Autoresearch and Automated Experimentation

References

A synthesis of selected conference talks and technical references. Citations link to the source material; they do not imply that every talk on this subject is included.

1. Multi-fidelity Bayesian Optimisation with Continuous Approximations

   Multi-fidelity optimization uses cheaper approximations to guide spending on an expensive target evaluation. The paper's training example varies dataset size and iteration count: smaller runs inform search, while the objective remains performance at the specified full setting. Its method shares information across fidelities and concentrates expensive evaluations in promising regions. Approximation quality can vary across the search space, so cheap evaluations are imperfect evidence about the target objective rather than interchangeable measurements.

2. Equivalence Tests: A Practical Primer for t Tests, Correlations, and Meta-Analyses

   A nonsignificant difference cannot establish no effect because the study may lack power. Equivalence testing instead specifies lower and upper bounds based on the smallest effect of interest and tests whether effects outside those bounds can be rejected. Statistical detectability and practical importance differ: a large sample can detect an effect too small to matter.

3. Randomized policy comparison and experimental uncertainty

   Randomly assigning eligible experimental units to control and treatment supports causal comparison of policy variants. For a scalar outcome, estimate the absolute effect as mean(Y_treatment)-mean(Y_control). Under independent units and appropriate large-sample conditions, an approximate 95% confidence interval is the estimated difference plus or minus 1.96 standard errors. Power depends on outcome variation, sample size and the effect one wants to detect. Choose an overall evaluation criterion before interpreting results, and examine operational metrics as well. The experiment measures outcomes during its observation period; short tests do not automatically identify long-term effects.

4. karpathy/autoresearch: program.md

   The inspected instructions permit editing train.py while prohibiting changes to prepare.py and its evaluator. They specify a baseline run, fixed five-minute training experiments, result inspection, and keeping or reverting candidate commits. A results table records commit, validation score, memory, status, and description, with keep, discard, and crash statuses. The protocol also specifies a run timeout while directing the overall search to continue until human interruption.

5. First Steps Toward Automated AI Research

   The speaker decomposes automated research into ideation, implementation, and validation.

6. Heuristics for Empirical Discovery

   Pat Langley, Herbert Simon, and Gary Bradshaw’s June 25, 1984 report develops BACON’s search for quantitative empirical laws while explicitly distinguishing those regularities from explanatory theories. Their chemistry demonstration supplied both the variables to investigate and their values. They proposed additional systems that could discover qualitative relationships, design experiments, and construct explanations from empirical laws; resulting theories could then guide further tests. The authors contrast their relatively general, data-driven heuristics with expert systems containing extensive domain-specific knowledge.

7. First Steps Toward Automated AI Research

   Under the speaker's definition, recursive self-improvement (RSI) requires modifying the research system itself; optimizing a separate small model is only autoresearch.

8. Strong Inference

   Platt proposes explicitly developing alternative hypotheses, designing experiments whose possible outcomes can exclude alternatives, executing them clearly, and refining the remaining possibilities. A hypothesis functions as a proposed explanation with testable consequences; a discriminating experiment connects its possible observations to competing explanations. The next experiment follows what remains unresolved rather than simply repeating activity.

9. The preregistration revolution

   Nosek and colleagues distinguish generating explanations from existing observations from testing predictions against new observations. Both support scientific progress, but presenting a post-result explanation as an advance prediction overstates its evidential status. Preregistration records research questions and analysis plans before outcomes are observed, making the distinction inspectable. Unexpected findings can generate new hypotheses without retrospectively becoming confirmatory tests.

10. BACON: A Production System That Discovers Empirical Laws

    Patrick W. Langley’s 1977 BACON system at Carnegie Mellon searched for empirical laws in a programmer-defined artificial universe. Rules detected constants and trends, constructed derived attributes such as acceleration from velocity, and tested generalizations. Detected regularities then directed further data collection. Counterexamples caused stored hypotheses to acquire restrictions or exceptions rather than disappear. This demonstrates an early connection between representation construction, hypothesis formation, and subsequent observation selection.

11. The Automation of Science

    King and colleagues describe a robot scientist that generates hypotheses, designs experiments testing their consequences, executes experiments, interprets observations, and repeats. Adam connected biological knowledge, experimental planning, laboratory execution, recorded data and metadata, and statistical interpretation. The researchers followed indirect growth-based findings with more direct enzyme assays. Literature checking also revealed that some apparently new findings were already known but absent from Adam's database.

12. NIST: Completely randomized designs

    A randomized experiment assigns levels of a primary factor to experimental units and compares a response across replicated runs. For an agent failure, an illustrative intervention replaces only suspected faulty tool feedback with verified feedback, while keeping the task, initial state, model, instructions, tool implementation and budget fixed. Compare subsequent verified success against the unchanged condition, with randomized run order and repetitions. Improvement supports the tested feedback explanation under those conditions; it does not prove that planning or other interacting components were faultless.

13. A Mobile Robotic Chemist

    Benjamin Burger and colleagues’ 2020 Liverpool study combined a mobile robot, laboratory instruments, and batched Bayesian optimization. Researchers supplied five hypotheses defining ten mixture variables. The robot prepared and illuminated samples, measured hydrogen by gas chromatography, and used observations to select subsequent batches. Two baseline samples per batch monitored stability. Across 688 experiments over eight days, the search rejected initially promising ingredients and found a formulation producing 21.05 micromoles of hydrogen per hour, versus 3.36 for its baseline.

14. Mathematical discoveries from program search with large language models

    Published online December 14, 2023, Romera-Paredes and colleagues’ FunSearch paired a frozen language model with executable evaluation and evolutionary program selection. A human-supplied program skeleton restricted which logic changed. The system found a construction containing 512 points in eight-dimensional modulo-three space with no three distinct points collinear. Researchers inspected generated programs, identified symmetry, and restricted subsequent searches accordingly, obtaining stronger constructions. This connects an automatically found artifact to human interpretation and productive follow-up.

15. The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

    Chris Lu and colleagues’ AI Scientist connects idea generation, code modification, experiment execution, plotting, manuscript writing, and automated review. Its demonstrations begin with human-written experimental templates in diffusion modeling, language modeling, and learning dynamics. The paper reports implementation failures and warns that successful execution can conceal an incorrectly implemented idea. Limited experimental budgets also prevented adequate control of parameter counts, computation, or runtime in some comparisons, producing misleading conclusions.

16. JCGM 200:2012 — International vocabulary of metrology

    The measurand is the quantity intended to be measured, including the relevant state and conditions. Measurement can itself alter the system, making the observed quantity differ from the intended one. A measurement procedure specifies the operations and calculations in enough detail to perform the measurement. A result generally combines a measured value with measurement uncertainty. Reference measurement procedures provide accepted comparisons for assessing other procedures. These distinctions support specifying the intended quantity, operating conditions, calculation, reference checks, and uncertainty before interpreting an experimental reading.

17. DataComp-LM: In search of the next generation of training sets for language models

    Jeffrey Li and colleagues’ June 2024 DataComp-LM study fixes model and training settings within each scale so dataset interventions can be compared through trained-model performance. At its 1B-1x scale, models have 1.4 billion parameters and consume 28.8 billion training tokens. Table 4 reports Core scores of 27.5 for the RefinedWeb reproduction and 30.2 for fastText filtering trained with OpenHermes 2.5 and ELI5 positive examples. Core averages 22 task accuracies after rescaling each task so random guessing is zero and perfect accuracy is one.

18. NIST AI RMF Playbook: Measure

    Construct validity asks whether an indicator measures the concept it claims to measure; external validity concerns generalization beyond development conditions. NIST calls for documented operating conditions, measurement assumptions, limitations and variance. Evaluations using human-subject data should reflect the population in the context of use. Applied to agent evaluation, define the deployment population and scenario dimensions before sampling, document exclusions, and compare sampled conditions with intended users, tasks and operating environments. A split within an unrepresentative dataset does not establish deployment coverage.

19. NIST randomized blocks: applying controlled comparisons to RAG evidence

    NIST describes holding nuisance factors constant within blocks and randomizing remaining variation. Proposed RAG application: for each fixed query, compare the recorded baseline with a candidate-boundary intervention replacing candidates with independently verified sufficient source passages, then run the unchanged reranker, assembler and generator. Separately replace only the final evidence block with sufficient passages, bypassing retrieval and assembly. Keep corpus/ACL/time snapshot, question, prompt template, model/version, decoding settings, token budgets and evaluator fixed; record passage identities, order and every resulting context. For assembly diagnosis, replay the same candidates through original versus evidence-preserving assembly. Match evidence length and position where feasible; otherwise the treatment changes those too. Compare boundary coverage and answer support, not answer wording alone.

20. The Target Trial Framework for Causal Inference From Observational Data: Why and When Is It Helpful?

    A causal question asks how outcomes would differ under alternative interventions, not merely which patients are likely to have an outcome. The target-trial framework first specifies the hypothetical randomized trial that would answer that question and then attempts to emulate it using observational data. This makes the comparison explicit and can prevent design-induced bias. It cannot repair inadequate data or make unmeasured confounding disappear; effect identification still requires assumptions such as measured baseline confounders.

21. What exactly is ‘N’ in cell culture and animal experiments?

    An experimental unit is independently assigned to a condition; an observational unit is the entity measured afterward. Several observations can belong to one experimental unit. Treating those observations as independent experimental units creates pseudoreplication and exaggerates apparent evidence. Genuine replication also requires independently applied interventions and attention to units influencing one another. The paper distinguishes evidence within a particular subject from evidence across subjects: adding measurements within existing subjects does not supply additional subjects for a population-level claim.

22. Guardrails and interference in online experiments

    Guardrail metrics represent outcomes experimenters seek not to degrade, distinct from the primary success metric. Standard treatment-effect analysis assumes a unit's outcome does not depend on other units' treatment assignments. Social connections, competition for shared marketplace resources or shared infrastructure failures can violate that assumption. Then a treatment-control difference need not estimate the effect of deploying one policy to everyone. Cluster randomization groups interacting units, while switchback designs alternate treatment over time; both require design-specific analysis. Clustering can reduce interference while also reducing effective sample size and statistical power.

23. NIST Engineering Statistics Handbook: One variable at a time

    NIST shows that optimizing one input while fixing the others can miss interactions: the effect of changing one factor can depend on another factor's setting. Its example contrasts an additive response with a response containing a product of the two inputs. Empirical investigation of useful operating settings does not require first deriving a complete mechanistic model.

24. OpenThoughts: Data Recipes for Reasoning Models

    OpenThoughts compares data-source, filtering, answer-generation and teacher choices through downstream student performance. In its answer-filtering experiments, random selection outperformed tested quality filters for math. The unfiltered arm retained twice as many examples and was explicitly not compute-controlled. Teacher choice also mattered: QwQ-32B-generated supervision outperformed DeepSeek-R1-generated supervision in the tested settings despite the teachers’ opposite standalone benchmark ordering. Selecting strategies by average performance did not yield the strongest result in every domain.

25. NIST Engineering Statistics Handbook: Full factorial designs

    A full factorial design includes every combination of the selected factor settings. Two factors with two settings each therefore produce four treatment combinations. This supplies the design structure for crossing a component's enabled or disabled state with the presence or absence of a hypothesized workload condition.

26. NIST: replication and randomized experimental runs

    Replication at repeated factor settings estimates process variability independently of the fitted model and supports checking lack of fit. Randomizing run order prevents systematic ordering from aligning factor levels with drift, wear or environmental changes; it does not eliminate that variation. Robotics application: predefine the response and measurement window, deliberately set timing/contact factors, keep the controller, task and measurement protocol comparable, and repeat each combination with fresh trials. Reset relevant state or record and account for carryover; changing conditions only after hardware warms up confounds treatment with time. Four single observations can fit an intercept, two main terms and an interaction exactly, leaving no residual degrees of freedom to estimate noise. Attribution therefore needs valid assignment, controlled remaining differences and suitable experimental units, not just a crossed table. A controlled simulator comparison identifies behavior of that simulator under its assumptions, not demonstrated physical-robot effects.

27. How Autoresearch Is Changing ML Research — Zhengyao Jiang, Weco AI

    Aiden reportedly obtained a leaderboard improvement by combining Gated Attention, quantization, and a tokenizer change after the first two together barely improved the score.

28. NeurIPS Paper Checklist Guidelines

    The checklist asks researchers to align claims with experimental scope, disclose assumptions and limitations, and provide a reproducibility route. It requests exact execution commands and environments, experimental settings and selection methods, sources of variability behind error bars, and compute and memory requirements. Total research effort should disclose preliminary or failed experiments omitted from the reported results, rather than reporting only the winning runs' cost.

29. Practical Guide to Controlled Experiments on the Web

    Concurrent randomized treatment and control groups support causal comparison when assignment produces comparable populations, users receive consistent variants, and interactions or interference do not undermine the design. Choose the evaluation criterion before inspecting results. Define eligible users and actual exposure; unexposed users can dilute effects, but exposure filtering must preserve comparability. Statistical power is the probability of detecting a specified real effect under the design. For independent observations, standard error decreases approximately as 1/sqrt(n); detecting smaller effects requires substantially more data. A migration test that finds no significant difference can still miss a harmful change when power is low.

30. Adaptive Designs for Clinical Trials of Drugs and Biologics

    The guidance describes group-sequential designs with prospectively planned interim analyses and stopping criteria. Futility means the trial is unlikely to demonstrate the intended effect. Repeated conventional tests require adjustment to control erroneous conclusions; stopping rules and analysis timing affect the design's guarantees. Binding and nonbinding futility rules have different consequences for those guarantees.

31. Generalization in Adaptive Data Analysis and Holdout Reuse

    Selecting later analyses using earlier results makes the selection depend on the reused dataset. Aggregate scores can therefore influence selection even when individual test cases remain hidden; ordinary fixed-analysis guarantees no longer follow automatically. Engineering application: after using evaluation failures or scores to revise a coding agent, retain useful failures for regression testing but evaluate the revised configuration on fresh, independently sampled untouched cases for an ordinary holdout claim. Reuse is possible under specialized safeguards: Thresholdout compares training and holdout averages using noisy thresholds, releases controlled answers and stops when its overfitting budget is exhausted.

32. The Miranda Hypothesis: How Hamilton (the Musical) Poisoned Your Persona Evals

    The speaker recommends preregistering the experimental matrix, questions, rubric, and directional predictions, and clearly labeling predictions separately from observations.

33. PROV-DM: The PROV Data Model

    PROV separates entities with identifiable fixed aspects, activities that use or generate entities, and agents responsible for those activities. It represents derivations, revisions, usage, generation, and responsibility. An experimental application can represent protocol and dataset versions as entities, each execution attempt as an activity, and observations and analyses as generated artifacts. This separates a planned condition from the activities that actually used its inputs. PROV also distinguishes merely using an artifact from establishing that another artifact was derived from it.

34. Author Correction: An autonomous laboratory for the accelerated synthesis of inorganic materials

    The authors clarify that their novelty language meant new to the prediction platform, not necessarily new to science. Manual reanalysis supported 36 of 40 reported successes; four remained inconclusive from X-ray diffraction alone and were excluded from successful synthesis counts. Another compound was removed because it had mistakenly appeared in training data. The revised analysis underwent post-publication peer review.

35. An autonomous laboratory for the accelerated synthesis of inorganic materials

    A-Lab connects computationally selected targets, literature-derived recipes, robotic synthesis, and physical X-ray diffraction measurements used to identify product structures. Its corrected account reports 353 experiments over 17 days, with 36 successful targets, four inconclusive targets, and 17 unobtained targets. Observed reaction intermediates guide subsequent recipe selection and help avoid repeating routes leading to the same unsuccessful products. Search stops when the target becomes the majority product or available recipes are exhausted. Researchers subsequently changed preparation and heating procedures outside the automated search's capabilities and obtained two additional targets.

36. Reproducibility and Replicability in Science: Understanding Reproducibility and Replicability

    The report defines computational reproducibility through consistent results using the same data, computational steps, code, and analysis conditions. Replication addresses the same scientific question with newly collected data; generalizability concerns consistency across other contexts or populations. It distinguishes making artifacts available from actually recomputing results and emphasizes reporting planned versus unplanned analyses, preparation methods, and uncertainty.

37. PyTorch documentation: Reproducibility

    PyTorch does not guarantee identical results across releases, platforms, or CPU and GPU execution, even with identical seeds. Controlling random generators and avoiding nondeterministic algorithms are separate measures. Libraries can maintain separate generators requiring their own seeds. Deterministic-algorithm mode uses deterministic alternatives where available and raises errors for known unsupported operations. Such settings can reduce single-run performance while making debugging and comparisons easier.

38. OWASP Access Control

    Authentication establishes identity; authorization decides which actions that identity may perform on particular resources. A user allowed to initiate a transfer must still be authorized for the source account. Least privilege limits the authority of running code and service accounts, while centralized checks reduce inconsistent enforcement. In an AI application, tool availability and a model-produced argument are therefore insufficient grounds to execute a business operation; the application must apply resource- and action-level policy.

39. Guide, Verify, Solve: The Engineering Discipline Agentic Development Demands

    The talk's 'zero trust' verification means applying a consistent, auditable verification regime regardless of code origin and using methods different from generation.

40. First Steps Toward Automated AI Research

    Recursive describes proposing changes, implementing them, running experiments, and checking promising results for variance and evaluator exploits. In GPU-kernel search, invalid candidates cached outputs, relied on persistent state, or exploited timing details. Increasingly strict correctness checks became part of candidate admission. Humans inspected outputs with AI assistance, while acknowledging possible missed kernel errors. For fixed-budget NanoChat training, the report gives validation scores of 0.9372 and 0.9109 bits per byte for its comparison and selected solution. It explicitly cautions that assembling effective techniques does not establish independent rediscovery because underlying models may know public work.

41. Quantifying infrastructure noise in agentic coding evals

    Anthropic varied resource allocation across Terminal-Bench configurations while holding the model, harness, and tasks constant. Resource enforcement changed infrastructure failures and task success. The report distinguishes additional headroom that reduces transient container failures from larger allocations that enable different solution strategies. In one example, installing a data-science dependency stack exhausted memory before solution code was written, whereas a leaner approach was possible. Thus an unsuccessful trajectory can reflect both an action strategy and the environment's constraints rather than a model-only defect.

42. Effective harnesses for long-running agents

    Anthropic reports coding agents leaving partially implemented work undocumented or declaring completion after seeing only partial progress. Its experimental approach records feature requirements with explicit passing status, progress notes, and version-control history for subsequent sessions. Agents work incrementally and test features before marking them complete. A published feature fixture specifies observable interface behavior rather than merely requesting code changes. The report also describes cases where unit tests or HTTP checks missed failures visible through end-to-end browser interaction.

43. Optimal Bayesian Design for Model Discrimination via Classification

    Experimental design first specifies what information the experiment should provide. For discriminating competing models, controllable conditions are selected to reduce expected uncertainty about which model generated the observations, or to reduce expected model-selection errors. Because outcomes are unknown when choosing the experiment, the criterion averages over possible observations under the specified models and prior assumptions. This gives an explicit objective for choosing an informative experiment, distinct from maximizing the measured performance of a candidate system.

44. On Optimal Game-Tree Search Using Rational Meta-Reasoning

    A bounded agent chooses between acting now and performing a computation that might change its eventual action. The computation's value comes from improved external decisions, with delay reducing utility. A simplified one-step application is VOC(c)=E[U(action selected after c)]−U(current action)−cost(c), expressing cost and benefit in common utility units. Continue when a feasible computation has positive expected net value; otherwise act or stop. Information gathering can be assessed similarly, including its direct costs and effects. Estimation must account for possible results and whether they would change the decision. Partial computations can enable valuable later computations, so a purely one-step stopping rule can miss their combined value.

45. Random Search for Hyper-Parameter Optimization

    Bergstra and Bengio formulate search as choosing trial configurations from an allowed space, evaluating them, and selecting the best observed configuration. Random search independently samples configurations and supplies a simple baseline for adaptive methods. Their analysis explains why a grid wastes trials when only some dimensions strongly affect performance: many grid points repeat the same values along the important dimensions.

46. How Autoresearch Is Changing ML Research — Zhengyao Jiang, Weco AI

    The speaker argues that codebase abstractions bias autoresearch much as neural-network architecture biases learning.

47. Autonomous Agents for Scientific Tasks - Sina Shahandeh, RADiCAIT

    The speaker identifies hypothesis generation, or research taste, as a bottleneck distinct from implementation in open-ended scientific work.

48. Practical Bayesian Optimization of Machine Learning Algorithms

    Bayesian optimization builds a probabilistic model of an objective over candidate settings and uses an acquisition function to choose the next evaluation. The paper describes balancing predicted performance and uncertainty, modeling evaluation duration, and choosing expected improvement per second. For parallel work, it accounts for possible results of pending evaluations; simply repeating the unchanged acquisition decision can select an already-running experiment.

49. National Academies: inference with missing outcomes

    If O indicates an observed outcome and Y indicates success, the observed rate estimates P(Y=1|O=1), not automatically P(Y=1). Selective feedback can make these differ. Missing completely at random means missingness is independent of measured and missing values; missing at random permits dependence on observed information but not remaining missing outcomes after conditioning. MAR therefore does not justify an unadjusted complete-case success rate. Adjustment or weighting needs justified models and adequate observation probabilities. Dependence on unobserved outcomes requires additional assumptions and sensitivity analysis. Derived bookkeeping: with S successes, F failures and U unresolved eligible cases, the eventual success fraction lies between S/N and (S+U)/N, where N=S+F+U.

50. Accounting for Variance in Machine Learning Benchmarks

    The study separates variation from data sampling, initialization, data ordering, augmentation, and hyperparameter search. It varies individual sources while holding others fixed and estimates data-sampling variation through bootstrap resampling. In its studied tasks, data sampling was a substantial source of variability that initialization-only repetitions did not capture. Consequently, repeating training with different initialization seeds addresses a narrower uncertainty than varying the sampled data and the full selection procedure. Contributions from different sources cannot simply be added because they are not independent.

51. NIST experimental-design terminology: replication, interactions and confounding

    NIST defines replication as repeating a treatment combination to estimate random error; an interaction occurs when one factor’s effect depends on another. RAG application: repeat each baseline/intervention condition across the same queries with a prespecified sampling policy, randomize execution order, and report paired success-rate changes with uncertainty. Record seeds when available, but a seed is not proof of deterministic service behavior. A reproducible improvement estimates the effect of that defined replacement under the fixed setup, not the sole cause of the original failure. No improvement cannot exonerate retrieval: assembly or generation can still fail. Compare candidate and assembly repairs individually and together when multiple faults are suspected. Final-context substitution bypasses both upstream stages, so its success alone cannot distinguish them.

52. On Over-fitting in Model Selection and Subsequent Selection Bias in Performance Evaluation

    Cawley and Talbot demonstrate optimization of a finite-sample selection criterion continuing to improve that criterion while independent test performance deteriorates. Their analysis separates a criterion's bias from its variance: even an approximately unbiased estimator can be exploited by selecting favorable noise. The selected development score therefore need not estimate the selected method's repeatable performance.

53. Best practices for the human evaluation of automatically generated text

    The authors recommend defining separate assessment criteria, reporting participant selection and demographics, and measuring agreement among multiple annotators. Practice examples spanning good and bad outputs can help reviewers establish a common rating scale. Random ordering or counterbalancing reduces presentation-order effects; fatigue and practice can otherwise confound system comparisons. Removing participants or missing responses can introduce selection bias. The paper distinguishes exploratory analysis from confirmatory hypothesis testing and explains that testing many hypotheses increases the chance of false-positive conclusions.

54. Time-uniform, nonparametric, nonasymptotic confidence sequences

    A confidence sequence provides intervals whose coverage guarantee holds jointly across observation times. Howard and colleagues explain that continuously monitoring ordinary fixed-sample intervals can inflate error, whereas their time-uniform procedures permit arbitrary stopping under the stated assumptions. This flexibility generally requires wider intervals than methods relying on a fixed analysis time or stronger assumptions.

55. Generalization in Adaptive Data Analysis and Holdout Reuse

    Repeatedly choosing hypotheses or configurations using earlier holdout results can overfit the holdout itself. The paper identifies hyperparameter tuning, feature selection, stopping decisions and analyst feedback as routes through which a supposedly held-out sample influences the selected system.

56. BioWatch PCR Assays: Key Components of a Performance Standard

    The report distinguishes controls that check the measurement process from unknown test samples. A positive control contains reference material expected to produce a signal, helping distinguish a negative sample from measurement failure. A negative control lacks the target and helps detect contamination or nonspecific signals. Known samples also check instrument and operator consistency. Expected acceptable control results should be specified before testing.

57. ALPHALAB: Autonomous Multi-Agent Research Across Optimization Domains with Frontier LLMs — Brendan Rappazzo

    AlphaLab feeds worker postmortems back to a strategist so subsequent experiments can respond to observed model behavior.

58. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

    The paper validates model judges by comparing their decisions with human judgments on the same responses, including human–human agreement as a reference. Its controlled MT-Bench study uses 58 expert-level labelers; a separate Arena analysis uses sampled crowd votes. Under one setup excluding ties, GPT-4–human agreement reaches 85%, versus 81% between humans. The authors test position bias by swapping answer order, verbosity bias through redundant longer answers, and self-enhancement bias through preferences for a judge's own outputs. These results support scoped use as preference estimators, not acceptance evidence independent of model errors.

59. OWASP LLM10:2025 Unbounded Consumption

    Uncontrolled inference can consume shared capacity or money without crossing a content-policy boundary. Long inputs, repeated requests, and expensive operations can make request count a poor proxy for work. OWASP recommends input limits, per-user quotas, resource management, timeouts, throttling, graceful degradation, and bounds on queued and total actions. For an agent, the engineering implication is to account for the whole task, including generated tokens and downstream calls, and enforce limits before repeated work expands beyond its budget.

60. What Parameter Golf taught us

    Parameter Golf supplied a baseline, fixed dataset, evaluation scripts, artifact-size constraint, and training-time constraint. Organizers report independently reproducing every record-track leaderboard submission. They also describe agents copying unusually strong but invalid approaches, propagating the same rule violations. Submission volume exceeded their manual inspection capacity, prompting an automated triage system that flagged submissions for human review. This documents how faster experimentation can increase verification and attribution work alongside useful optimization.

61. First Steps Toward Automated AI Research

    The speaker invokes Karl Popper's evolutionary view of science: proposed explanations must compete and face rigorous empirical testing.

62. The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search

    Yutaro Yamada and colleagues’ April 10, 2025 report replaces fixed code templates with staged tree search. Nodes retain plans, code, measurements, errors, and review feedback. Debugging, refinement, ablation, and repeated-seed runs have distinct roles.

63. Autonomous Agents for Scientific Tasks - Sina Shahandeh, RADiCAIT

    The speaker argues that models can miss small scientifically important image features because they lack appropriate scientific training data.

64. How Autoresearch Is Changing ML Research — Zhengyao Jiang, Weco AI

    In this competition, the speaker traces almost all of Aiden's record-producing ideas to human sources; its strength was finding, implementing, and combining them.

65. LLM Scientific Reasoning: How to Make AI Capable of Nobel Prize Discoveries

    Restricting retrieved papers to a historical cutoff did not prevent the model from using knowledge learned after the discovery.

66. Autonomous Agents for Scientific Tasks - Sina Shahandeh, RADiCAIT

    Explicitly decompose the problem into linked hierarchical documents before asking the agent to propose improvements.

67. ALPHALAB: Autonomous Multi-Agent Research Across Optimization Domains with Frontier LLMs — Brendan Rappazzo

    A Jira-style or Kanban board provides concrete intervention points for autonomous research.

68. ALPHALAB: Autonomous Multi-Agent Research Across Optimization Domains with Frontier LLMs — Brendan Rappazzo

    A flawed evaluation can invalidate the entire optimization process; the speaker recommends starting with a clear way to measure success.

69. How Autoresearch Is Changing ML Research — Zhengyao Jiang, Weco AI

    Aiden couples public-source research and experiments with a quality gate before publishing a pull request.

70. Verifiable Environments for AI in Biology — Kenny Workman, LatchBio

    Data analysis may provide a verifiable substrate for agentic biology, analogous to code in software engineering.

71. Verifiable Environments for AI in Biology — Kenny Workman, LatchBio

    SpatialBench decomposes analysis workflows into smaller tasks with prepared data, scientific prompts, and deterministic graders.

72. Verifiable Environments for AI in Biology — Kenny Workman, LatchBio

    SpatialBench-Long evaluates whole workflows contextualized by experimental design, literature, and the original scientific goal.