THE RESEARCH LIBRARY
Browse 152 works across eight questions in Open World Discovery.
Eight research questions. One connected discovery process.
152 works
Organizes evolving agents by component, update timing, learning mechanism and application domain, covering model, context, tool and architecture changes.
Supplies a systematic map of mechanisms that can improve and retain discovery capabilities.
Compares agent-generated research ideas with human papers and follow-on research using measures of exploration breadth, distance from seed literature and alignment with future work.
An empirical audit of whether scientific ideation expands exploration, useful for evaluating open-ended discovery claims.
Carries LoRA updates across tasks and distills a frozen teacher conditioned on verifier-accepted trajectories, optionally removing invalid-action turns.
A concrete alternative to treating every successful or failed generated token as a direct training target.
Uses a multi-agent hypothesis-generation, critique and refinement system with scientist-provided objectives and feedback.
A reference for hypothesis-search orchestration and the separation between proposal ranking and external experimental evidence.
Generates linear-Gaussian SCMs with an observe/intervene/submit protocol and separate skeleton, directed-graph and intervention-efficiency scores.
A direct environment for testing causal experiment selection separately from observational structure inference and overconfident edge commitment.
Trains agents within their actual runtime harness, explicitly handling prompt reconstruction, token-prefix compatibility, variable sample counts and rollout-level normalization.
A direct checklist for converting long scientific-agent traces into correctly aligned training samples.
Decouples agent execution, environments and RL components, and increases the allowed interaction horizon in stages to balance early exploitation and later exploration.
A training infrastructure baseline for horizon curricula and controlled comparisons across discovery environments.
Runs budgeted design-build-test-learn campaigns, judging both the quality of measured designs and prediction/ranking of held-out designs.
A directly relevant benchmark for the separability of finding promising experiments and learning a reusable model under a limited measurement budget.
Rotates a small active set of harness variants according to mastery and learnability rather than training on a fixed prompt or uniformly large prompt pool.
Suggests evaluating discovery skills across paraphrased and structurally varied harnesses instead of measuring them only under the training interface.
A simulation-pretrained causal model conditions on observational and diverse intervention regimes to predict causal queries using intervention-aware representations.
Helps formalize discovery as narrowing an identified set, which can be useful even when the entire causal graph remains ambiguous.
Augments a persistent multi-agent scientific environment with supervisory guidance and periodic meta-reflection to maintain exploration without a single dense intermediate metric.
A concrete integrated system and evaluation setting for sustained computational scientific investigation.
Samples hidden SCMs, permits bounded interventions and asks agents for both held-out predictions and an explicit graph/mechanism description.
A close benchmark for measuring whether successful answers are backed by recovered mechanisms and appropriate experiments.
Advocates keeping causal claims tied to explicit assumptions, data and formal algorithms while agents assist data inspection, method selection and interpretation.
A clear counterpoint to LLM-generated causal edges and a useful provenance contract for scientific harnesses.
Partitions high-dimensional variables, combines LLM/human prior knowledge with data-driven discovery in local graphs, and merges results through hypothesis/critic agents.
Exposes a testable boundary between domain-knowledge proposals, statistical tests and final graph commitments in a harness.
Treats inference strategies as operators on a base policy and derives strategy-dependent SFT/RL gradient weightings for pass@N, voting and best-of-N selection.
Motivates matching a discovery policy's training objective to the deployed search/selection harness instead of optimizing pass@1 in isolation.
Builds controlled data-integrity, analysis-verification and hypothesis-revision tasks around real datasets, including constructed conflicting-evidence dossiers.
A benchmark for evidence discipline and scientific judgment, complementary to task-completion or equation-recovery metrics.
Evolves agent source code through an archive of candidate agents, selecting self-modifications using downstream coding performance.
Makes harness/code improvement a distinct object from model-weight training and preserves stepping stones beyond the current best agent.
Defines discovery foundation-model systems through revisable research states and seven coupled capabilities, with state-conditioned operations, experimental gating and validated cross-task discovery skills.
Connects problem discovery, representation construction, experiments, verification and learning as one discovery-system agenda.
Evolves evidence-grounded hypothesis trees, compiles investigations into executable tasks, and reuses research histories through contribution credit, harness replay and subsequent rediscovery rounds.
Connects learned hypothesis development, executable research workflows and cumulative improvement of discovery procedures.
Coordinates existing CLI agents through permission boundaries, persistent artifacts, execution budgets and human supervision interfaces.
A concrete harness baseline for testing whether reliable state, evaluators and budget handling explain improvements attributed to agent intelligence.
CANOPY combines broader rollout exploration with reference-policy regularization to preserve useful outcome-reward variation and limit policy drift.
An empirical counterpoint to assuming long tasks necessarily require a process reward: first test coverage and reward-group signal.
Organizes harness candidates and matched outcome records as explicit interventions, producing both deployable designs and provenance-linked training examples for later model updates.
A useful data schema for jointly studying harness variation and the model-training data those variations generate.
Externalizes candidate pools, document memory, evidence graphs and curation state while training the policy to make semantic search and selection decisions.
A concrete division between reliable harness bookkeeping and trainable evidence-seeking decisions.
Trains a harness proposer at test time while keeping the executor fixed; generated prompts, skills and execution controls guide budgeted discovery rollouts whose feedback updates proposer weights.
A direct test-time learning mechanism for improving the procedure used to search for discoveries.
Integrates task solving and meta-level modification into one editable program, allowing archived agents to change both task behavior and the procedure that generates future modifications.
Makes the improvement procedure itself an experimental object for recursive discovery-system development.
Fifteen author-proposed, unpublished follow-up directions supply original papers, code, datasets, and task specifications. Agents submit a follow-up paper, code, and execution logs. Expert-verified direction-specific and integrity rubrics are graded by an agentic judge, separating integrity, completeness, and sufficiency. The main metric requires all three strict gates.
A concrete evaluation protocol for pursuing open scientific problems without a reference solution, with executable artifacts and evidence traceability. The primary contribution is a benchmark and evaluation method, so it belongs in Discovery Evaluation.
Learns during a single target problem using a tail-emphasizing entropic reward objective and state reuse that favors promising candidates while preserving exploration.
Distinguishes producing one exceptional verified artifact from maximizing average answer quality, and makes test-instance-specific policy learning explicit.
Mutates canonical physical-law expressions and embeds them in increasingly complex interactive model systems to test identification beyond familiar formulas.
A control against memorization and a way to evaluate active experimental design under known hidden equations.
Updates an episode-specific adapter on agent, environment or summary text while downweighting spans repeatedly seen in earlier updates.
A baseline for learning from an evolving evidence stream without repeatedly reinforcing a stuck trajectory.
Co-evolves a challenger and solver initialized from a base language model, using solver uncertainty and repetition controls to create a curriculum and self-consistency-derived training targets.
An enabling curriculum-generation and self-training mechanism for systems that expand their learning challenges.
Converts a paper and its code into an executable MCP server using environment setup, tutorial execution, reusable tool extraction, iterative testing and server assembly; the scientific LLM consumes the resulting tools, resources and prompts.
A concrete example of improving the discovery harness and reusable scientific tools without updating base-model weights. It motivates controlled model-by-harness ablations and checks of execution fidelity versus scientific validity.
Combines literature and data-analysis agents to suggest therapeutic hypotheses, interpret laboratory outputs and propose subsequent hypotheses.
Provides a real laboratory-feedback case with an identifiable hypothesis/experiment/update cycle.
SPROUT checkpoints both execution state and matching agent context at turn boundaries, then branches failed attempts through verifier-guided continuation selection.
A concrete harness requirement for reusing partial experimental progress during long-horizon RL.
Constructs data-analysis tasks with hidden evidence DAGs, synthesizes trajectories aligned with evidence primitives, and uses verified potential increments for turn-level RL.
A close precedent for collecting training trajectories and crediting intermediate scientific evidence.
Provides seven Gym-compatible scientific simulations with separate simulation and equation-problem interfaces for data collection and interpretable equation discovery.
A training interface for separating experiment collection, control reward and equation recovery within one discovery task.
Searches a large dataset for multiple testable mechanisms using an agentic residual-driven discovery process that removes explained variation before subsequent searches.
Directly studies choosing what to investigate and accumulating multiple evidence-supported hypotheses.
Alternates harness search proposed by the current policy with RL updates under the selected harness, carrying hard tasks and failure evidence into subsequent rounds.
A concrete co-evolution mechanism for jointly improving discovery policies and the harness that elicits their trajectories.
Retrieves training trajectories, synthesizes a task-specific skill for a fixed-parameter solver, and optimizes benchmark-specific synthesis prompt slots from rollout feedback.
An important harness-adaptation baseline that can improve use of past experience without updating the solver's weights.
Evolves a swarm of hypotheses, programs and experimental evidence through feedback-guided mutation, complementary component recombination and reuse of a shared research-state graph.
Connects scientific agent collaboration with retaining partial progress and reusing evidence across improvement rounds.
Co-optimizes a test-question-conditioned curriculum synthesizer and solver using self-consistency and a difficulty signal favoring partially solvable generated questions.
A candidate for generating locally useful training problems when a target research problem is too difficult to produce positive rollouts directly.
Shapes selected MLP projections for next-context-state supervision and fits a prompt-specific fast-weight perturbation by ridge regression before generation.
An architectural memory/adaptation comparison for keeping long research evidence usable, separate from reward-driven agent learning.
Orchestrates idea generation, novelty checking, code edits, executed experiments, plotting, manuscript writing and automated reviewing around supplied research templates.
An end-to-end harness baseline and a source of candidate trajectories for studying research-stage failures.
Research resource covering autoresearch, experiment loop.
Uses an evaluated archive to select a coding agent that edits its own implementation, tools and oversight mechanisms, retaining changes according to downstream benchmark utility.
An enabling example of improving the executable discovery harness through evaluated self-modification.
Coordinates literature/implementation analysis, idea generation, algorithm implementation, validation and writing through iterative specialized-agent feedback.
Useful for separating implementation success, idea quality and empirical improvement when evaluating a research harness.
One language model proposes and solves executable reasoning tasks, using validity and learnability signals to generate its own curriculum.
A concrete mechanism for collecting new verifiable tasks without a prewritten question-answer corpus.
Specialized agents perform literature review, experimentation and report writing, with optional human feedback and a pool of scored experimental programs.
Shows how successful and failed execution feedback can be captured separately from the final report for training data collection.
Records agent execution through a common transition interface and separates training from runtime orchestration, decomposing multi-call trajectories for credit assignment.
An infrastructure baseline for training an existing scientific-agent harness while preserving tool feedback and call provenance.
Combines curiosity-guided task synthesis, experience-conditioned exploration and step-level attribution rewards to train tool-using agents efficiently.
Provides training and credit-assignment components for agents learning from synthesized tasks and accumulated interaction experience.
Allows independent agent laboratories to publish/retrieve reports from a shared archive and build on previous computational research.
A model of population-level research memory and reusable discovery artifacts beyond one agent's context window.
Reviews single- and multi-agent scientific workflows, implementations, datasets and task-dependent evaluation metrics.
Provides entry points into domain-specific systems and reliability questions across the research lifecycle.
Evolves executable candidate programs with an LLM ensemble, a program database and task-specific automated evaluators, using successful candidates and feedback in later prompts.
A strong frozen-model search baseline when evaluating the benefit of weight updates in discovery.
A meta-agent writes and refines executable agent designs, storing prior candidates and their evaluations to guide later proposals.
Treats workflows, prompts and tool orchestration as searchable representations instead of assuming the harness is fixed.
Uses foundation-model embeddings to search simulations for semantic targets, historical novelty and diverse collections of behaviors.
An explicit example of representation-dependent novelty: changing the observer embedding changes what counts as a new behavior.
Synthesizes scientific agents as orchestration across scientists, natural language, code and physical processes, spanning the research lifecycle and levels of autonomy.
A complementary organizing perspective connecting agent interfaces with scientific evidence and human participation.
Uses an LLM agent to select genetic perturbations in batches, incorporating prior observations and optional literature, critic and gene-search tools.
A concrete scientific evidence-acquisition workflow and a useful test of whether a harness component actually benefits a given policy.
Lets a scientist agent query a hidden probabilistic world, revise a model and communicate an explanation to a novice whose predictions evaluate that explanation.
A useful decomposition of evidence acquisition, model quality and reusable communication, rather than scoring only a final answer.
Manually specified SCMs generate ARC-like observational, interventional and counterfactual demonstrations, enabling several reasoning and adaptation protocols.
A controlled comparison of reasoning with different causal evidence types and a task-family generalization test.
Architect/technician agents interact through modules that validate experimental procedures, partition plans, enforce dependencies and retain structured experiment knowledge.
A concrete harness contribution: experimental control and provenance can be evaluated independently of policy parameter updates.
Separates a correctness-reward RL-only reasoning experiment from a practical multi-stage pipeline combining cold-start supervision, RL, rejection sampling and distillation.
A baseline for verifiable-outcome optimization and for distinguishing successful-answer reinforcement from learning how to acquire unknown evidence.
Constructs discovery questions and reference hypotheses from published data-analysis workflows, with a separate synthetic dataset-generation route.
A fixed-data benchmark for hypothesis formulation, dataset interpretation and multistep analysis.
Research resource covering harness, long running agents.
Combines workflow stages, autonomy levels and architectural mechanisms into a domain-oriented organization of agentic science.
Helps choose coverage categories and identify gaps between a workflow claim and the components actually evaluated.
Classifies LLM roles as tools, analysts and scientists across stages of the scientific method, pairing methods with benchmark and domain examples.
A navigation framework for the awesome list and for separating individual research-stage competence from end-to-end autonomy.
Data-analysis and literature agents share a structured world model while pursuing multi-cycle research objectives and producing source-linked reports.
Motivates long-horizon memory, claim-level provenance and discovery-quality-versus-compute measurements.
A frozen LLM proposes executable equation skeletons; numerical optimization fits constants, and scored programs are retained in a clustered experience buffer for iterative refinement.
A useful frozen-policy baseline against which to measure whether RL actually improves equation search efficiency or generalization.
Combines transformed known equations and synthetic scientific terms to reduce simple formula recitation, evaluating both data fidelity and symbolic agreement.
Separates memorized scientific priors from fitting less familiar mechanisms and motivates hidden mechanism families in discovery evaluation.
Treats a small trainable model as recurrent hidden state, updated by a learned self-supervised loss, while outer training learns the surrounding projections and update behavior.
Provides a trainable memory mechanism that could consolidate observations during long discovery histories.
Packages 75 Kaggle competitions with recreated grading, agent execution environments and leaderboard-derived human reference scores.
A practical long-horizon experimentation benchmark with executable outcomes and compute-sensitive comparisons.
Provides modular Gym-style research tasks, shell-execution environments, datasets and agent interfaces across a 13-task ML benchmark.
A starting interface for collecting research-agent rollouts and testing alternative policies/harnesses under controlled task definitions.
Uses seven research-oriented competitions with baseline-relative effectiveness, implementation evaluation and explicit runtime/memory constraints.
A strong counterweight to evaluating discovery solely through novelty prose or paper-review scores.
Uses language/vision models to generate task descriptions and executable environments, train agents, and retain learned/failed tasks in an evolving archive.
A concrete model for generating discovery-training worlds and curricula from a persistent record of competence and failures.
Agents implement research papers from scratch and are graded against hierarchical rubrics, including execution and result-match requirements after clean-environment reproduction.
Useful for instrumenting long-horizon research dependencies and measuring reproducible work rather than a plausible report alone.
Trains multi-turn trajectories with state, reasoning, action and reward records; analyzes collapsed within-group reward variance and stabilizes updates with filtering and critic-related changes.
Useful for measuring whether exploration diversity and reward variance survive long discovery rollouts.
Evaluates agents and human experts on seven bounded AI R&D tasks with quantitative scoring and explicit time/compute conditions.
Shows why long single trajectories, repeated short attempts and total cost must be compared under a common budget.
Builds research-grounded programming tasks with heterogeneous datasets, expert-provided task knowledge and execution-based validation.
A benchmark for reliable implementation of individual scientific workflow steps before making end-to-end autonomy claims.
Optimizes search-interleaved LLM rollouts with PPO/GRPO and masks externally returned retrieval tokens from policy optimization.
A practical precedent for training tool selection from terminal rewards while respecting which trajectory tokens were actually generated by the policy.
An outer RL loop selects generated training data and update directives; an inner supervised update applies those self-edits to model weights, rewarded by downstream adaptation quality.
A direct template for training how evidence is converted into learning updates, not only how evidence is gathered.
Estimates labels by majority agreement among sampled answers, then uses agreement-based rewards for RL on otherwise unlabeled test questions.
A useful self-consistency baseline when external labels are unavailable.
Combines idea generation, progressive experiment-tree search, analysis and manuscript generation, with experiment management and visual feedback.
A system baseline for the complete research workflow and for collecting traces of experimental branching and reporting.
Uses computer-controlled physical chambers with actuators and sensors to generate heterogeneous datasets and interventions over well-characterized systems.
A bridge between inexpensive synthetic discovery and controlled physical experimentation with known mechanisms.
Builds temporary adaptation data from few-shot demonstrations, with leave-one-out tasks and domain-specific transformations, then combines updated weights with prediction aggregation.
Suggests testing whether task-local examples support a useful inner-loop objective before introducing a complex RL adaptation policy.
A principal-investigator agent coordinates specialized scientist/critic agents through meetings, producing a computational nanobody design pipeline with human guidance.
A real experimental-validation case for distinguishing proposal generation, computational scoring and physical evidence.
Updates a neural associative memory with gradient-based surprise, momentum and forgetting, combined with attention and persistent learned memory.
A memory-design option for long evidence histories and selective retention.
Organizes scientific agents around planners, memory, action spaces and verifiers, distinguishing prompt-native planning from learned planning.
A mechanism-level index for separating harness design, policy learning and evidence verification in a discovery agent.
Uses unsuccessful web attempts to generate new curriculum instructions, learns an outcome reward model and controls update drift while reusing trajectory data.
A data-collection pattern for training from failures while tracking environment feedback and task difficulty.
Reconstructs three Kosmos radiobiology analyses and compares proposed gene signatures with random-gene or permutation null models.
A concrete example of claim-level discovery validation and of retaining negative findings rather than rewarding persuasive hypotheses.
Collects teacher-model interaction trajectories, filters by task reward and mixes accepted agent demonstrations with general-domain instructions for supervised tuning.
A concrete data-collection and SFT baseline before RL, with explicit environment transcripts and retention controls.
A ReAct-style LLM invokes specialized chemistry/search/computation tools, with cloud-connected synthesis and a human-query tool.
Illustrates the difference between a capable scientific tool harness and a policy trained to choose informative experiments.
Alternates policy sampling with reward annotation and reward-weighted likelihood training, instantiated with correctness-filtered math/code solutions.
A simple self-training baseline that isolates whether expensive multi-turn RL is needed beyond verified trajectory reuse.
Behaviorally probes language models on multiple causal tasks and discusses connecting text-based prior knowledge with statistical and logical causal methods.
Useful for defining which parts of causal reasoning can be elicited from pretrained knowledge and which require new data or intervention evidence.
Provides interactive simulated investigations with instrument use, interventions, hypothesis formation and explanatory outputs.
A training/evaluation substrate for complete long-horizon investigations with separate process and knowledge measurements.
A frozen LLM proposes reward programs; inner-loop RL trains control policies and resulting task fitness/reflection guides reward-program evolution.
A template for learning or searching the training interface while keeping the final discovery evaluator externally specified.
A frozen LLM proposes programs while an evaluator and island-based program database retain useful candidates and construct subsequent prompts.
A necessary search-only baseline: externally verified discoveries can occur without any test-time parameter training.
Combines estimated learning progress with a foundation-model judgment of human interestingness to sample training tasks.
A way to distinguish useful, learnable research questions from unlimited novelty or raw prediction error.
Defines open-endedness relative to an observer through continuing novelty and learnability, then develops a position on foundation-model-based open-ended systems.
Provides a theoretical lens for asking whether discovery continues to yield learnable novelty over time.
Generates candidate trajectories, retains successful ones and uses a reflector with environmental feedback to repair failed outputs before adding them to training.
A specific answer to how failed attempts can yield training data: preserve feedback and validate a correction instead of imitating the original error.
Retrieves related training sequences and fine-tunes a language model on their next-token objective at inference time, instead of placing all neighbors in the prompt.
A baseline for choosing what evidence to consolidate into weights rather than only storing it in context.
Trains a recurrent policy across synthetic SCMs to choose budgeted interventions and edit an explicit graph estimate, rewarded by terminal graph accuracy and an intervention bonus.
Directly demonstrates that a discovery training problem can learn experiment selection and belief/graph updating jointly, beyond learning a representation alone.
Uses a proposed causal DAG to guide adaptive interventions, with guarantees that depend on advice quality while retaining robustness to poor advice.
A useful design for combining LLM-generated causal hypotheses with experiments that verify or correct those hypotheses.
Combines literature-trained recipe proposals, robotic synthesis, XRD interpretation and thermodynamics-guided active learning to revise unsuccessful recipes.
A prospective physical discovery loop and a direct example of failed experiments serving as evidence for choosing the next experiment.
Connects an LLM planner to web/documentation search, code execution and laboratory automation, including chemistry and hardware-specific interfaces.
Shows why physical execution, documentation grounding and validation interfaces belong in a scientific-agent harness.
Alternates solving tasks, compressing found programs into reusable library abstractions, and training a recognition model on replayed and generated tasks.
Separates improving a representation language from improving search within that language, making representation growth an explicit learning target.
Learns policy/value-guided assembly-program search with separate correctness/latency value predictions and explicit instruction constraints.
A concrete long-horizon construction problem where dense correctness feedback, search and terminal performance have distinct roles.
Meta-trains a large memory-based RL policy across a procedurally generated task space with an adaptive curriculum, enabling within-context adaptation to unseen embodied tasks.
An enabling foundation for learned experiment-like exploration and rapid adaptation in unfamiliar worlds.
Asynchronous expression populations combine mutation, crossover, age regularization, migration, simplification and continuous constant optimization under accuracy/complexity tradeoffs.
A strong non-LLM search baseline and a concrete example of how a representation grammar and an optimizer jointly determine what can be discovered.
A transformer maps observational/interventional datasets to an autoregressive distribution over graph adjacency entries, trained on synthetic graph/data pairs.
A learned inference baseline that separates recovering structure from selecting which intervention data to acquire.
Interleaves language reasoning, environment actions and returned observations in one trajectory, using prompting and optional fine-tuning.
A basic action/observation protocol for instrumenting discovery trajectories and comparing reasoning-only with interactive evidence acquisition.
Separately prompted search/relevance/grounding steps generate trajectories that are scored and reused for self-training and smaller-model distillation.
Evidence against a blanket rule that only superficially correct intermediate steps should be retained; selection must match end-to-end objectives.
Stores verbal self-reflections from evaluated trials in memory and conditions later attempts on them without gradient-based policy updates.
An essential memory-only baseline for distinguishing harness adaptation from test-time weight training.
Graph-network energy predictors filter candidate crystals; DFT relaxation/evaluation adds verified computational labels for subsequent active-learning rounds.
Shows that discovery training can be supervised active learning with a computational oracle, providing a useful alternative to LLM-agent RL.
An LLM proposes exploration tasks, writes executable skills and refines them from environment/error feedback; validated skills enter a growing retrieval library.
A model of persistent skill accumulation and endogenous task generation that can be transferred to scientific tool use.
Combines recursive problem simplification, neural-guided search, large-scale pretraining, genetic programming and linear-model tokens in one modular symbolic-regression system.
A strong precedent for training a prior across problems and adapting search to a new discovery instance; useful for decomposing gains from model training and search harnesses.
Amortizes graph inference with a permutation-aware neural model trained on sampled graph/mechanism/data distributions under a forward-KL objective.
A useful model of learning an inference procedure across worlds; separates amortized causal inference from learning a policy to gather data.
Constructs DAGs sequentially with a generative flow network trained to sample graphs proportionally to Bayesian scores rather than returning only a maximizing graph.
An alternative to reward maximization when discovery requires diverse plausible hypotheses and uncertainty-aware experimental design.
Learns visual causal representations from temporal transitions with observed intervention targets and an intervention-conditioned latent transition model.
Shows why observation representation and available intervention diversity must be specified before measuring discovery-policy ability.
Uses weight/augmentation-averaged teacher predictions and stochastic restoration toward source parameters while adapting across changing domains.
A precedent for retaining a stable reference policy and testing revisited worlds after online agent updates.
An AlphaZero-style policy/value network guides tensor-decomposition search; training combines played games, synthetic factorization demonstrations and problem symmetries.
A direct discovery-RL precedent showing how to obtain valid synthetic training instances when generating solutions is easier than solving the forward problem.
Selects reliable, non-redundant samples for entropy adaptation and regularizes important parameters to reduce forgetting.
Suggests separate controls for update usefulness, compute cost and retained capability in a discovery-agent learning stream.
Evaluates batch acquisition policies and counterfactual outcome estimators over curated genetic-intervention screens.
A reproducible finite-pool baseline for sample-efficient experiment selection using biological measurements.
Predicts both transition means and variances, subtracting a learned aleatoric term from prediction-error curiosity.
Useful when designing discovery rewards that should favor learnable uncertainty rather than irreducible noise.
Combines posterior uncertainty over graphs and mechanisms with information-based acquisition, selecting both intervention targets and values. Greedy and soft batch selection trade acquisition quality against computation.
A baseline for the experiment-action space: choosing what to manipulate and by how much are separate decisions.
Iteratively generates rationales and fine-tunes on correct answers; rationalization conditions on known answers to recover additional training examples.
A reasoning-training comparison for success-filtered self-training, especially when contrasting known-answer tasks with discovery lacking answer labels.
Simulates elementary science tasks through interacting objects, simplified physical processes and executable textual actions.
A tool-use and experimental-execution baseline before testing novel mechanism discovery.
Organizes autonomous learning around novelty detection, characterization, self-initiated data gathering, adaptation and continual incorporation of knowledge through human and environment interaction.
A direct conceptual foundation for agents that initiate investigations and learn from unknowns during deployment.
Adapts an image encoder to each test input with masked reconstruction while retaining a recognition head.
Illustrates both per-instance adaptation and the importance of selecting an auxiliary objective compatible with the downstream scientific task.
Uses paired observations before/after unknown interventions to learn latent causal variables and mechanisms, with explicit and implicit causal-model parameterizations.
A precise precedent for representation-level discovery with weak experimental supervision, rather than treating every new embedding as a discovered variable.
Optimizes a design policy over entire experimental histories using sequential contrastive lower bounds on information gain. History pooling supports fast adaptive design after upfront training.
A direct precedent for learning experiment selection rather than merely solving a fixed question, with train-time amortization and test-time evidence gathering separated.
An autoregressive recurrent policy emits expression trees and is optimized using the conditional reward of samples above a reward quantile.
Direct precedent for optimizing the best discovered result rather than average rollout reward; important when comparing discovery RL with ordinary task-solving RL.
Introduces a latent differentiable distribution over DAGs and performs approximate joint graph/parameter posterior inference with particle-based variational methods.
Supplies an uncertainty-bearing belief model for selecting interventions rather than collapsing immediately to one guessed causal graph.
Uses a recurrent expression generator to seed short genetic-programming runs and trains the generator using its own samples plus selected high-performing evolved expressions.
Shows how trajectories or candidates improved by a non-neural search process can become training data for a discovery proposal policy.
Combines self-supervised adaptation with stored source feature moments and online alignment, using contrastive supervision and feature queues.
A direct reason to validate that a test-time training signal aligns with discovery quality instead of assuming any self-supervised update helps.
Updates normalization statistics and affine parameters online by minimizing prediction entropy on unlabeled target inputs.
A compact label-free adaptation baseline and a warning that optimizing confidence differs from acquiring reliable new evidence.
Connects structural causal models, independent mechanisms and sparse distribution changes to the problem of learning causal variables from low-level observations.
Makes variable construction and mechanism learning distinct research targets; the representation is part of the discovery problem.
Formalizes open-world learning with and without labels, connecting novelty detection, clustering, instance management and incremental recognition in a modular learner.
Provides foundational vocabulary and evaluation distinctions for learning unknown categories and managing accumulated knowledge.
Maintains a population of paired environments and agents, mutating challenges and transferring solutions, with a behavioral characterization separated from the environment encoding.
A curriculum-generation baseline in which stepping stones and transfer can produce tasks that a linear difficulty schedule misses.
Pools datasets while explicitly representing their contexts as variables, then supplies context-to-system background constraints to suitable causal-discovery algorithms.
Provides a representation for accumulated experiments with heterogeneous interventions rather than treating all samples as exchangeable observational data.
Jointly trains supervised and self-supervised tasks, then adapts shared parameters on an unlabeled test input using the auxiliary loss before prediction.
Suggests learning an update rule whose auxiliary signal can be obtained from observations when the desired test label is unavailable.
Learns a trajectory encoder and generative decoder for a posterior over task embeddings, then conditions a policy on that uncertainty to optimize online adaptation return.
A baseline for evaluating the value of information through later task performance while keeping posterior inference and action selection explicit.
Meta-trains a recurrent policy to gather information before a rewarded causal query, with within-episode adaptation carried by recurrent state rather than new weight updates.
A concrete training formulation for choosing informative interventions and using their results to answer later causal questions.
Updates a Transformer-XL on already observed sequence segments before predicting later segments, treating weights as a short-term adaptation memory.
Establishes the causal ordering needed for test-time learning from newly acquired evidence without training on future targets.
Trains a class-independent meta-classifier over nearest examples, then recognizes enrolled classes and rejects unseen classes without retraining that meta-classifier for every class addition.
An early open-world knowledge-acquisition mechanism that makes recognition and incorporation of new categories explicit.
Represents acyclicity through a smooth matrix-exponential equality and optimizes a penalized structural-equation score with an augmented Lagrangian.
An interpretable structure-learning component or baseline inside an agent's model-building loop.
Replays failed goal-conditioned trajectories against alternative goals that their achieved states satisfy, recomputing the corresponding reward.
Inspires reuse of failed experiments as evidence for a different, explicitly stated target.
Tests whether a target's conditional predictive relationship is invariant across environments and intersects accepted predictor sets to obtain conservative causal statements.
A way to ask agents for mechanisms that survive environment changes rather than explanations that only fit one dataset.
Adds an approximate KL change in a Bayesian dynamics model to the external reward, using variational neural-network inference and a practical second-order approximation.
An implementation baseline for training evidence-seeking behavior from belief updates rather than raw environmental surprise.
Formalizes recognition with unknown rejection and subsequent category incorporation, deriving bounded open-space-risk conditions and a Nearest Non-Outlier classifier that supports incremental updates.
An early operational foundation for detecting unknowns and extending an agent's recognized knowledge over time.
Uses independence between causes and additive noise to identify causal direction under restrictions; RESIT searches for sinks by regression-residual independence and then removes unnecessary edges.
Demonstrates why observational discovery can work only after a mechanism-family assumption adds information beyond conditional independence.
Defines causal queries through autonomous structural mechanisms and intervention operations, separating model specification from statistical identification and estimation.
Requires an agent to state which intervention or counterfactual query its evidence supports and which causal assumptions are needed.
Intrinsic reward compares old and improved compressors on the same accumulated data; the controller seeks expected future compression progress. The appendix separates compression length, runtime cost and delayed compressor-improvement feedback.
A candidate discovery objective is improvement in an agent-relative explanatory code, not novelty or prediction error alone.
Contrasts two-part description length with refined universal coding and normalized maximum likelihood, exposing model-complexity costs that depend on a specified model family.
Provides an explicit fit-versus-complexity objective for candidate explanations, and a way to discuss changes in an effective hypothesis language.
Characterizes potential-based shaping with the discount-consistent difference of successor and current potentials, preserving optimal policies under the stated MDP conditions.
A starting constraint when adding dense progress feedback to a sparse discovery objective.
No works match these filters.
