THE DISCOVERY OBSERVATORY
Papers, discussions, and releases moving discovery research forward.
Feed refreshed Publication window –
162 updates
1 / 3SageAI announces Sage-A1.5, a 35B MoE model with 3B active parameters, and reports a training-data flywheel using scientific discovery workflows and tens of thousands of synthesized environments. Its figures illustrate parallel research delegation and evidence-driven correction. These establish a concrete research-agent connection, but do not demonstrate recursive self-improvement.
1 / 4Yuandong Tian announces Flint-40B-A2B, a local coding model he says was built using an automated research system and trained from scratch. He reports benchmark performance and local throughput, and says the complete training pipeline was built in four months. Video not reviewed.
Xaira announces X-Scientist, an agentic drug-discovery platform. The author reports benchmark gains over other agents and roughly 10× faster target nomination than human curation with a comparable wet-lab hit rate. Video not reviewed.
VX-One combines compact decision heads with an empirical evaluation of when unknown detection and abstention remain useful under distribution shift.
The survey organizes physical self-improvement by where experience leaves a retained change and how that change benefits subsequent behavior or learning. It proposes controlled comparisons for testing whether acquired repair knowledge improves an improver and its subsequent self-revision.
SCDM shares a nonlinear predictor and causal-strength matrix across target-variable prediction tasks. Episodic training samples targets to reduce computation, while separate graph readouts produce directed predictive scores.
1 / 4Robert Lange compares two coding-model releases produced through agent-assisted research. His substantive contribution is distinguishing Mirendil’s researcher-directed experiments from Recursive’s stated ambition to automate idea generation, experimentation and subsequent research decisions through its Eureka Machine.
Richard Socher announced Flint, Recursive’s first local coding model, which he says was built using the Eureka Machine automated research system in under four months. He also announced open-weight availability. Video not reviewed.
LOSCAR extends RESIT by estimating both conditional location and scale before testing residual independence. Pretrained TabPFN quantile predictions supply flexible scale estimates without additional model training.
1 / 4Recursive released Flint, an Apache-2.0 coding model that it says was built with substantial contributions from its Eureka Machine automated research system. The announcement describes experiments in architecture selection, quantization and distillation, and hardware-specific kernel optimization. This provides a concrete model-development artifact of automated research, but does not establish that Flint improves the research system’s own capabilities.
Mirendil reports that one researcher working with AI agents pretrained Mirendil-Base in 3.5 months for under $10 million in compute-related spending, including agent tokens and data processing. The company says agents wrote over 99% of the code, ran experiments and diagnosed failures, while the researcher directed the work and judged results. This is an account of agent-assisted model research; Mirendil explicitly stops short of claiming autonomous recursive self-improvement.
PreCoMem maintains confidence-weighted user beliefs and separates tentative hypotheses from committed facts. New evidence reinforces, provisionally extends, or corrects memory through a reliability-adjusted mismatch signal.
The author interprets Recursive’s reported Flint release as competition between automated research systems: future advantage may come from the machines that design models, rather than individual model releases.
OMAR organizes previously unseen objects by inferring shelf conventions from observed arrangements. It combines multi-view object grounding, persistent semantic scene graphs, language-model placement reasoning, and geometric execution checks.
Memory-Centric Continual Learning organizes agent memory by storage substrate, reusable knowledge asset, and management policy. It connects memory-state updates and policy adaptation to behavioral retention, adaptation, and transfer.
Future Chunk Prediction supplements autoregressive video denoising with prediction of temporally separated chunks. It aims to reduce reliance on local continuity and improve learned representations of scene evolution.
Laurence Moroney proposes evaluating Flint on three real backlog tasks in a familiar repository, measuring time spent correcting its output. He describes the local coding model as produced through Recursive’s automated research system.
1 / 4Recursive reports that its Eureka Machine helped produce Flint by proposing, implementing, testing, and combining experiments across model development. The announcement gives concrete examples in architecture search, quantization with distillation, and hardware-specific kernels. This is an author-reported application of automated research to a released coding model; it does not establish sustained recursive self-improvement.
The author reports that Infinity Autoresearch doubled Nemotron 3 Super token throughput on the HP ZGX Fury (GB300). Video not reviewed.
Heng Wang announces a future-aligned research-ideation framework, motivated by learning from scientific discovery history to train language models for future research. He identifies the accompanying paper as an EMNLP 2026 Findings paper.
Harness Scaling trains policies across structured variations in prompts, tools, observations, context management, and workflows. It reduces specialization to a single execution setup.
GuideEvolve separates task evaluation from evolution control and adds mechanisms for deliberate modification, recovery of previous successful solvers, and autonomous termination. It investigates whether this organization makes agent self-improvement more reliable and efficient.
SERA organizes a high-level vision-language planner around a fixed robot executor. Its harness maintains task context within episodes and is revised between development attempts to improve progress assessment, planning guidance, and recovery.
ExpEvolve converts an archive of evaluated programs into interpretable idea factors and uses those factors to guide subsequent experiments. It combines exploitation of promising ideas with deliberate variation when search plateaus.
DeepEvolve couples a live forecasting-task generator, an evidence-gathering forecaster, and an evolution engine. Forecasting trajectories support parameter training, reusable skill extraction, and generation of harder future tasks.
The paper evaluates complete deep-search and deep-research traces rather than relying solely on final answers. It separates execution and infrastructure failures from failures in exploration, evidence use, reconciliation, and synthesis.
Behnam Neyshabur reports that Mirendil pretrained a coding model from scratch with one AI-assisted researcher in 3.5 months using less than $10 million in compute, including research experiments and final training. He presents this as progress toward Mirendil’s RSI mission; independent evaluations remain forthcoming.
An empirical audit shows how a learned experimental controller can appear successful by repeating a good fixed protocol. A second synthetic task demonstrates observation-dependent choices, but simpler adaptive controllers match the learned behavior.
TraceForge synthesizes reasoning in the student's model-family style while retaining a stronger teacher's answers. Controlled student-training comparisons suggest these synthetic traces can outperform the teacher's original traces on coding tasks.
Slot-JEPA reorganizes dense visual features into stable semantic slots and changing spatial masks. Transition memory and transport-based goal costs make predicted action consequences more useful for visual planning.
Sage-A1.5 develops a 35B mixture-of-experts agent whose learned orchestration and verification behaviors support sustained scientific investigation across tools and parallel research tasks.
S-Agent 2.0 equips a general coding agent with geometric perception, persistent scene memory, and spatial-language readouts so answers remain tied to acquired evidence.
LASAGNA studies whether improvements to individual RAG components survive combination and corpus growth. It finds different behavior for retrieval and faithfulness, and shows that pipeline rankings can change with scale.
A frozen surrogate ranks proposed code edits before execution, directing a coding agent's search toward promising self-modifications.
Sakana AI introduces QD-Rec, an open-ended search method that builds an archive of candidates paired with explanatory hypotheses. It combines bandit-based exploration, LLM mutation and recombination of earlier hypotheses, and local comparisons that preserve distinct ideas. Applications include research-paper recommendations explaining connections to a researcher’s interests. Sakana reports higher average LLM-rated quality and larger archives in many tested settings, rather than independently validated improvements in discovery. Video not reviewed.
OverSeer jointly predicts multiple agents’ views and stationary auxiliary views, using the latter as a shared spatial reference for preserving environmental history and cross-agent consistency.
WaferHDC detects abnormal wafer maps and progressively organizes recurring anomalies into persistent defect modes. It provides a concrete industrial instance of continual open-world discovery.
PAPAV compares digital agents, robotic systems, and multimodal embodied agents through five functions: Perceive, Anticipate, Plan, Act, and Verify. It explains how physical uncertainty changes evidence acquisition, prediction, recovery, and outcome judgment.
MaD-RSI uses behavioral differences among frozen language models as training evidence for a shared harness. It aligns successful and failed trajectories, selects informative cases, and proposes harness revisions that can transfer across executors.
Video-Point Model jointly predicts future camera views and persistent 3D point trajectories. Shared motion supervision improves geometric coherence, visual forecasting, and an action-generating robot extension.
RiseHarness improves terminal-agent reliability by evolving interaction mechanisms around a fixed model. It compares successful and failed executions, connects observed problems to harness code, and retains changes supported by repeated evaluation.
Agent S4 evolves repositories containing executable interfaces and usage knowledge for computer-use agents. A meta-agent learns from ungraded synthetic-task trajectories, while a separate controller selects repositories using validation performance.
CulturalRiddles tests whether language models can combine local knowledge with indirect clues across cultures. Native speakers construct and verify riddles, then help evaluate valid answers that exact matching overlooks.
ChartEvo improves chart reasoning by evolving an external harness around a frozen model. It converts recurring failures into reusable reasoning guidance, executable measurements, and persistent local toolchains, then freezes the resulting harness for evaluation.
LinADMG connects identifiable representation learning with causal discovery under latent confounding. It uses variation in noise distributions across domains to recover directed structure and, under an additional no-cancellation condition, the full mixed graph.
AGENT4SCI provides 25 tasks from 20 scientific benchmarks across 11 domains. It measures how coding agents improve verified baselines under fixed resources and investigates whether those improvements involve domain insight.
A literature audit measures whether recursive self-improvement claims receive appropriate experimental support and proposes reporting requirements for each kind of claim.
AW-0 jointly generates code-based plans, future video, and robot actions. An annotation pipeline extracts recurring routines into reusable skills, and an optional hybrid executor combines generated programs with learned action chunks.
Hugging Face publishes a curriculum covering supervised traces, distillation, reinforcement learning, stateful environments, and training through real agent harnesses. It connects each stage to tutorials and implementation resources.
PhysRSI-eval measures how agents improve physical manipulation through persistent code, tools, and notes while model weights remain fixed. It separates starting ability, improvement timing, attained performance, physical budget savings, and recovery from failure.
Guanghan Ning discusses WITNESS as an interactive testbed for forming, testing, and revising hypotheses, and its connection to DiscoveryRSI. The post focuses on hypothesis-testing environments and the relationship between the two projects.
VLDA predicts compact changes between current and future states, then uses those changes to guide action generation. Local transition selection and global consistency objectives reduce attention to static visual content.
VibeTraining-Bench measures whether research agents can coordinate changes to training data, trainers, and inference harnesses within fixed budgets, selecting improved systems through validation feedback.
TEAL evolves a reusable tool library for a fixed coding agent operating over a data lake. TEAL-Q supplies synthetic analytical questions when real training questions are scarce.
Stateful Harness separates executable world dynamics from video rendering. A persistent semantic ledger and editable spatial scene determine action outcomes, while a frozen video model depicts those outcomes and visual observations inform subsequent state.
Retrospective Self-Distillation trains agents from their own completed interactions by giving a teacher richer historical and future context than the acting student received.
The study measures whether repeated deep-research reports reproduce the same entities, sources, and entity-source associations, treating consistency as a structured evaluation problem.
Ropedia W0 learns egocentric predictions across humans and robots through a shared camera-and-skeleton interface. Human experience supplies transferable supervision for embodied prediction, policy evaluation, and planning.
TokenWorth learns query-dependent memory usefulness from task outcomes while keeping the language-model backbone frozen. The same learned score guides immediate retrieval and admission, replacement, and eviction in a bounded memory bank.
PVBench measures physical plausibility using estimated object states and physical-law residuals. It separates visual appeal from geometric, kinematic, material, and mechanical consistency across paired text-to-video and image-to-video scenarios.
PAR makes recurrent language-model iterations represent successive operations by combining progress-aligned supervision with recurrent-state norm anchoring. Prefix self-distillation supplies intermediate targets when only final answers are available.
OpenVLHarness combines consistent tool interfaces, interpretable evidence rendering, and a per-query multimodal artifact registry. Agents can revisit images, masks, arrays, and web references and compute over earlier results during iterative investigations.
OpenSaC lets research agents express retrieval workflows as inspectable Python programs. Agents can batch independent searches, process and retain sources, and return selected observations before deciding what to investigate next.
Anti-Persistence retains execution records and outcomes but generates temporary guidance for each new task. It separates accumulating experience from permanently storing inferred rules, while adapting how retrieved experience is used.
The Research Progress Graph represents how research ideas evolve through literature, experiments, insights, challenges, and localized revisions, preserving the evidence and reasoning behind each transition.
MemT integrates an evolving memory state into language-model computation. It learns to compress historical information and use that state for later queries, including procedural guidance for coding agents.
MMEnterprise-Bench tests evidence-grounded question answering over simulated multimodal workspaces that preserve corrections and historical versions instead of overwriting them.
MIRA makes multimodal web research responsive to the requested deliverable. It combines explicit report requirements, recorded textual and visual evidence, iterative submission checks, and training trajectories aligned with this workflow.
Knowledge Manifold Optimization searches a continuous literature map for underrepresented regions supported by neighboring research. It preserves competing opportunity criteria through Pareto selection and supplies auditable evidence for subsequent hypothesis generation.
Holi-World converts egocentric recordings into a shared representation of persistent scene structure and changing object states. Automatically generated grounding and question-answer supervision trains a vision-language model, while independent human annotations evaluate record quality and learned reasoning.
HarnessRL jointly adapts task sampling, privileged rollout guidance, and reward rubrics during policy training. A memory-backed proposer and critic seek to maintain useful learning signals as policy capabilities change.
HarnessPro recursively improves agent instructions, tools, and runtime policies while retaining several performance-cost operating points. It replaces direct trajectory judging with fixed executable tests synthesized from public task evidence.
Fréchet Forcing trains streaming world models on their own generated trajectories using real-data statistics. It combines conditional generation supervision with scene fidelity and perceptual realism objectives to reduce accumulating rollout errors without a teacher or long training recordings.
EvaOmni evaluates audiovisual answers together with the temporal evidence supporting them, exposing weaknesses obscured by answer accuracy.
DecisionPivot diagnoses decision models through related cases whose correct outcomes change, supplemented by established public tasks and probability-based reliability metrics.
Self-evolving agents improve aggregate task performance but struggle to retain compatible knowledge across changing topics, workflows, and domains. Shared prompt edits overwrite prior instructions, while similarity-based memory retrieval can select experiences with incompatible workflows.
AnyGoal tests world-model generalization beyond expert training goals while preserving the environment's physical rules and marginal state/action coverage. Its analyses separate failures of dynamics prediction, visual representation, and downstream planning.
AdaExploreBench asks agents to design probes that reveal hidden decision rules and recognize when those rules become stale. It separates proactive exploration from reasoning over supplied evidence and measures adaptation throughout an episode.
ActionRoPE represents camera movement through persistent world coordinates rather than learned action features. Coordinate-Conditioned Noise Scheduling uses the resulting observed-region partition to improve scene consistency.
The perspective distinguishes recoverable information in JEPA representations from understanding the predictive mechanisms of their latent worlds. It proposes a progression from grounding variables to testing causal roles and improving behavior.
ScaleNav learns goal-directed visual dynamics from Internet egocentric videos and transfers them into robot navigation actions using simulated supervision. A shared model predicts future observations and motion chunks while retaining initial and recent visual context.
NeuroAct couples an inner research loop with an outer experience loop. Specialized roles investigate competing explanations, implement experiments, review outcomes, and retain conditional guidance for later studies.
Collective Experience Learning improves a shared robot policy using fleet-generated execution data after expert-supervised initialization. Failures identify relevant successful behavior to revisit, while an asynchronous system sustains collection and learning.
Jointly optimizes a coding agent’s policy and its surrounding harness, treating prompts, tools, workflows, and model behavior as coupled components rather than improving either in isolation.
1 / 4Google Research reports progress toward human-directed autonomous wet labs through VERID, a stateful laboratory digital twin, and agent-guided molecular discovery workflows. The authors report that three experimental rounds improved aptamer hit rates to 33.5% for myoglobin and 42.1% for α-thrombin after addressing experimental artifacts. They also describe faster protein-binder design using epitope excision, improved computational hit rates with human–agent collaboration, and FAP binders with approximately 3–10 nM affinity. Video not reviewed.
This paper introduces DiscoveryRSI, which builds testable hypothesis trees through multi-agent evolutionary search and improves its research workflow using contribution credit, harness replay, and rediscovery. It also introduces DFM-Bench to evaluate hypothesis quality and cumulative research progress.
Introduces Heuresis, a benchmark of anonymized simulated datasets designed to test whether agents diagnose and repair causal-discovery errors rather than exploit semantic variable names or merely select an existing algorithm.
SciΠvot evaluates whether research agents make sound decisions at pivotal moments, using 248 cases derived from human interventions across 26 projects. Frontier models often preferred the original agent plan over the human correction, exposing a major scientific-judgment gap.
This work tests acceptance criteria used by automated discovery systems and shows that seemingly successful descriptive tests can admit scientifically empty explanations. It proposes evaluating null rejection, positive acceptance and independently verified consequences together.
Data World generates synthetic data environments from explicit underlying mechanisms, along with tasks and auditable rubrics. This provides scalable, verifiable training and evaluation material for agents that must discover insights across heterogeneous files.
Google presents a researcher-guided workflow combining protein database searches, functional predictions, structural comparisons and literature checks. The identified antimicrobial peptide candidate is undergoing wet-lab validation.
Val Alexander proposes evaluating whether self-modifying agents retain required behavior, memory provenance, and responsiveness to correction. The article outlines longitudinal controls and explicitly identifies its evaluation protocol as unimplemented and unvalidated.
SwarmRSI evolves the research states shared by scientific agents. It preserves unsuccessful experiments and partially useful programs so later agents can revisit hypotheses, repair implementations, and combine complementary solutions. Self-improvement occurs through inherited research experience under fixed search operators.
The authors find that longer or parallel agent searches plateau as accumulated histories anchor subsequent exploration. Periodically clearing search state while retaining the best candidate restores exploration and improves results on programming, kernel-optimization and model-training tasks.
1 / 4Google Research announces a multi-agent orchestrator for reproducing and adapting existing scientific methods, reporting catheter designs that performed better at preventing bacterial infection in simulation. It also announces wider access to Deep Review, an agentic system for checking technical documents.
Trains an LLM’s action-selection reasoning with information-gain rewards so it can conduct sequential Bayesian experimental design more cheaply and generalize its information-gathering behavior to unseen categories and clinical dialogue.
Formalizes persistent revision of prompts, memory, workflows, tools, code, and plugins while model weights remain fixed, and analyzes admission, retention, validation, and safety requirements for accepted harness changes.
This paper defines a setting where an agent receives a large table and an outcome but no predefined research question. HypAgent interactively proposes, tests and refines multiple hypotheses while removing explained signal so later searches target additional mechanisms.
Evaluates continuous, multi-stage research workflows at both claim and stage level. Its experiments find that execution quality trails planning, repeated attempts rarely improve results, and added skills or research orchestration do not reliably close the gap.
Fan Nie announces OSVerify, which separates proposing errors from validating them to filter mathematical solutions for expert review. The authors report reduced review burden and say pairing it with a math agent helped resolve three Langlands conjectures and contributed to resolving the Manin constant conjecture, providing verification signals for scientific discovery.
Into the Unknown introduces 15 open computational research problems contributed by scientists. It evaluates whether agents can turn a supplied research direction into an honest, complete, scientifically sufficient follow-up paper, without relying on a known reference solution.
Autonomous Science Bench supplies six budget-constrained design-build-test-learn campaigns based on published biology, chemistry and materials experiments. It evaluates agents separately on the quality of measured discoveries and their ability to rank unseen designs.
Sakana AI introduces MASS, a self-improvement loop that searches and evaluates multi-agent workflows, then trains a shared model on selected team executions. The authors report that two cycles with a 27B open-weights model raised score per output token to 1.2–1.6 times baseline across four research benchmarks. They caution that self-evaluation can reinforce shared blind spots. Video not reviewed.
HarnessTTT trains a proposer at test time to produce executable combinations of prompts, skills, tools, middleware and execution controls while leaving the discovery executor frozen. Outcome feedback improves later harness proposals under a fixed search budget.
D-World represents a robot's environment as an executable digital twin that agents update from physical measurements. Proposed parameter changes are accepted only when replay does not worsen measured residuals, while action selection balances task progress against revealing model errors.
Ashish Kumar analyzes code-evolving research systems and argues that current recursive improvement lives mainly in editable scaffolds, search policies, memory, and tools. The post highlights transferred gains from AIDE² but also stresses evaluator gaming across systems.
TagBO extracts coarse task-aware relationships from LLM reasoning and combines them with structural descriptors inside a graph-based Bayesian-optimization representation. Experiments target mixed categorical and continuous chemistry and materials searches under limited evaluation budgets.
This work proposes auditable accounting that distinguishes content generated with designer-supplied mechanisms from genuinely machine-authored operator types. Its physics runtime reports held-out validation of newly composed operators while documenting protocol flaws and limits to autonomy.
Uses predictive world models to guide action and learning-data selection, while observed feedback controls policy or workflow improvement and conditional world-model adaptation.
OpenSocialJax varies layouts and hidden rules across tens of millions of possible social-dilemma environments. Repeated trials test whether language-model agents can discover environmental rules while adapting to teammates and maintaining cooperation.
JANUS organizes autonomous research around a growing graph of proposals, implementations, predictions, and experimental outcomes. It combines experiment selection with feedback-driven improvement of research skills and predictive judgment.
Harness^2 uses task-level recursion to revise an agent's harness without ground-truth verifiers, then extracts experience from those revisions into a continual-improvement harness for future tasks. The reported evaluations cover previously unseen professional tasks.
Introduces an embodied-simulation benchmark centered on compositional generalization to measure whether self-evolving agents accumulate experience and transfer it to new tasks.
SciTaste-Bench evaluates paper assessment across six disciplines using 300 selected papers, seven scoring dimensions, and four external-information conditions.
Presents MASS, an RSI loop where a shared model acts as task solver, workflow optimizer, and evaluator through multiple agent instances, then trains on selected team trajectories to internalize collective improvements.
The paper formalizes abductive autonomous scientific discovery with possibility theory, allowing agents to generate and test hypotheses while acknowledging that better explanations may remain undiscovered. It introduces an abductive-utility objective and an anytime-valid frontier-search algorithm.
A frozen LLM continually builds an explicit world model as a revisable natural-language rulebook, compiles it into executable code, and admits updates only after semantic and replay-based verification.
The paper tests whether frozen language models can use prior experiment records to forecast the effects of unseen research interventions, including transfer to environments with no local outcome records.
LLM-IDEA couples a language-model proposer to deterministic identifiability diagnostics. It also audits whether discovery benchmarks reward distinctions that their observation interfaces can actually reveal.
EvoBelief maintains an external world model of falsifiable rules while keeping the underlying LLM frozen. Learned claims gain permission to affect decisions only after validation and can lose that permission when observations contradict them.
Offers reusable training, inference, rollout, evaluation, and sandbox services for data, harness, and joint improvement. It introduces RSI-Index to compare how much of the remaining performance gap research agents close.
Uses construction, reproduction, and review agents to convert published papers into executable research environments with automated evaluators, producing 210 environments across 18 fields.
Kernaut lets coding agents invent kernel functions as programs while construction contracts guarantee mathematical validity. A quality-diversity archive and functional novelty screening retain distinct candidates whose inductive biases transfer to unseen optimization and enzyme-kinetics tasks.
Procedurally generates synthetic biobanks with concealed but known causal structure, enabling end-to-end training and evaluation of AI scientists for causal drug-target discovery without exposing private participant data.
Anthropic announced a $150 million, three-year commitment to support AI-assisted scientific discovery through the federal Genesis Mission. It says it will provide Claude access, API credits, training, and technical support for several hundred research projects, with collaboration priorities including fusion energy and quantum computing.
ST-Bench tests whether optimized multi-agent workflows improve scientific data analysis across hydrology, agriculture, and wetland methane tasks.
Station places multiple agents in an open-world scientific ecosystem and adds supervision plus periodic meta-reflection to sustain research when intermediate metrics are absent. Agents are evaluated by how many findings from withheld research papers they independently rediscover.
Studies why independent autoresearch runs settle into different local idea regions and proposes branching and selective context removal to encourage broader exploration beyond an agent’s current trajectory.
Guidance-TTT adapts a compact model at test time to suggest high-level strategy changes while a stronger frozen model implements complete executable solutions. This separates strategic learning from implementation and reduces the cost of discovery-time policy updates.
1 / 3The author reports using AI agents and quantum simulations to design YBaMnFeO₅ and recognize previously synthesized KV[Cr(CN)₆] as candidates for compensated magnetic semiconductors with spin-selective band edges. The oxide faces an atomic-ordering obstacle during synthesis. The previously known compound has reported magnetic order above room temperature, but its predicted band gap and spin sorting remain experimentally unmeasured. Video not reviewed.
Introduces MechHypoBench, which tests whether AI scientists can infer mechanisms from complex empirical observations across 14 fields and 17.98 million records. Generated hypotheses are evaluated through predictions under withheld conditions, revealing a substantial mechanism-recovery gap.
Tests whether AI-scientist answers actually follow supplied evidence by withdrawing or reversing evidence in benchmark data. High ordinary accuracy fell sharply under an evidence-grounded metric, showing that prior knowledge and shortcuts can masquerade as scientific reasoning.
Andrew Wharton proposes evaluating AI-generated connections across scientific fields using typed correspondence cards that identify objections and relevant specialists. His proposed experiment pairs hostile expert review with random-pairing controls to test whether the connections offer useful leads. He presents this as an untested discovery strategy and acknowledges that specialist attention remains a bottleneck.
Develops an anytime-valid detector for deciding when iterative self-evolution should stop and a change-point method for selecting an earlier artifact when late updates overfit. Reported experiments substantially reduce compute while preserving unseen-test performance.
VCLMU represents cells through reusable Latent Mechanism Units and models genetic interventions as transitions of their observation-specific states.
Jeremy Bengtson contributes a four-level framework for assessing recursive self-improvement, distinguishing supervised coding and research automation from autonomous successor training. His synthesis contrasts company-reported progress with skeptical research and separates faster AI invention from slower business adoption, arguing that current evidence supports a human-supervised loop rather than runaway self-improvement.
Presents an autonomous forecasting-research system that diagnoses baselines, proposes evidence-grounded architecture changes, executes them, and accumulates results. Deterministic authorities constrain edits, evaluation, and promotion while agents retain open-ended hypothesis generation.
Benchmarks agents that design executable evaluators by selecting tasks, implementing graders, and repairing them through pilots. The best agent-built evaluators approximate target-benchmark rankings but still suffer failures such as truncation and budget exhaustion.
CURIO augments test-time reinforcement learning for program-based discovery with curiosity derived from prediction errors in the proposing policy's representations.
Allows an optimizer to revise both an executor's harness and its own prompts, skills, tools, hooks, and notes while keeping model weights frozen. Execution-based verification, replay, and regression tracking produced held-out and out-of-distribution gains.
1 / 3Ai2 announced AstaBrief 8B, an open-weights model and training-data release for cited scientific reports in its Asta research agent. Its recipe combines supervised fine-tuning, direct preference optimization and citation-focused data filtering. Ai2 reports about 3.5× faster end-to-end report generation than Claude-powered Thinking mode, with mixed quality results across benchmarks.
1 / 3Ai2 released AstaBrief-8B, an open model trained with supervised fine-tuning and preference optimization to turn retrieved scientific literature into cited reports within its Asta research platform. Ai2 reports improved synthesis scores over its base and SFT models and a 72% LLM-judged win rate against Claude-powered Thinking mode on held-out computer-science questions, alongside weaker DeepScholarBench results.
Introduces persistent state for hypotheses, gates, evidence pointers, failures, and recovery decisions across long research trajectories. Its separated controller and structured reflection improve end-to-end research scores and provenance diagnostics.
PRISMS prioritizes candidate molecules, materials and other designs using pairwise expert preferences instead of data-hungry absolute-score regressors. It selectively escalates uncertain comparisons to more expensive experts using a Fisher-information criterion.
Studies runtime selection and coordination among agents drawn from existing AI-scientist systems. Runtime agent selection improved observed mean scores, while extra contracts and verification had mixed effects under constrained budgets.
Combines open-ended simulation experiments with a small number of physical trials to build and repair reusable robot tools. The reported system acquired manipulation skills without demonstrations and with ten minutes of real-robot interaction.
Edison reports a deployment combining clinical-protocol design, regulatory drafting, and source-grounded error checking. Its case study describes a custom benchmark and training process, reporting 81.5% on writing tasks and detection of 98% of planted errors.
Builds a continually growing library of decision-relevant experimental experience. An agent retrieves prior evidence and runs a low-cost pilot only when that evidence is insufficient before committing to expensive full evaluation.
Evaluates whether repeatedly rewritten agent skills transfer beyond the tasks used to evolve them. It focuses on the gap between training-task improvement and genuine held-out generalization across several self-evolving skill methods.
Shows that scalar feedback can hide tradeoffs among competing model behaviors. ConflictGuide exposes behavior-level conflicts so an AutoResearch loop can pursue edits that resolve them rather than merely improving a single aggregate score.
SPEED-AE transforms disentangled representations into coordinates suited to sparse ODE discovery and analyzes the remaining ambiguity in recovered variables.
Creates matched familiar and altered synthetic worlds to test whether scientific agents follow new evidence when it conflicts with recognizable prior laws. Reported laws are checked both by execution and mechanism recovery.
Organizes large-scale biomedical resources around causal mechanisms and links them to a multi-million-edge evidence network. The proposed foundation is designed to evolve as additional population evidence becomes available.
DeepRewind adds commitment control to a deep-research agent. It tracks evidence dependencies, predicts the consequences of accepting conclusions, blocks risky commitments, and repairs affected reasoning when contradictory evidence arrives.
COEVOLVE runs multiple principle-evolution branches in parallel and transfers measurements through a coordination layer. Value-of-information routing and context-sensitive evidence weighting are intended to expand exploration while limiting harmful cross-branch transfer.
The author reports that ten Claude Sonnet 5.5 agents produced Lean proofs of optimality and uniqueness for the seven-electron Thomson problem in about 15 hours. The article describes their coordinated proof search, numerical certificate discovery, exact verification, and checks that the agents preserved the fixed target statements. These are author-reported results; the associated video was not reviewed.
MultiEcho studies frozen world models as experimental systems. It separates the predictability of their responses, recovery of intervention parameters, and correspondence to physical dynamics.
ALDER discovers equation-based world models through interaction. It combines structural proposals, numerical coefficient fitting, targeted experiments, and protected validation, returning a model only when admission criteria are met.
Joeran Beel argues that AI scientists are progressing from tool-using research loops toward systems that can choose and prioritize research directions. The post compares 2026 examples by output, duration, parallelism, cost, and human validation rather than treating paper generation alone as discovery.
GameBoyWorlds benchmarks experience-driven learning through short execution tasks and extended playthroughs. Its unfamiliar fan-made games test knowledge acquisition beyond recall of familiar walkthroughs.
Edison announces licensed access to specialist Nature-portfolio journals inside Kosmos. The integration links research to the publisher's current record, including corrections, retractions, and editorial notes.
No updates match these filters.
