AI Night-Scientist, adaptive alloy optimization, TCSAlgBench proofs

Reinforcement Learning for Scientific Ideation: Researchers created AI Night-Scientist, a framework that trains language models using Group Relative Policy Optimization to diverge from predictable reasoning. By categorizing ideation across action, process, and outcome dimensions, the model increased the volume of explored research directions by 27.8 percent and improved proposal originality scores by 66.2 points over the baseline.
Adaptive Bayesian Optimization in Materials Design: A study assessed language-model-based adaptive policies for multi-objective alloy discovery under shifting laboratory budgets and timelines. Tested on high-entropy and soft magnetic alloys, the adaptive routines adjusted batch allocations using real-time campaign data, reducing predictive uncertainty on Pareto-optimal compositions without decreasing overall search speed.
Automated Proof Coverage on TCSAlgBench: Testing across 398 challenges drawn from the STOC and COLT 2026 conferences evaluated agent performance in theoretical computer science. An agent workflow using planning with GPT-5.5 xhigh achieved a verifier-accepted proof coverage rate of 25.4 percent, while a 10-round discussion setup using GPT-5.6 Sol max attained 23.6 percent.
Evidence Governance in Research Search Spaces: Researchers introduced EPOCH, an architecture designed to regulate how AI research agents evaluate mathematical and algorithmic solutions. The system incorporates formal task contracts, typed memory, falsification protocols, admission criteria, and replay verification, yielding mean normalized scores of 0.65 on AlgoTune and 0.57 on Math14.
Multi-Stage AI-Augmented Peer Review: An evaluation of an AI-augmented review workflow separated tasks into distinct stages focused on mathematical proofs, framework concordance, chemical validity, and revision checks. Applied to a theoretical chemistry manuscript, the system issued 64 review actions that achieved a 93.7 percent author adoption rate across revision cycles.
Agent-Directed Polymer Discovery: A proposed framework for self-driving polymer laboratories uses a central coordinator to parse natural language research targets. The coordinator assigns tasks across specialized sub-modules handling property prediction, synthesis paths, and solvent selection to manage closed-loop experimentation.
Autonomous Photovoltaic Screening: Researchers developed PVmatAgent, a system linking 11 computational tools across nine modules to automate photovoltaic material analysis, from structural retrieval to device performance calculation. The agent incorporates a truncation protocol that stops execution when ungrounded numerical values are generated.
Systematic Review of Scientific Agents: A published systematic review categorized the integration and operational roles of autonomous AI agents within scientific data workflows.

Sources

Reinforcing Agentic Creativity in Scientific Ideation with Night Science
Large language models (LLMs) excel at structured, verifiable tasks, but their low-entropy bias can produce homogeneous and predictable outputs, limiting their utility for open-ended scientific ideation. Effective discovery, however, spans a broader creative spectrum: from structured day science to loosely structured, serendipitous night science that reaches ideas beyond those typically considered. We introduce AI Night-Scientist, an agentic framework that uses reinforcement learning to teach models when and how to depart from predictable reasoning. Grounded in cognitive science, we model…
Agentic Resource Allocation for Batch Multi-Objective Bayesian Optimization in Autonomous Materials Discovery
The discovery and development of advanced materials is a challenging process constrained by the high time and monetary costs of synthesis, processing, and characterization. The underlying design spaces can be enormous, often with multiple competing objectives. Bayesian optimization (BO) provides a principled approach for efficiently navigating such spaces, but most workflows rely on fixed exploration-exploitation policies that lack the capacity to adapt to shifting constraints in dynamic campaigns typical of self-driving laboratories. In this work, we develop a multi-objective BO framework…
TCSAlgBench: Benchmarking Automated Proving for Research-Level Theoretical Computer Science
Large language models perform strongly on competition mathematics, but their research-level reasoning remains difficult to evaluate systematically. Theoretical computer science (TCS) connects algorithm design to explicit guarantees and fundamental limits, providing a setting for evaluating whether models can justify computational improvements with arguments humans can inspect. We introduce TCSAlgBench, a benchmark and reusable pipeline for natural-language proof discovery, comprising 398 theorem-level challenges from 138 STOC and COLT 2026 papers. Expert-designed rules complete paper-specific…
EPOCH: Reliable Discovery through Evidence-Governed Search
AI research agents are increasingly used to search over programs, mathematical constructions, and proofs. However, existing systems typically optimize evaluator feedback without adequately governing how that feedback is interpreted, challenged, and reused. As a result, promising but fragile candidates can be promoted as discoveries, while benchmark improvements, finite certificates, and theorem-level claims are too easily conflated. We introduce EPOCH, an evidence-governed architecture designed to close this gap. EPOCH implements an evidence-governed discovery loop by combining explicit task…
AI-Augmented Peer Review by Orthogonal Constraint: A Retrospective Case Study of Mathematical, Concordance, Chemical, and Meta-Review
The increasing use of artificial intelligence in scientific writing and mathematical formalization creates a corresponding need for review methods capable of validating rapidly generated, multidisciplinary technical work. We present and retrospectively evaluate an AI-augmented peer-review architecture in which distinct review stages are assigned orthogonal objectives: mathematical correctness, concordance with an established formal framework, chemical admissibility of molecular claims, targeted verification of revisions, and final meta-review. The method is examined through the development of…
A roadmap for polymer informatics super-intelligence
Polymer informatics has matured from isolated property-prediction studies into an integrated discipline that couples data, models, and decision-making across the polymer design cycle. Yet it still falls short of a true intelligent system capable of inverse design on demand, causal reasoning across chemistry, processing, and performance, and closed-loop autonomous experimentation. This article traces a roadmap toward that goal, grounded in experience developing two complementary agentic and informatics platforms. Central to this vision is a modular, agent-directed architecture in which a…
PVmatAgent: A Large Language Model (LLM) Agent for Perovskite Photovoltaic Material Design and Analysis
ABSTRACT The rapid advancement of large language models (LLMs) has opened new opportunities for materials informatics. However, LLMs fall short in photovoltaic (PV) material design due to their lack of domain grounding, unreliable outputs, and inability to perform integrated computational tasks. To address these issues, this work presents PVmatAgent, an autonomous LLM‐based computational agent designed specifically for PV materials design and analysis. The system integrates 11 domain‐specific tools (organized into 9 functional modules), including the machine learning force field CHGNet for…

Also this week

From Raw Scientific Data to Trusted Discovery: Agentic AI for FAIR and Reproducible Data Science: A Systematic Review of Data Engineering, Semantic Data Models, Multimodal Analytics, Provenance, and Human Oversight

Full transcript

What does it take to push automated models beyond predictable hypotheses during scientific discovery? That question is our focus today on Agents in Research, tracking artificial intelligence across the sciences. We start with AI Night-Scientist. When language models handle early-stage scientific ideation, standard training pushes them toward low-entropy outputs. They reproduce expected solutions instead of testing unfamiliar territory. Which is the problem AI Night-Scientist targets. The framework structures creative generation across three axes from cognitive science: action, process, and outcome. And it relies on Group Relative Policy Optimization to steer models toward uncharacteristic hypothesis spaces while retaining structured reasoning. On benchmarks, that approach broadened the scope of explored research directions by 27.8 percent, alongside a 66.2-point rise on proposal originality metrics relative to the baseline. Moving past hypothesis generation into formal verification, we have TCSAlgBench evaluating how models navigate mathematical proofs. That benchmark collates 398 theorem-level problems taken from 138 STOC and COLT 2026 papers. Under matched computational budgets, structure mattered as much as raw model size. With GPT-5.5 xhigh, a planning-focused workflow led the five-run verifier-accepted proof coverage at 25.4 percent. Outperforming setups that relied on task decomposition or direct prover-verifier dialogue. Though among models operating without external scaffolding, GPT-5.6 Sol max reached 23.6 percent verifier acceptance across a ten-round discussion setup. Even so, raw acceptance rates hide an ongoing failure mode: models treating finite empirical certificates or local benchmark improvements as formal mathematical proof. Which brings us to the EPOCH architecture. It places the search loop behind explicit task contracts, typed memory, active falsification routines, admission checks, and independent replay verification. It forces the system to verify findings independently rather than trusting intermediate metrics. In testing, EPOCH recorded a 0.65 mean normalized score on AlgoTune, 0.57 on Math14, and generated verified solutions for ten open discovery challenges. That separation of verification stages also appears in manuscript evaluation. Researchers tested a multi-stage review pipeline on a theoretical chemistry paper, splitting the work into orthogonal checks: mathematical proofs, chemical validity, revision tracking, and concordance with established formal frameworks. The distinct modules showed different correction profiles. The mathematical agents located formal flaws in derivations, while the concordance checkers flagged departures from defined structural rules. Out of 64 discrete review actions executed, authors adopted 93.7 percent across revision cycles. In physical experimentation, the focus shifts to operating within hardware constraints. A recent study examined language-model-driven adaptive policies managing multi-objective Bayesian optimization during sudden disruptions, like unexpected budget drops or truncated schedules. Tested on high-entropy and soft magnetic alloy runs, the adaptive policy reallocated sample batches based on real-time run metrics and resource caps. It lowered predictive uncertainty along Pareto-optimal compositions without dropping overall optimization velocity compared to static baselines. In semiconductor work, we see similar specialization with PVmatAgent, designed for perovskite photovoltaic materials. It organizes eleven domain-specific tools, including the CHGNet force field and MEGNet neural networks, across nine functional modules. Those modules handle crystal retrieval, bandgap estimation, formation energy, and device-level metrics. It also includes an automated cutoff: execution halts immediately if ungrounded numerical values appear during screening runs. Polymer science is adopting a comparable modular structure. A new framework places a central coordinator over tools for inverse molecular design, synthesis planning, property prediction, and solvent selection. Converting natural-language goals directly into execution sequences for self-driving laboratories, linking molecule generation straight to physical testing protocols. Alongside these implementations, a newly published systematic review mapped out adoption trends and measured utility across agentic artificial intelligence implementations in scientific data workflows. We will be back next week with more reports on research automation. From Agents in Research, thanks for listening.