% of relevant sessions, and still frequently misdiagnose the system. Targeted probes reveal where experimental reasoning breaks down. Models reliably interpret decisive evidence and revise hypotheses when the causal comparison is explicit, and can locate that evidence within a full experimental record. However, they struggle to predict the outcome of the experiment under each competing cause, a necessary step in using the observed result to distinguish between them. Agents can therefore gather the right evidence and still infer the wrong cause. As AI systems take on a growing role in synthesising scientific evidence, these findings highlight the need to evaluate how they use evidence to reach causal conclusions, alongside whether their interventions succeed.

Figures & tables

Figure 1: Controlled experimental environment. (a) A synthetic organism maps environmental inputs through an unknown gene regulatory network to three observable phenotypes; one example hidden fault is highlighted. Agents cannot observe the network directly and must learn about the system through perturbations. (b) Each session proceeds under a fixed experimental budget. All sessions include exploration and prediction; disease sessions additionally require diagnosis of a hidden fault followed by an intervention to rescue the phenotypes.

Panel of Figure 2 — Figure 2: The diagnosis–rescue gap. (a) A downstream gene can provide a plausible diagnosis and rescue target, restoring the phenotype without correcting the true fault. (b) Diagnosis and rescue at medium difficulty (20 seeds) for seven LLMs and three non-LLM baselines. Lines connect the two scores for each agent. CIs and prediction accuracy are in Appendix B.1 (Table 2 ).

Figure 3: Exploration does not simply explain diagnostic success. (a) Models differ substantially in both the number of experiments they run and the ratio of genes they probe. (b) Within each model, sessions above versus below its median experiment count show no consistent advantage from greater exploration.

Panel of Figure 4 — Figure 4: Alternative hypotheses and discriminating experiments. (a) Two different faults can produce the same abnormal readout, leaving both as plausible causes. (b) An experiment is discriminating if its outcome differs depending on which fault is present. Values in (a) and (b) are illustrative. (c) We classify the 60 organisms by whether an alternative fault fits the initial observation and, if so, whether an available experiment can distinguish it from the true fault.

Figure 5: Agents usually encounter discriminating evidence, but barely use it. Each point shows, for one organism with a resolvable alternative (

n=43

), how often agents choose discriminating experiments relative to random choice. Agents show only a small preference above chance (24.1% vs. 19.2%). Nevertheless, 91% of sessions encounter discriminating evidence before diagnosis, yet diagnosis accuracy in those sessions is only 0.30.

Figure 6: One agent holds the evidence and commits against it. Experiments 5–7 probe the agent’s eventual diagnosis, G008, in three directions. In each case, the observed readout of one phenotype (black) matches the true fault’s prediction (green) rather than G008’s (red). The agent nevertheless commits to G008 14 experiments later.

	true cause perturbed
model	never	one direction	both directions
Opus 4.6	0.00 (3)	0.00 (19)	0.84 (38)
GPT-4.1	0.00 (25)	0.06 (16)	0.84 (19)
Sonnet 4.6	0.00 (2)	0.00 (24)	0.71 (34)
Qwen 3.5	0.00 (8)	0.13 (30)	0.64 (22)
GPT-4o	0.00 (43)	0.08 (12)	0.60 (5)

Table 1: Diagnosis accuracy by how agents perturb the true cause , with session counts in brackets. Testing in both directions is associated with higher accuracy for every model. Five cells contain fewer than ten sessions, so individual estimates should be interpreted cautiously.

Figure 7: Evidence use breaks in context. (a) Each probe shows more of the investigation and gives less help with it: the two competing causes, the readout each one predicts, both, or neither. (b) Accuracy on decisive evidence, each gray line represents a model; per-model values are in Tables 5 and 6 . Accuracy stays close to 1.00 for as long as the predicted readouts are given (Probes 1–2 and 3a), falls once the models must work them out (3b–c), and falls further in live sessions.

Figure 8: Tracing the diagnosis–rescue gap through the investigation. Successive analyses rule out insufficient exploration and, for most systems, missing evidence, while bidirectional testing (dashed box) marks a correlate of successful diagnosis. Probes localise the remaining failure: models interpret and revise on isolated evidence, and still locate it inside a full record, but cannot derive what each competing cause predicts.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

agent	prediction acc.	diagnosis	rescue
Opus 4.6	$0.850\pm 0.056$	$0.550\pm 0.239$	$1.000\pm 0.000$
Sonnet 4.6	$0.917\pm 0.054$	$0.350\pm 0.229$	$1.000\pm 0.000$
GPT-4.1	$0.817\pm 0.084$	$0.250\pm 0.208$	$0.650\pm 0.229$
Qwen 3.5	$0.792\pm 0.066$	$0.200\pm 0.192$	$0.900\pm 0.144$
DeepSeek V3	$0.708\pm 0.061$	$0.200\pm 0.192$	$0.900\pm 0.144$
GPT-5.1	$0.783\pm 0.116$	$0.150\pm 0.171$	$0.400\pm 0.235$

Appendix

Table 2: Diagnosis, rescue, and prediction accuracy with 95% confidence intervals. Medium disease, 20 seeds per agent, mean

\pm

95% CI. Seven LLMs, then the three non-LLM baselines.

class	organisms	committed sessions	diagnosis [95% CI]
unambiguous	14	98	0.31 [0.16, 0.46]
resolvable ambiguity	39	273	0.30 [0.23, 0.37]
structural ambiguity	7	49	0.12 [0.00, 0.29]

Appendix

Table 3: Diagnosis accuracy by ambiguity class. Bootstrap 95% CIs clustered by organism (organisms resampled, not sessions), since sessions on one organism share a task.

model	experiments	budget used	coverage	bidir. rate	diagnosis	rescue
Opus 4.6	24.1	84%	1.00	0.20	0.53	0.98
Sonnet 4.6	22.9	78%	1.00	0.26	0.40	0.94
Qwen 3.5	21.5	75%	0.83	0.36	0.30	0.94
DeepSeek V3	18.3	67%	0.52	0.43	0.22	0.93
GPT-5.1	17.1	59%	0.91	0.23	0.15	0.46
GPT-4.1	12.7	47%	0.40	0.37	0.28	0.73

Appendix

Table 4: Exploration axes per model , mean per session over all 420 sessions (60 per model). Diagnosis and rescue are exact-match, pooled over the three disease sweeps.

model	P1 real	P1 filler	P2 revise	P2 retain	P2 filler	revise-on-noise
Opus 4.6	1.00 ( $\geq$ 0.95)	$0.35\pm 0.14$	1.00 ( $\geq$ 0.90)	1.00 ( $\geq$ 0.90)	$0.63\pm 0.14$	5/43
Sonnet 4.6	1.00 ( $\geq$ 0.95)	$0.44\pm 0.14$	1.00 ( $\geq$ 0.90)	1.00 ( $\geq$ 0.90)	$0.51\pm 0.14$	5/43
GPT-4.1	1.00 ( $\geq$ 0.95)	$0.33\pm 0.13$	1.00 ( $\geq$ 0.90)	$0.97\pm 0.07$	$0.63\pm 0.14$	5/43
GPT-4o	1.00 ( $\geq$ 0.95)	$0.30\pm 0.13$	1.00 ( $\geq$ 0.90)	$0.94\pm 0.09$	$0.40\pm 0.14$	13/43
GPT-5.1	1.00 ( $\geq$ 0.95)	$0.53\pm 0.14$	1.00 ( $\geq$ 0.90)	1.00 ( $\geq$ 0.90)	$0.35\pm 0.14$	8/43
DeepSeek V3	$0.97\pm 0.05$	$0.37\pm 0.14$	$0.94\pm 0.09$	$0.74\pm 0.14$	$0.58\pm 0.14$	10/43

Appendix

Table 5: Probes 1–2: item accuracy per model. Real items

n{=}69

(revise 34, retain 35), fillers

n{=}43

. Filler accuracy means answering UNDETERMINED; revise-on-noise counts filler items answered REVISE — switching hypothesis on an uninformative result.

model	3a (predictions given)	3b (predictions derived)	3c (open diagnosis)	live
Opus 4.6	1.00 ( $\geq$ 0.92)	$0.65\pm 0.14$	$0.58\pm 0.14$	$0.58\pm 0.14$
Sonnet 4.6	1.00 ( $\geq$ 0.92)	$0.58\pm 0.14$	$0.47\pm 0.14$	$0.44\pm 0.14$
GPT-4.1	$0.95\pm 0.07$	$0.70\pm 0.13$	$0.49\pm 0.14$	$0.28\pm 0.13$
GPT-4o	$0.98\pm 0.06$	$0.74\pm 0.13$	$0.37\pm 0.14$	$0.09\pm 0.09$
GPT-5.1	$0.91\pm 0.09$	$0.56\pm 0.14$	$0.42\pm 0.14$	$0.12\pm 0.10$
DeepSeek V3	$0.95\pm 0.07$	$0.51\pm 0.14$	$0.51\pm 0.14$	$0.16\pm 0.11$

Appendix

Table 6: Probe 3: the same evidence, inside an investigation record.

n{=}43

per cell,

\pm

95% Wilson half-widths; cells at the ceiling show the one-sided bound. Probe 3a supplies the candidate predictions for every logged experiment, 3b withholds them. The replayed logs come almost entirely from Opus 4.6 sessions, so 3c measures reasoning over that evidence, not what each model’s own investigation would have produced.

Figure 9: The full session record behind Figure 6 . Observed readouts (black) against the predictions of the true fault (green rings) and of the agent’s eventual diagnosis, G008 (red crosses); gray dots are multi-intervention calls, for which no single-fault prediction is defined. The shaded band marks the three G008 probes of Figure 6 ; the dashed line marks the commit; the red segment at experiment 22 is the post-commit refutation.

Scientific Reasoning in Language Models

Scientific reasoning in language models studies how AI systems use evidence, domain knowledge, and tools to solve problems and support discovery.

See all 161 papers

LUCA: A Synthetic World for Experimental Reasoning
Nov 2, 2026Anna Gogleva, Thomas Charman, Martin Balla +1Process-Based Agent Evaluation LLM Agent Evaluation
Relation Therapeutics
AI scientists produce results without reasoning scientifically
Apr 20, 2026Martiño Ríos-García, Nawaf Alampara, Chandan Gupta +5AI Agent Reliability Scientific Reasoning
Laboratory of Organic and Macromolecular Chemistry (IOMC), Friedrich Schiller University Jena, Humboldstr. 10, 07743 Jena, Germany · Department of Civil Engineering, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India · School of Interdisciplinary Research, Indian Institute of Technology Delhi, Hauz Khas, New Delhi 110016, India +3
CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists
May 25, 2026Junlin Yang, Dylan Zhang, Xiangchen Song +7Structural Causal Models LLM Agent Evaluation
Tsinghua University · University of Illinois Urbana-Champaign · Carnegie Mellon University +2

Nov 2, 2026J/K move · Enter open · S save

LUCA: A Synthetic World for Experimental Reasoning

Anna Gogleva, Thomas Charman, Martin Balla, Karl Gemayel

Relation Therapeutics

Process-Based Agent Evaluation LLM Agent Evaluation Autonomous Experimentation Scientific Reasoning AI Agent Benchmarks

Open paper AskPDF arXiv

An experimental agent can succeed for the wrong reasons. Evaluations, however, often focus on the outcome, not the experiments, evidence, and inferences that produced it. We introduce LUCA, a configurable synthetic world for evaluating this process under fully known ground truth. Agents learn unknown regulatory systems through budgeted experimentation, then predict unseen interventions, diagnose hidden faults, and attempt rescues. This setup allows us to evaluate experiment selection, prediction, diagnosis, and intervention separately. Across seven frontier LLMs and three baselines, these capabilities dissociate sharply. Agents frequently rescue malfunctioning systems while diagnosing the wrong fault; diagnosis varies 7x across models while prediction varies only 1.4x. Models also adopt surprisingly different experimental strategies, and greater exploration does not reliably yield better performance. In one failure mode, agents often perform the experiment needed to distinguish competing causes, but still fail to use the resulting evidence. LUCA turns these hidden failures into measurable behaviours, providing a controlled environment to study where experimental agents succeed and fail. Code: https://github.com/RelationRx/luca.

Abstract

Scientific Reasoning in Language Models

Explore similar work

LUCA: A Synthetic World for Experimental Reasoning

AI scientists produce results without reasoning scientifically

CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists