Adapt-1 Preview API unlocks two new learning rules today: Temporal Context Projection (TCP) and Counterfactual Utility Plasticity (CUP). To put it simply: they allow reward signals to update two structures that are usually fixed—the state representation during decision-making, and the relationships between internal predictors.

## Useful Components Can Fail Together

A planner can exploit small errors in a learned world model to search for a trajectory that looks optimal only within the model, then pass this impossible state to the value estimator, which confidently assigns it a high score. Every component works fine on its own, but the whole system breaks when combined. The problem doesn't lie in any single component—it's that the relationships between them have never been learned.


**Figure 1** Compositional outcomes. The world model, planner, and value estimator each produce useful predictions. In one scenario, short rolling horizon plus value estimation yields positive joint utility; in another, planner search amplifies model error, and the estimator rewards an unreachable state.

Conventional online learning updates parameters, but leaves the computational structure fixed. It can modify policies, fit models, and update stored states, but the architecture generally predefines what information goes into the current state and how internal predictors interact. Optimization only starts after these declarations are made.


**Figure 2** Declared structure. Online updates change memory statistics, model parameters, policy states, and source reliability. But the declared state representation and combination rules remain fixed.

## Observation-Equivalent Process States

In partially observable environments, the current observation may not uniquely identify the context that produced it. Two time steps can have identical observations but require different actions, because their transition histories produce different process states. State aliasing merges states that the task needs to keep separate.

A software agent receives the same terminal frame after two different causal trajectories. A failed test may require rolling back dependencies after a dependency update, but may require fixing an interface after a local interface edit. Optimizing only the final error leaves the missing process state unavailable.


**Figure 3** Process state ambiguity. A software agent receives the same terminal frame after two different causal trajectories. The trajectory determines the process state corresponding to the observation.

## TCP: Enabling Bounded History to Participate in Decision-Making

TCP solves this representation boundary problem. Before decision-making, it allows finite, causally ordered observations from the same episode to participate in the current representation. TCP introduces eligible history before decision-making, but does not pre-declare reward lags. Adapt-1's existing underlying infrastructure discovers which projected contexts predict measurements through conventional online updates.


**Figure 4** State construction. Maximum lag bounds the available causal history before decision-making. Conventional learning discovers which eligible history fields predict later feedback.

### Performance on RepeatPrevious

POPGym's RepeatPrevious turns state aliasing into a controllable task. On Medium difficulty, the correct action at time t is the observation at time t−32. Giving only the current input is inevitably insufficient. The domain exposes a general maximum time lag of 64, but does not declare reward lags. TCP incorporates bounded history into the current representation, and conventional online learning discovers this 32-step dependency from outcomes.

Five TCP-only runs of 10 episodes each give a total of 1030 interactions. The final episode has an average return of 0.9444. Run with seed 0 entered frozen evaluation (with feedback, writing, and adaptation turned off), scored 1.0 on 16 held-out episodes, and showed an invariant state fingerprint.

Easy difficulty provides a causal ablation. The condition with only current input ends up with an average return of -0.5109 and action accuracy of 0.2445. All five TCP runs reach 1.0 return and accuracy, and frozen evaluation retains perfect performance. TCP + CUP also reaches the same upper bound. Since TCP-only already hits the upper bound with lower probabilistic loss, the return gain is fully attributed to TCP.


**Figure 5** Causal state results. TCP changes the decision representation and produces return gains. TCP-only already reaches the upper bound, so CUP has no causal attribution to the 32-step solution.

Published results provide context for difficulty and sample budget. In the original POPGym study, 12 out of 13 models had negative MMER after 15 million time steps on Medium, with LMU being the only positive exception. A later TMLR table reports that vTransformer and GRU reach 1.0 after 1 million environment steps on Easy. Adapt-1 TCP reaches 1.0 after 510 online interactions. Protocols differ across optimization, sources, metrics, and uncertainty reporting, so direct ranking requires normalized evaluation.

| System | RepeatPreviousEasy | Train & Eval Protocol |

| --- | --- | --- |

| Adapt-1 TCP | 1.000 ± 0.000 SD | 510 online interactions/run; 16 frozen episodes × 5 runs |

| vTransformer | 1.000 ± 0.000 SE | 1,000,000 environment steps; 16 test episodes × 5 seeds |

| GRU | 1.000 ± 0.000 SE | 1,000,000 environment steps; 16 test episodes × 5 seeds |

| Mamba | 0.993 ± 0.001 SE | 1,000,000 environment steps; 16 test episodes × 5 seeds |

| Published No Memory | −0.434 ± 0.013 SE | 1,000,000 environment steps; 16 test episodes × 5 seeds |

In matched experiments, changing the representation boundary makes alias mapping learnable. TCP provides temporal identity, and Adapt-1's existing underlying infrastructure learned the 32-step dependency.

## CUP: Enabling Relationships Between Predictors to Be Learned

Robot failures expose the composition problem. Relationships between predictors can add value beyond the sum of their parts, but they can also duplicate evidence or introduce systematic conflict, and their effects change with the environment. Even accurate predictors can still replicate the same upstream error when combined. Breiman's random forest analysis formalizes this strength-correlation tradeoff.

Reliability scores for individual predictors cannot capture these conditional effects; the missing utility belongs to the relationship itself.

CUP enables learnable relationships between the prediction sources registered by an agent. Conventional posterior learning estimates the credibility of each source individually. CUP goes further: after sources submit predictions and outcome feedback for the decision link arrives, it asks whether a combination performs better or worse than what its members and all smaller subsets can already explain. It accumulates relationship-specific evidence over time.

When evidence is strong enough, beneficial relationships are consolidated into bounded persistent states and can change subsequent posterior fusion. Harmful relationships are marked for suppression, and old evidence decays to allow revision when the environment changes. CUP not only learns which sources are reliable, but also which sources are complementary, redundant, or jointly harmful.

### Persistent Marginal and Joint Utilities on RepeatPreviousMedium

RepeatPreviousMedium exposes CUP state across multiple registered predictors. TCP and the adaptive underlying infrastructure are causally responsible for the 32-step solution. The exported CUP state retains the marginal and joint relationships of two correlated sources: contextual_memory and learned_model.

In seed 0, learned_model is consolidated with positive utility, while contextual_memory × learned_model is suppressed. In seed 3, contextual_memory alone is suppressed, while the joint combination is consolidated. The marginal utility of a source can have the opposite sign to its conditional utility.


**Figure 6** Inspectable topology. CUP learns independent persistent utilities for marginal and joint relationships. One run suppresses the joint combination, while another consolidates it.

Across five paired ten-episode Medium runs, TCP+CUP increased the average learning curve return from 0.2839 to 0.3167, a relative improvement of 11.5%, and won four out of five seeds. The average of the last five episodes rose from 0.7578 to 0.8067. The paired 95% confidence interval crosses zero, and both variants reach 1.0 in frozen testing on seed 0.


**Figure 7** Acquisition interval. Across five paired runs, CUP improves the average acquisition curve. The paired 95% interval crosses zero, so the acquisition effect remains undecided.

Five runs are not enough to resolve the acquisition effect, but they establish the structural result: while the main learner continues operation, CUP learns persistent marginal and joint effects between attributable predictors.

### Adaptive Topology for Simple Control

The Wisconsin breast cancer diagnostic dataset isolates online predictor combination with no long memory requirement. The first 12 declared features generate singleton and pairwise coalition predictors. Across 30 randomized prequential orders, each probability is issued before the label arrives. The first 100 observations are excluded as warmup, and feature selection is independent of labels.

Only the combination rule changes across the matched ladder: exact disable fallback, uniform fusion over the same coalition library, selecting the single coalition with the lowest past prequential loss, or CUP's adaptive topology.

CUP improves probabilistic mass over both the disable path and fixed uniform fusion. For uniform fusion, it reduces average log loss by 0.0201 across all 30 orders. Fixed uniform fusion retains a balanced accuracy advantage of 0.0131.


**Figure 8** Matched combinatorial controls. CUP improves probabilistic mass over uniform fusion in every random order. Selecting a single coalition by past loss achieves lower log loss, and fixed uniform produces the highest balanced accuracy.

Best coalition selection has an average log loss of 0.2493, while CUP has 0.3087, and selection has lower loss across all 30 orders. Its balanced accuracy advantage over CUP is small and statistically undecided. In this stable, highly redundant setting, selection is more appropriate for log loss.

Selection does not erase CUP's structural result. CUP learns source relationships online and improves log loss over fixed fusion in every order. The control shows that adaptive topology must be compared against both fusion and selection.

## TCP and CUP in Agent Problems

TCP handles process state ambiguity in agents. Long-running agents encounter identical interface observations produced by different process states. A failed test after a dependency update may require rolling back dependencies, while the same test after a local schema edit may require interface fixing. A latency alert after deployment and a latency alert after a traffic surge also require different interventions.

Transcripts retain history as stored data. TCP makes bounded committed history available before decision-making, allowing the underlying learner to discover which temporal distinctions identify the next action. Coding agents can learn which earlier edit defines the current failure; embodied agents can use recent contact to disambiguate the current sensor frame. Trajectory information erased from the current observation thus enters the decision state.


**Figure 9** Learned process states. TCP makes bounded committed history available to downstream learning before decision-making. The learned representation can distinguish terminal observations produced by different trajectories.

CUP acts on interaction-specific predictive utility between registered sources. Sources submit predictions before outcomes. Measurement feedback evaluates their combinations and writes the resulting relationships into persistent topology.

Shared upstream signals can cause two predictors to repeat the same error. Tool-derived sources can fix blind spots in learned models for specific mechanisms. Marginal source utility can differ from conditional utility. CUP assigns persistent state to every relationship. Agents can use it to discount correlated agreement, consolidate combinations that add predictive value, suppress harmful combinations, and revise relationships when the environment changes.

Every relationship retains its participants, order, state, support, and uncertainty. When debugging an agent, you can inspect learned interactions, the direction of utility, evidence supporting the state, and responses to mechanism changes.


**Figure 10** Persistent interaction utility. CUP writes interaction-specific utility between registered sources into persistent topology. Outcome updates assign supported or suppressed states, and revise outdated relations after utility changes.

Hidden process state ambiguity and interaction-specific source utility exist in software development, incident response, scientific control, long-horizon navigation, and robotics. Outcome feedback can update both structures while the base learner continues operating. During continuous deployment, outcome updates can admit previously missing temporal dependencies into the current state, or change the persistent impact of predictor relationships. Domain constraints bound both updates; exported states retain the evidence needed for inspection.

Reproduce the experiments: https://github.com/0xReisearch/adapt1-behavioral-study

Original blog post: https://reilabs.org/blog/unlocking-plasticity-rules