Results · Downloads · Progress · Status · Research · Model · Experiments · Run · References
CDM studies how a model can keep perceiving, reasoning, acting, and learning while the world changes. Internal state persists across interactions. The model learns external actions and the choice of its next observation or computation. A conversation is one event in that continuing process.
An observation can become outdated while inference is still running. The research question is how to preserve useful computation, incorporate new observations, and spend additional computation where it improves subsequent decisions.
Current research
Current model and research direction · Results and corrected calculations · Closest methods and comparisons
The completed joint-learning comparison trains response hypotheses from subsequent action returns together with the shared core and both task interfaces. Twelve full continuations compare three conditions across four repetitions from two parents. The fixed test has 1,008 primary task instances, including untrained response mechanisms, target ratios and ingredient counts, plus 144 paired naming replays. Original physics and chemistry supply 5,760 native retention executions.
Broader training improves new-task performance: the four joint models complete 952–961/1,008 tasks at 241–246 mean operations, against the parents' 867/883 completions at 327/321 operations. Matched inherited continuation reaches 952–966 completions at 236–249 operations. The new action-gradient connection does not establish a stable advantage over these simpler continuations. Within the joint models, the learned response forecast is more accurate on selected experiments than the inherited forecast. Native costs generally increase. The original model with twelve ledgers remains the recommended native reference.
!Broader-task transfer and native retention
Dots are endpoint means; lines show the range across runs, not confidence intervals. The full results include completion counts, paired intervals, task groups and all controls.
v0.17.0: models, training inputs and complete records · Download and reproduce
The preceding function-belief study learns coherent alternatives for unfamiliar relationships. Its small response models improve on attentive conditional predictors and reach the strong CDM's digital performance range. Hypothetical observation updates help a finite-cost planner, while a cheaper information rule remains competitive. v0.16.0 preserves that frozen-core comparison.
The response-law study completes six joint trainings, two 576-task digital cohorts and 3,840 native executions. It expands training to five response mechanisms and tests a withheld combination. Adding twelve experimental ledgers to the original models is an effective initialization; previous memory continuation is not needed to obtain that benefit. On the primary mixed-law cohort, retained observations reduce operations relative to removing them from the same trained models.
The new hypothesis-set action head does not improve on the marginal selector and harms original chemical tasks. Removing it restores chemical performance. On independent digital streams, the resulting models average 155.18 / 156.93 operations versus their parents' 161.47 / 160.93, with both paired intervals including zero. The complete results preserve these limits, the simpler continuation controls and the stronger known-formula reference.
v0.15.0: response-law models and records · Download and reproduce
The integrated direction retains persistent inference, active experiments, learned computation and long-term consolidation. Response inference now enters joint action training. Stable decision gains, native retention and operation under an independent clock remain the next capabilities to establish.
v0.14.0: models and experimental records · Download and reproduce
The experimental-memory study asks how a continuing model can recognize when past experimental knowledge is useful again. It archives actual input–outcome records, reconstructs response predictors, and compares those predictors against new observations before choosing the next experiment. Current-task learning remains available when past relationships do not explain the observations.
Four joint continuations and 5,280 native executions are complete on twenty twelve-task streams. On returning chemical relationships, memory reduces mean operations from 88.67 to 59.33 and 83.07 to 64.23 relative to separately trained no-memory models. Both memory and no-memory models complete all 240 chemical tasks. New relationships incur extra cost, and only one continuation establishes an overall advantage. Fixed evidence weighting performs similarly to learned reading, and neural-state resets preserve the return benefit. Complete results include new/return effects, physical outcomes, direct recipe retrieval and the known-family Bayesian reference.
Four further joint continuations and 2,400 native executions test direct action-return training of the memory reader. That change does not establish an improvement over prediction-only continuation or the memory parents. Complete follow-up.
Archive retention is a separate, effective intervention. Expanding six task ledgers to twelve preserves relationships that were previously evicted before their return. Both continuations need fewer operations on long returns, with the same weights and evidence rule; short returns incur a small extra cost. Capacity comparisons report the initial cohort and new instances separately. New relationships remain the main source of inappropriate reuse.
This follows the completed response-learning comparison. A learned relation-conditioned update makes an experiment improve predictions for related experiments. After joint training, retaining that update reduces chemical operations by 8.5% and 9.8% relative to removing it from the same weights, while completing all 120 tasks in each condition. Superiority over a separately trained calibration reference remains unestablished. The new memory study tests whether useful learning can survive task changes.
The project measures three complementary capabilities: using ongoing computation to improve action, learning from the model's own returns, and recovering useful knowledge after interference. Their individual results are tracked in HANDOFF.md; a combined advantage is still a research objective.
Earlier measured results
The scientific-task studies train the continuing core to choose what to measure and when to test a setting, expand practice through alternative calibration relations, and jointly learn physical prediction and experimental action. Earlier activity scheduling and online cost-prediction studies remain part of the record.
Results · Method and reproduction · Models and experiment records
| Study | Result | Interpretation | | --- | --- | --- | | Physical prediction and action, two independent parents | 117/120 and 116/120 native completions, with 105.63 and 106.93 actions per task; matched parents both complete 115/120 with 117.18 and 125.91 actions | The complete training method improves native action efficiency. Same-weight memory interventions do not establish reliable old-context recovery. | | Active measurement, two training seeds | Both persistent models complete 52/52 new instances from the training relation families, using 100.15 and 90.69 native actions on average; full measurement needs 242.13 | The trained policy learns to acquire useful evidence before collecting every measurement. | | Continuing-state comparison | Across all 64 instances, persistent models complete 53/64 and 60/64; separately trained ledger controls complete 51/64 each | Continuing state affects experimental decisions. Resetting the same trained weights gives 51/64 and 54/64. Transfer to the reserved relation varies across training seeds. | | Alternative calibration practice, two independent parents | On a separate 64-instance cohort, both rehearsal models complete 62/64, versus 54/64 and 60/64 for matched extra training | Both match full measurement's exact completed instances, with 53.1% and 54.6% fewer native actions. Both complete all 12 instances of the relation without original native training examples. | | Experience across experimental instances | On 184 instances in 20 changing-relation streams, carrying trained state completes 179/184 and 176/184, with 32.9% and 31.7% fewer actions than resetting the same weights at task boundaries | State reuses recent experimental experience. There is no consistent overall advantage over the frozen parents, and no clear efficiency gain on the first return to an earlier relation. | | Joint activity and learning, two continuations | On a new 184-instance cohort, full models complete 179/184 and 174/184, using 148.58 and 193.76 actions per instance; calibration rehearsal completes 179/184 with 116.32 | The shared training path runs, but the new models do not improve the quality–cost tradeoff over this reference. Online cost calibration and experimental decision quality must be evaluated separately. | | Learning during process control | Retained online updates give error 0.153 ± 0.003, versus 0.307 when discarded at matched compute, over three training seeds | Training through the updates used during deployment makes online learning useful in the controlled process tasks. |
The scientific experiments use nine explicit affine hypotheses and a numerical setting tool. Rehearsal holds recorded measurements and public calibrations fixed, generating alternative target relations for training. The relation without native training examples is therefore practiced through generated hypotheses. All policies run from task reset and pay for navigation, measurement and adjustment. The matched controls, failures and instance-level data are retained in the results.
Earlier scientific studies established autonomous execution and learned selection between complete setting experiments. They also found that additional forecasting or computation did not reliably improve decisions. The full record preserves those findings. The task-stream comparison measures experience carried between instances. The subsequent joint comparison connects activity selection and parameter learning to the same policy. Useful adaptive computation, rapid recovery after interference and learning that improves later experimental decisions are now studied through the physical-belief comparison. Useful autonomous computation and recovery across broader task families remain open.
Model
The joint-belief scientific-task checkpoint stores 47.96 million parameters; 32.85 million are trainable in its continuation. It shares core weights between physical control and chemistry and retains a separate neural state for each interface. These comparisons use twelve experimental ledgers, with up to 32 returned observations per ledger. The response-function model has 106,500 parameters and the added belief-to-action head has 37,761. This path uses structured public measurements and numerical tools. The earlier image/text model is a separate experimental configuration described below.
A shared-core implementation connects the physical and chemical interfaces to one set of core weights, with either continuing or separately retained task states. The completed shared-core and active-feedback studies jointly train both interfaces and evaluate native task performance. Earlier task-specific results remain separate comparisons.
The active-experiment policy reads structured public measurements and operation feedback. Training updates 26.65 million parameters, including 23.11 million in the inherited GRU and Transformer workspace. State continues across measurements and the two experimental targets. The frozen visual encoders are not used in this particular comparison.
The physical-learning model jointly updates that action core and 73 shared parameters defining physical priors, uncertainty and context retrieval. Bayesian posterior updates run during deployment and change the setting predictions. The action network uses nine physical-context tokens when deciding what to measure. This study fixes one workspace refinement per event and uses structured inputs.
The earlier integrated activity-and-learning model stores 42.93 million parameters, of which 27.82 million participate in joint training. Its per-stream fast and slow adapters add 128 online-updated scalar values. The new policy chooses external experiments and internal refinement, retrieval and learning activities; the numerical tools still fit physical relations and execute precise settings.
The earlier multimodal policy combines frozen SigLIP image/text features with a persistent state, an action reader and exact experimental records. Its decision network has about 0.053B parameters; frozen perception adds about 0.375B. The successful early-reader comparison trained only 442,753 parameters; later studies jointly trained the recurrent core and its task interfaces.
DESIGN.md specifies the broader model, including activity selection, episodic retrieval and continued learning. The on-demand Qwen3.5-2B semantic interface is implemented separately; it is not used by the successful scientific policy. Qwen3.6-35B-A3B is a proposed offline teacher and has not generated the current demonstrations. TRAINING.md distinguishes these routes from the completed experiments.
The existing trained studies use these implementations:
- Sequence core (cdm/model.py): a GRU context over observations, goals, environment descriptors, previous actions and elapsed time; a workspace of latent slots refined by one shared Transformer block; a bounded record bank that stores intermediate work with its generating context, goal, observation-prefix index, parameter version and depth; conditioned reading with slotwise reuse, revision and initialization gates; action, activity and value heads.
- Spatial planner (cdm/planner.py): a convolutional workspace over a grid, refined by one gated block and read out as an action and a distance for every cell. After a change it can restart, carry, retrieve stored workspaces, or read them with learned attention; it can read its workspace at the goal cell, rebuild channels through a learned keep-or-rebuild gate, split its workspace into map and goal parts, and read the age of every cell of a partial map.
- Continual agent (experiments/continual_agent): the spatial planner with an information state (the last seen map and the age of each cell) and an activity head that chooses each tick how many refinements to run, whether to request a full observation and whether to move, trained on reward minus computation and observation costs. Computation it leaves unused pays for online learning of the planner.
- Continual core (cdm/core.py, experiments/lifetime): one model that runs a process for a whole lifetime. A GRU fast state, a persistent workspace refined by a shared Transformer block over recent experience, and predictions of the next reading for every action through an instance-owned adapter feed heads that choose, every tick, the control, a precise measurement, and whether the compute until the next tick goes to refinement or to an update of the adapter. Trained with recurrent PPO over instrument and thermal process lifetimes; also run against processes on their own wall clock.
Implemented and measured
| Study | Implementation | Result | | --- | --- | --- | | Neural pilot | Five seeds, three model variants, supervised sequential contextual recall | Query accuracy: memory 56.20%, persistent workspace 36.03%, restarted workspace 38.44% | | Dynamic grid | Five seeds, three change regimes, two computation budgets; world advances during planning | Reuse helps at low budgets; this local factorization does not outperform nearest-context retrieval | | Numerical transport | Changing quadratic objectives and finite-depth state reconstruction | Tested correction loses to strong solvers; reconstruction quality and task quality differ | | Changing maze | Neural latent planners act every tick while doors, obstacles and goals change; 14 variants with five seeds and 3 partial-view variants with three; shared budgets | A read of the workspace at the goal cell lets a goal change reach every cell in one refinement (affected cells right: 92.6% against 43.0% for carry); a learned keep-or-rebuild gate makes 83% right before any refinement without a hand-made split; at the training size the best planner earns the optimal policy's return within 0.01 with one refinement per tick | | Continual agent | One running process over a stream of changing sites: partial view, asynchronous surveys, requests with idle periods; PPO-trained choice of refinements, surveys and moves; learning from unused compute | Three PPO seeds per planner: the learned activity head beats the best hand-written schedule on two of three planners (+0.048, +0.014) and ties on the third, with 42–51% fewer refinements and more surveys. Learning paid from unused compute raises the running agent's reward on larger mazes with a full view (+0.026, six paired runs); with a partial view it has not improved reward | | Lifetime | One 35.9 M-parameter model over instrument, thermal and vehicle lifetimes with unannounced, returning condition changes. First round: eight settings, five seeds of 2·10⁷ interactions each. Second round: the model trained through its deployed learning steps, workspace across events and copies of processes, three seeds; held-out dynamics; three families; control against an independent clock with power | First round: prediction supervision lowers tracking error from 0.146 to 0.109 without the policy reading a prediction; refinement and online learning add nothing, and the instance's updates slightly hurt. Second round: trained through the learning steps it actually takes, the model's own updates halve its error against discarding them at matched compute (0.153 against 0.307), on held-out dynamics and against a real clock. Unrolling the workspace and an adapter bank did not help; activity-value guidance was not effectively tested; retention of returning conditions and proactive work are not yet achieved (phase review) |
STATUS.md maps the full objective to code, measured results, and remaining work. Physical-learning models and complete records are in v0.11.0. Joint-activity models, prepared data and complete comparisons are in v0.10.0. Task-stream models are in v0.9.0. Active-learning models and experiment records are in v0.8.0. Earlier checkpoints and evaluations are in v0.2.0, v0.2.1 and, for the continual core, v0.3.0 (first round) and v0.4.0 (second round).
The neural pilot trains action prediction with a fixed memory-write schedule. The changing-maze planners and the continual agent run on GPUs; their training, evaluation and analysis commands are in the maze and agent READMEs.
Run
Python 3.12 is the recorded environment. The numerical experiments require NumPy; the neural core uses PyTorch. Installation, CPU reproduction, CUDA commands, and profiling are in RUNNING.md.
python experiments/continual_grid/run_experiment.py --output-dir runs/continual-grid
python experiments/neural_pilot/train.py --device cpu --output-dir runs/neural-pilotEach command writes configuration and raw measurements to its run directory. Archived measurements remain alongside their experiment.
Research map
Start at HANDOFF.md: what the experiments established, what failed and why, what is implemented against what is specified, and the next steps.
| Document | Contents | | --- | --- | | HANDOFF.md | Research progress, evidence, unfinished questions and next steps | | STATUS.md | Capability coverage, measured performance, released artifacts, remaining research | | research/PHASE_REVIEW.md | Where the project stands after two rounds: goals, results, unsuccessful methods, hypotheses and the next phase | | RESEARCH.md | Objectives, proactivity, continuous operation, learning, applications | | DESIGN.md | State, model operations, interfaces, and execution | | TRAINING.md | Data, losses, activity learning, adaptation, GPU training | | EXPERIMENTS.md | Current studies, primary comparisons, instrument task, scaling | | RUNNING.md | Executable commands and next implementation steps | | ref/ | 122 source records, reading depth, and closest related methods |
Target applications include robotics, electromechanical control, autonomous experiments, edge devices, and digital research. They require timely decisions, persistent context, and learning across continued interaction. Eidolve is a related knowledge and memory project; CDM can also connect to other tool and memory environments.
Original code and documentation use the MIT License. External papers and implementations retain their own licenses. CITATION.cff identifies this project; include the commit used when reporting an experiment.