The Greatness of Science Cannot Be Planned: Agentic Auto-Research is Fuzz Testing

The paper argues that autonomous research agents should be viewed as operating under the same control loop as coverage-guided greybox fuzzing. The authors contend that the current generate-and-rank paradigm, which scales generation and uses learned judges to rank samples, misses the problem of sparse feedback. The paper's central position is that auto-research should be instrumented, feedback-directed search, where each experiment exposes a cheap, dense signal of epistemic progress, the search policy uses that signal to choose its next intervention, and final validation, protected from adaptive reuse, determines what counts as a discovery.

The paper draws a direct analogy between fuzzing and auto-research, mapping components such as the seed corpus to prior results and literature, mutation operators to LLM proposers, coverage to intermediate progress signals, and the crash/sanitizer oracle to protected validation. The authors state, "Fuzzing reveals the missing design choice. An autonomous research agent generates a candidate experiment, executes it, observes feedback, and chooses what to try next. Coverage-guided greybox fuzzing runs the same control loop."

The paper derives three main lessons from fuzzing. The first lesson is to make research progress observable, arguing that final validation gives sparse feedback and that an intermediate signal is needed to report partial progress on each candidate. The second lesson is to use feedback to search, meaning the next experiment should depend on what previous experiments revealed, rather than only ranking completed samples. The third lesson is that guidance is not a verdict, requiring a protected validation step to certify discoveries, as an optimized progress signal cannot certify a finding.

The paper makes three separable, falsifiable predictions. First, "Signal prediction: with the proposer, task, and candidate pool fixed, a useful intermediate metric should predict protected outcomes and select more validator-confirmed discoveries under a fixed validation budget than an uninformative or terminal-only signal. Second, Search prediction: with the proposer, task, protected validator, and total budget fixed, a policy that uses intermediate experimental feedback to choose its next intervention should yield more validator-confirmed discoveries per unit cost than repeated sampling. Third, an epistemic benefit prediction: with the candidates and search trace fixed, evidence protected from adaptive reuse should reduce the false-discovery rate relative to reporting the optimized guidance score."

The paper discusses related work, including the AI co-scientist, which uses an LLM tournament (Elo ratings) for guidance and physical experiments for final validation, and systems like FunSearch and AlphaEvolve that use machine-executable objectives. It also addresses six alternative views, including objections about LLMs' abductive capabilities, the relationship to Bayesian optimal experimental design, the lack of a perfect oracle in research, the bitter lesson about scaling computation, and the instability of research search spaces. The conclusion reiterates that the position stands or falls on the three controlled comparisons and locates the human role in declaring the problem and grounding the verdict, while the agent runs the loop between them.

Improvements to AI systems:

1. Add an intermediate epistemic progress metric to any autonomous research agent. The system should compute a cheap, dense signal (e.g., novelty of intermediate results, reduction in hypothesis entropy, coverage of unexplored sub-questions) after every experiment, not just at the end. This enables the agent to report partial progress and prioritize experiments that reduce uncertainty, even if they don't yet yield a final discovery.

2. Replace generate-and-rank with a feedback-directed search policy. The system’s next experiment should be selected based on the observed intermediate signal from prior experiments (e.g., which sub-hypotheses were falsified, which parameter regions showed unexpected variance), rather than merely re-ranking a static pool of candidates. Concretely, the agent should maintain a dynamic search tree where each node is an experiment, and the policy chooses the next node with the highest expected information gain per unit cost.

3. Implement a protected validation oracle separate from the guidance signal. The system must keep two distinct pipelines: (a) a fast, noisy guidance metric used only for search (e.g., a learned proxy for plausibility), and (b) a slow, rigorous, and immutable validator (e.g., a held-out experimental protocol or a formal proof checker) that is never exposed to the search policy. The final discovery claim must be based solely on the protected validator, and the system must log all adaptive reuse of the validator to prevent overfitting.

4. Add a coverage map of the research space. The system should track which regions of the hypothesis space have been explored, which have shown high variance in intermediate signals, and which remain unexplored. This coverage map guides the search policy to prioritize unexplored or high-uncertainty regions, analogous to coverage-guided fuzzing’s edge coverage.

5. Integrate a mutation operator that proposes experiments by perturbing prior results or literature. Instead of generating from scratch, the system should use LLM proposers that take the current best hypotheses, prior experimental outcomes, and the coverage map as input, and propose mutations (e.g., change a parameter, combine two hypotheses, negate a condition) that are likely to produce new intermediate signals.

6. Add a budget-aware stopping criterion. The system should decide when to stop searching based on the rate of new intermediate signal acquisition (e.g., if the coverage map stops expanding or the entropy reduction per experiment falls below a threshold), and then switch to the protected validation phase for the top candidates.

7. Implement a false-discovery-rate controller. The system should track the number of times each candidate was selected based on the guidance signal, and when reporting final discoveries, apply a correction (e.g., Benjamini-Hochberg) that accounts for the adaptive reuse of the guidance signal. This reduces the chance of reporting a false positive that merely overfits the noisy metric.

What the improved AI system can do:

• Run long-horizon autonomous research (e.g., drug candidate screening, materials discovery, theorem exploration) with a clear, measurable notion of making progress at every step, not just at the end.

• Dynamically reallocate its compute budget toward experiments that reduce uncertainty the most, rather than wasting resources on redundant or low-information candidates.

• Produce final discoveries that are statistically robust against the bias introduced by its own search process, because the validation step is protected from the guidance signal.

• Explain its search trajectory by showing the coverage map and the intermediate signals that drove each decision, making the agent’s reasoning auditable and interpretable to human researchers.

• Scale to large research spaces where terminal-only validation is too sparse or expensive, by using cheap intermediate signals to prune the search space efficiently.

• Avoid the false discovery trap common in LLM-based research agents that report the highest-scoring candidate from a noisy judge, by separating guidance from certification.

Autonomous research agents can generate experiments faster than researchers can validate them. Researchers have responded by scaling the proposer and ranking more samples with a learned judge or human reviewers. We argue that this *generate-and-rank* paradigm misses the problem of sparse feedback. Within a declared research problem, an agent follows the control loop of a greybox fuzzer: it proposes a candidate, executes it, observes feedback, and chooses what to try next. A fuzzer rarely finds a bug, but coverage makes partial progress observable on every execution. Fuzzers then use that signal to mutate inputs and allocate effort, rather than only to rank completed runs. Auto-research needs the same two capabilities. First, each experiment should expose a cheap, dense signal of epistemic progress before final scientific validation is available. Second, that signal should determine the next intervention so that the agent searches rather than repeatedly samples. Because the optimized progress signal is guidance rather than a verdict, final validation must still decide what counts as a discovery using evidence protected from adaptive reuse. We propose controlled tests of whether candidate signals predict validated progress, whether feedback-directed search yields more validated discoveries per unit cost than repeated sampling, and whether protected validation reduces false discoveries. Feedback architecture, not only generation, is a central bottleneck in auto-research.