Experimental memory across changing tasks

This release contains shared physical/chemical response models, experimental-memory models, the matched action-credit continuations, and their complete experiment records.

Experimental memory reconstructs predictions from actual past input–outcome ledgers and tests their applicability against current observations. In the primary comparison, returning chemical tasks use 59.33 / 64.23 operations versus 88.67 / 83.07 for separately trained no-memory models. Both memory and no-memory models complete all 240 chemical tasks. New-task costs increase, and fixed evidence reading remains competitive.

With the same trained weights and evidence rule, retaining twelve rather than six task ledgers reduces overall operations to 69.77 / 72.43. A second 120-task cohort reproduces the retention effect: 78.02 / 85.38 to 69.97 / 75.02, with all tasks completed. The main gain is recovery after longer interruptions; short returns incur a small extra cost. Direct action-return training of the reader does not establish an additional gain over prediction-only training.

The model packages and experiment records include the fixed endpoints and negative controls, complete physical outcomes, paired stream intervals and cached pre-return predictions. These are known simulator families using structured observations and numerical tools. Broad multimodal transfer, autonomous archive writing and neural consolidation are not established by this release.

Assets include:

• Shared-response parents and their records.
• Both experimental-memory continuations and separately trained no-memory controls.
• Both action-credit and prediction-only continuation pairs.
• Native execution records, Bayesian and retrieval references, fixed cohorts, and developmental planning comparisons.
• Training inputs and initializer weights; no earlier model release is needed for inference from a shared checkpoint.

Check out v0.14.0 and follow the download and reproduction guide. The method and complete comparison includes all conditions and paired intervals. Archives restore repository-relative paths. For split records, concatenate the .part* files in filename order and decompress with zstd -dc --long=30 | tar -x. The follow-up records retain one shared initial environment per instance; per-condition copies are omitted. Preserve the included upstream DiscoveryWorld Apache 2.0 license with its derived records. CDM code and checkpoints retain the repository license.

Learning useful response updates from selected experiments

This release studies whether feedback from one experiment can improve predictions and decisions for later experiments. It separates response learning from cost learning and compares stored representations, refreshed representations, relation-conditioned correction and probability calibration.

Sixteen update trainings and 1,040 native task executions are complete. With identical model weights, retaining the learned relation correction reduces mean operations from 117.38 to 109.95 and from 119.93 to 111.75, with all 40 tasks completed in each condition. Paired intervals for both reductions exclude zero. Prediction improvements include mixtures not previously tested in the current task.

The relation method does not yet show a stable advantage over independently trained calibration. Four full shared-core physical/chemical continuations are running, with native comparisons queued; their unfinished endpoints are not included in this release.

The archive contains two complete shared-core parent models, all sixteen fitted response rules and loading instructions. The repository includes methods, training and evaluation commands, paired predictions and every native task outcome.

• Study and reproduction
• Complete numerical results
• Transfer to previously untested mixtures

Chemical experiment learning and native comparisons

This release provides the learned chemical response models and their native experimental experience.

• Four self-contained checkpoints: frozen-core warmup and joint endpoints for training seeds 753 and 754.
• 320 native training instances with 10,610 returned experiments and their public action records.
• Training code, split definitions, fixed endpoint records and all 360 native confirmation executions in the source repository.

All six confirmation conditions complete 60/60 tasks. The two joint models use 143.35 and 278.37 mean native operations, against 367.38 for primitive-recipe enumeration. Removing direct experiment history from the first model raises cost to 343.75. The complete results preserve both training repetitions, recipe-group uncertainty and the weaker evidence for the additional recurrent-state readout.

The models use structured public observations and represent a chemical-task adaptation. Autonomous computation allocation, cross-task retention and online neural learning are separate ongoing research questions.

See the download and reproduction guide and complete chemical study. The source also preserves the completed twelve-condition physical comparison and the revised execution-aware training implementation.

Physical prediction and experimental action learning

Physical predictions that affect experimental decisions

The model now carries uncertain physical relationships between tasks and uses them for measurement selection and device settings. Nine relation tokens enter the same continuing Transformer workspace. Six joint trainings update the 26.65M-parameter action core together with the physical belief; Bayesian posterior updates operate during execution.

On the same 120 new native Reactor Lab instances:

Method	Parent A: completed / mean actions	Parent B: completed / mean actions
Frozen calibration-rehearsal parent	115/120 · 117.18	115/120 · 125.91
Ordinary continuation	114/120 · 122.40	115/120 · 137.80
Joint physical learning with memory	117/120 · 105.63	116/120 · 106.93
Same trained weights, memory discarded	118/120 · 113.58	117/120 · 107.65

The complete method reduces actual actions by 9.9% and 15.1% against the corresponding independent parents, including failed tasks. Whole-stream 95% intervals for action differences exclude zero in both repetitions. Completion differences are small and their intervals include zero.

Memory alone has a mixed result: retaining it loses one completion in each repetition, and the first returning task does not improve. Saved observation traces identify noisy calibration and extrapolation as concrete remaining difficulties even when the correct relationship is recognized. The release also preserves the physical-prediction transfer tradeoff and the negative comparison of interval feedback against a history-matched failure mask.

Contents

• Four physical pretrainings and all six final joint policies, with inference and optimizer-bearing model packages.
• Thirteen native conditions on 120 shared instances (1,560 executions), matched parent and ordinary-training controls, and same-weight memory interventions.
• Complete training curves, development records, simulator randomness, native trajectories and public-feedback replay.
• Editable result figure, per-instance tables, physical inference implementation and reproduction commands.

Method and full results · Downloads and commands

The study uses structured observations, nine explicit affine hypotheses and one fixed workspace refinement per event. It does not yet demonstrate general autonomous computation scheduling or cross-domain scientific learning. All 19 focused inference and integration tests pass.

Joint activity and learning across experimental tasks

This release jointly trains the scientific-task core, internal activity selection, experience retrieval and fast/slow prediction updates. It includes all eight final models, training histories and twenty complete native evaluation conditions.

Results

Every condition uses the same 184 instances in twenty changing-relation streams. Training repetitions start from one shared parent and are reported separately.

Policy	Completed	Native actions / instance
Full model, repetition 201	179/184	148.58
Full model, repetition 202	174/184	193.76
Calibration-rehearsal reference	179/184	116.32
Explicit relation-memory reference	180/184	128.42

Learned activities save operations relative to the fixed complete schedule in both repetitions. The full models do not improve on the strongest reference. Continuing state helps some consecutive instances of an unchanged relation, but increases the cost of the first instance when an earlier relation returns.

Online prediction updates execute during deployment and improve later operation-cost forecasts. They can change experiment choices, but retained and discarded updates complete exactly the same instances in all four same-weight comparisons. This separates prediction calibration from useful experimental learning. Physical relation fitting and precise settings still use the numerical tools.

All 3,680 native instance executions completed without execution errors. The five exploratory deployment checks are distinguished from the fifteen primary conditions. Model size is 42.93 million stored parameters, with 27.82 million trainable shared parameters; this study uses structured observations.

Model, learning objective and commands · Complete results and paired intervals · Download and execution guide

Downloads

• seed201-inference and seed202-inference: four final policies per archive and their numerical prior, approximately 467 MB each.
• The corresponding models archives retain outer optimizer state.
• cdm-joint-activity-records: all training histories, native trajectories, predictions, comparison plans and shared simulator randomness.
• cdm-joint-activity-data-records: the prepared measurement data and calibration worlds.

Archives restore project-relative paths. Each has a file list. The inference package plus records suffices to reproduce a trained native policy; retraining starts from the parent in v0.9.0. Visual encoder and language-model weights are not required for this experiment.

v0.9.0 — Continuing state across experimental instances

This release trains the continuing core across sequences of scientific experiments. Each instance has fresh objects, measurements and calibration coefficients. Only the GRU and Transformer state carry over; numerical records restart.

Four matched trainings are followed by ten native evaluations on the same 184 Reactor Lab instances in 20 changing-relation streams.

Condition	Repetition 191: completed / mean actions	Repetition 192: completed / mean actions
Continuing state	179/184 · 138.79	176/184 · 141.52
Same weights, task-boundary reset	177/184 · 206.88	174/184 · 207.34
Frozen rehearsal parent	175/184 · 117.16	177/184 · 118.10
Explicit last-successful-relation memory	177/184 · 132.22	Shared evaluation

Retaining state reduces actions by 32.9% and 31.7% compared with resetting the same weights. Savings concentrate in subsequent instances of the current relation. First-return performance does not establish recovery of an earlier relation after interference. The new trainings do not consistently outperform the frozen parents or the compact explicit-memory reference. Native traces locate most added work in tests and device adjustment.

Full results and paired comparisons · Training method · Download and reproduction commands

Downloads include four final policies with and without optimizer state, the numerical prior, training histories, all native trajectories and shared simulator randomness. The 449 MB inference package uses structured public observations and needs no visual encoder or language model. The record archive preserves every original relative path while storing identical random-state files once.

The chemistry archive adds 15 successful reference trajectories and 262 recipe experiments, including unsuccessful trials. It is an environment/data resource, not a trained chemistry-policy result.

Neural parameters remain fixed during these native evaluations. Adaptive computation, deployment-time parameter learning and consolidation are not results of this comparison.

v0.8.0 — Active experimental learning

The continuing core now learns which measurements to obtain and when to test a setting. Training updates 26.65 million parameters, including the inherited GRU and Transformer workspace.

Calibration rehearsal constructs alternative tasks from the same measured quantities and public support calibrations. Two matched continuations from independently trained parents each complete 62/64 native Reactor Lab instances; ordinary continuations complete 54/64 and 60/64. Both rehearsal policies match the full-measurement control's exact completed instances with 53.1% and 54.6% fewer native actions. They can spend more exploration on cases already solved by ordinary training; all costs and failures are reported.

• Method and training commands
• Active-measurement comparisons
• Independent rehearsal repetitions and paired data
• Download, extraction and native execution

Start with cdm-active-inference.tar.zst (253 MB) to run either final policy. It includes both policy weights and the numerical prior, with optimizer state omitted. The compact execution example runs on CPU without downloading vision or language models.

The model and prepared-record packages contain the starting models, trained policies, datasets, training curves, native trajectories and shared evaluation randomness. The separate evidence archive preserves original public measurement histories. Join numbered archive parts before extraction; each archive has a content list. Paths are relative to the repository.

These experiments use structured observations, nine explicit affine hypotheses and a numerical setting tool. They establish active acquisition and transfer from generated relation practice. Adaptive computation and learning retained across scientific tasks are ongoing work.

v0.7.0 — Learning complete experimental choices

CDM now learns choices between complete experimental plans from their observed success and execution cost. A plan persists through setting a device and testing it; its result informs the next choice.

The study includes 302 matched experimental branches, four trained selectors, and 264 evaluations from task reset in DiscoveryWorld Reactor Lab. All outcomes, including failed choices, are retained.

Transfer without retraining

On 24 fresh paired instances with a 400-action budget:

Policy	Completed	Mean actions, including failures
Original action reader	19/24	267.88
Complete experiments, posterior order	21/24	262.13
Learned numerical selector, training seed 171	23/24	249.46
Learned numerical selector, training seed 172	23/24	249.46

Against posterior order, each learned selector gains three completions and loses one. Both trained weights produce the same action sequences on this cohort. The selectors transfer from a different action reader without further training.

On the earlier 24-instance cohort, the predictive parent completes 14/24, posterior-ordered experiments 16/24, and the two numerical selectors 17/24 and 18/24. Adding the continuing neural state gives 17/24 and 16/24; it provides no consistent advantage here.

Included

• All four 317,314-parameter selectors.
• Training curves, experimental designs, matched-prefix outcomes, and complete native evaluation records.
• Simulator observations and archive content lists.
• Code to collect experimental alternatives, train selectors, run both action readers, and regenerate the published result tables.

The continuing core, perception, and numerical prior are fixed in this study. Exact numerical tools remain part of execution. Adaptive internal computation and cross-task neural learning remain open research questions.

Results · Protocol · Download and run

The transferred policy uses the compact reference model package from v0.5.0 plus the new selector. Reproducing the predictive-parent comparison also uses v0.6.0. The release guide gives the exact commands. Pretrained perception is downloaded separately from its original publisher.

v0.6.0 — Action-consequence learning

This release adds joint prediction of action consequences, forecast feedback to a persistent workspace, and learning from paired native interventions.

Four matched training runs are complete, together with 216 from-reset native evaluations, 96 alternative-action continuations and six computation interventions. The decision network has 53.97 million trainable parameters; visual perception and the explicit numerical tool remain fixed.

Comparison	Result
First cohort, 24 instances	Reference 18/24; both prediction-training arms 16/24; removing extra workspace cycles from the feedback model 9/24
Forecast-reading intervention	Removing forecast scoring and feedback retains 16/24, with identical action sequences in 21/24 instances
48 paired training decisions	No pair changes final completion or progress; 34 differ by one action
Second cohort, 24 instances	Reference 17/24; predictive parent, matched replay and paired-consequence learning each complete the same 16/24

Paired learning improves return prediction but not task completion. These results distinguish reliance on additional computation from useful forecast-driven experimental choices. The original successful reader remains the reference. The two cohorts are separate and are not a longitudinal score comparison.

• Results and per-instance tables
• Methods and training commands
• Download and loading guide

The six archives contain model checkpoints, training and execution records, and observations for both studies, including failures. Each has a file list. Archives restore repository-relative paths below runs/discoveryworld/. The v0.5.0 reference package supplies shared source models; its original demonstrations are also needed to reproduce all training. Third-party pretrained weights and regenerable feature caches are obtained separately as described in the guide.

Scientific-task learning: policies, experiments and models

This release brings the completed visual scientific-task research into the main repository, with the methods, trained policies, comparisons and materials needed to continue it.

• Two action-reader seeds complete 14/16 and 13/16 Reactor Lab tasks from reset across the two reported cohorts. The method combines persistent state, observed experimental facts, an affine numerical model and a setting/probe tool.
• Evidence controls, workspace interventions, numerical-learning comparisons, feedback-based training and unsuccessful routes are included with per-instance results.
• Joint state/action training and a learned computation allocator are implemented and evaluated. The latest separate cohort does not establish an adaptive-computation advantage.
• Observation collection incorporates PR #3's in-memory rendering. The integrated code passes 225 tests.

Results · Download and reproduce

The archives contain saved checkpoints, structured experiment records and simulator observations. Concatenate numbered parts before extraction. They restore relative paths under runs/discoveryworld/. Regenerable feature caches and upstream pretrained weights are excluded. Earlier process-model artifacts remain in v0.2.0–v0.4.0.