A SELF-CORRECTING AI AGENT FOR AUTONOMOUS SCIENTIFIC LITERATURE REVIEW AND HYPOTHESIS GENERATION

Abstract

Though large language model (LLM) agents are used to review scientific literature, they tend to quote information inaccurately, cite nonexistent data, and offer unsupported hypotheses. In this paper, we introduce a SCLA (Self-Correcting Literature Agent) that implements the iterative process generate–verify–repair to conduct literature review and hypothesis generation, which is realized as a modular agent. SCLA extracts claims in a text with evidence, gets them checked by an independent critic, fixes or drops claims that are not supported by evidence, remembers failure patterns in a reflective memory, and filters novel hypotheses that are not found in previously retrieved work. After R repair rounds, we obtain a closed-form approximation of the residual error and test it against a Monte Carlo simulation of 120 research questions across six computer science subfields (5 runs each) against SCLA and five baselines, with one being a strong single-pass retrieval-plus-verifier approach. SCLA achieved claim faithfulness of 94.2% (95% CI [93.8, 94.6]) compared to 87.4% for the strongest baseline, and a rate of 78.4% of grounded and novel hypotheses compared to 46.5%. The verifier contributes the majority of the gains in faithfulness over the unverified systems, while the gains over the single-pass verification include the multi-round loop, re-retrieval and novelty screening. In all of the 400 random parameter settings chosen, SCLA received an overall advantage (median gain was 7.3 points above the best baseline option), with an advantage of approximately 1.7 times its token cost. All parameters are assumed and not measured which makes the results a snapshot of the architecture in absence of these parameters, and encourages evaluation of real models.