// Introducing Adapt-1 Preview: A Pretraining-Free Substrate for Test-Time Learning

A research release of Adapt-1 Preview, with formal academic publication to follow.

Adapt-1 Preview (Δ1) is the first implementation of ADAPT, a non-transformer neuro-symbolic substrate built to learn while it is operating. Core begins without task-specific pretraining, does not generate tokens, and does not require a language model in its decision loop. An application declares a Domain that names what may be observed, selected, predicted, and measured. Core forms persistent task-relevant state from the resulting stream.

Depending on the Domain, measured outcomes can change the next choice or connect an earlier action to a later consequence. The same persistent Core can operate under partial observation, explore alternatives, infer hidden task rules, adapt as conditions change, and reuse structure when conditions recur. These are not product modes or separate systems. The Domain determines which evidence may update the state and which measured outcomes may change a later result.

External models and solvers may supply perception or consume Core's structured state. They do not own its decision state or learning loop. Adapt-1 can serve an application directly, remain frozen while composing typed evidence, update online inside a partially observed stream, or retain grounded feedback across repeated interaction.

Adapt-1 is being released as a research preview for evaluation, experimentation, and early integration. This report covers its architecture, mechanisms, controlled evaluation, and three external evaluations.

1. Design Goal

Operational learning and inspectable state

Adapt-1 is designed to update after deployment while preserving the task definition, training signal, and provenance of later decisions. Its architecture follows three constraints:

1. Declared task contract. A Domain names admissible observations, alternatives, measured outcomes, prediction targets, and evidence gates. No learning-system tuning is required to begin; the Domain is the public task contract.
2. Outcome-gated updates. Evidence may support a decision, but only an eligible observation or an outcome attributed to the producing decision may train the state that produces the next one.
3. Inspectable state changes. The mutation state, evidence route, uncertainty, state version, and producing learner remain inspectable whether the Core is frozen or adaptive.

These are constraints on one architecture, not separate product modes. The same state must support a cold read, an online update, and a later replay without changing the meaning of the task or the ownership of its evidence.

2. Architecture at a Glance

Adapt-1 treats inference and learning as two views of the same persistent state. Domain defines the task contract and vocabulary. Core maintains the task state, evaluates evidence through installed learners and bounded relations, validates their influence, and selects, predicts, or abstains. Measured outcomes return through an attribution gate, allowing Core to update only the state implicated by the decision's provenance.

1. Task contract. A Domain names admissible observations, alternatives, measured outcomes, prediction targets, and evidence gates. Core manages learner selection, validation, calibration, and policy behavior internally, leaving specialized controls as service defaults rather than setup work for the user.
2. Core execution. Adapt-1 retrieves typed evidence, follows bounded relations, estimates supported alternatives or consequences, and returns either a selection or an explicit reason to abstain.
3. Outcome update. Eligible observations and attributed outcomes update only the state implicated by their provenance, so feedback can change a future result without allowing the system's own confidence to train itself.

Direct execution returns structured decisions, predictions, uncertainty, and provenance without an LLM in the loop. External models may consume the same state for flexible synthesis, but they are not required for Core execution or learning. HTTP, Python, compatibility, and Model Context Protocol clients all invoke the same session-owned Core rather than separate learners.

2.1 Mechanism map

Remembering an event is only the first stage of adaptation. A useful stateful system must decide what co-activation means, when a record becomes stable, how conflicting evidence should be represented, and which measured outcome is allowed to change a future decision.

A system can retrieve relevant episodes and still fail to discover the structure that connects them. It can rank a declared policy and still miss a new regime. It can report high confidence and still provide no evidence that confidence predicts correctness.

Adapt-1 addresses these problems through typed learning surfaces. Co-retrieval updates short-lived associations. Consolidation stabilizes selected episodes. Hypergraphs preserve multi-way context. Neurosymbolic reasoning combines continuous similarity with explicit paths. Domains define the observations, alternatives, rewards, and prediction targets that make adaptation executable. Autonomous structural memory, recurrent latent representations, explicit credit routing, versioned background learning, and prequential measurement extend that adaptation beyond the accumulation of records.

The update rule is asymmetric:

That rule prevents the system from treating its own retrieval confidence as proof that a result was useful. An internal state change counts as learning only when it can alter a future result and can be traced to eligible evidence.

Task Exposure Before the Result

A score is more informative when the task exposure required to produce it remains visible. Adapt-1 entered every reported external evaluation with clean Core state and no offline task training. Online interaction is reported separately from prior task training, and a frozen pass remains distinct from an adaptive run. This makes it possible to read each result in terms of both behavior and data efficiency.

Where an evaluation uses a pretrained external detector or a classical geometry solver, that support is named beside the result. It is not counted as Core task training, and it is not presented as task-naive perception.

Operating profile

• Non-transformer Core. Adapt-1's learning and reasoning state is not a language model and does not depend on token generation.
• Direct execution. Core can retrieve, explain, predict, select, or abstain without an LLM in the loop.
• Low ordinary-operation latency. In the controlled structural stream, p50 operation latency was approximately 95 microseconds; p95 was about 23.3 milliseconds with occasional induction work.
• Compact persistent state. Serialized state was approximately 105 KiB in the same controlled evaluation.

3. Plasticity, Consolidation, and Reconsolidation

Static retrieval asks which stored item resembles a cue. Adaptive memory asks three additional questions: which traces repeatedly appear together, which episodes deserve durable state, and when new evidence should revise an existing record.

3.1 Bounded Hebbian association

When an adaptive query returns a bounded support set, each pair in that final set can strengthen a Hebbian-inspired association. A newly created link begins near zero and approaches one under repeated co-retrieval:

The current learning rate is eta = 0.3. Without reactivation, the link decays continuously:

with lambda = 10^-4 per second.

The update is bounded in four ways. Only the returned support set is eligible, the number of updated memories per query is capped, each memory has a maximum neighbor count, and links below a floor are removed. These controls preserve associative plasticity without allowing a broad query to generate an indefinitely growing all-to-all graph.

A Hebbian link means repeated co-activation. It does not mean agreement, entailment, or causality. Contradictory records may become strongly associated because they are repeatedly consulted together. Adapt-1 therefore uses the link as a retrieval feature and retains each source record's own content and provenance.

Frozen reads disable this mutation. They retrieve from a stable state without changing access counts, consolidation, lability, or Hebbian association.

3.2 Fast and slow stores

New records enter a recent episodic layer. Promotion into consolidated state depends on surprise, importance, and repeated access. In the dual-store path, a record becomes eligible when the mean of surprise and importance reaches 0.7; access-driven promotion can also stabilize frequently used evidence.

This separation follows the computational motivation of complementary learning systems: detailed episodes can be acquired quickly while recurring or important structure is stabilized more selectively [1, 2]. Adapt-1 uses this as an engineering pattern rather than a claim of biological equivalence.

Consolidation changes storage status and future retention behavior. It does not transform an episode into universal truth. A consolidated record can still conflict with later evidence and remains addressable by its identifier.

3.3 Identity-gated reconsolidation

Near-duplicate writes are compared with a score that combines vector similarity and concept overlap:

where J is concept Jaccard similarity. Similarity alone is insufficient. Identifier-bearing tokens and metadata identities must also be compatible before the new write can revise an existing trace.

The identity gate matters because two linguistically identical observations may refer to different trials, devices, users, or moments. Without that gate, compression would silently merge distinct evidence.

Retrieval surprise can mark a memory as labile, but durable revision still requires an explicit correction or an identity-safe near-duplicate write. This maps the idea of post-retrieval lability from reconsolidation research into an auditable software operation [3, 4].

---

4. Retention Is a System Operation

Memory growth involves more than rows. A record may have a vector pattern, concept entries, hypergraph edges, Hebbian neighbors, consolidation state, and references from learned structures. Retention must therefore operate across dependent state.

Adapt-1 assigns each record a retention priority:

I is importance, C recency, A access support, and N novelty. Recency follows a configurable half-life, 30 days by default:

Access support rises with use and saturates rather than growing without bound.

Three processes should be distinguished:

• Association decay automatically weakens Hebbian links.
• Recency decay lowers one component of retention priority.
• Capacity pruning physically removes selected evidence and its dependent structures.

The default retention mode reports pressure. When pruning is enabled, exact duplicates and superseded evidence are structurally eligible. A unique memory becomes value-prunable only after measured execution feedback has been explicitly attributed to it. Unmeasured unique evidence is not assigned a fabricated neutral value.

Deletion is session-scoped and cascading. The primary record, vector entry, concept handles, graph references, Hebbian links, and derived-memory references are removed together. A memory clear applies the same cleanup to the owner's Core memory state without crossing into another session. Domain policy and adaptive state have separate reset or deletion operations because evidence reset, learner reset, and Domain-definition removal are distinct contracts.

---

5. Hypergraphs Preserve Multi-Way Context

A pairwise graph is natural for binary relations. Many events are not binary. An intervention, operating condition, component state, and outcome may jointly define one observation. Decomposing the event into unrelated pairs can remove the context that made it meaningful.

Adapt-1 maintains concept-level relations and typed hyperedges:

• recurring concepts become nodes;
• observations can generate typed edges or multi-node hyperedges;
• each relation retains supporting memory identifiers;
• local connected evidence can produce bounded abstractions;
• direct and depth-bounded searches expose candidate paths.

For a locally learned relation, support combines normalized recurrence and embedding compatibility:

Only relations above a threshold enter learned structure. A bounded one-hop graph encoder can then add structural context to memory ranking. The representation remains dominated by record content, with neighborhood evidence acting as a bonus rather than replacing the episode.

Hypergraph learning is related to formal work on higher-order relational data [5, 6], but the interpretation remains specific. A recurrent compatible edge supports association or prediction. It does not independently identify a causal mechanism.

Adapt-1 can construct causal-chain-shaped explanations from event types and temporal order. Their status remains explicit:

• an observed association records co-occurrence or a typed link;
• a predictive path records a useful temporal or relational pattern;
• a candidate cause requires intervention-before-outcome evidence;
• identified causal evidence additionally requires a declared interventional design and control or counterfactual support.

This hierarchy prevents a graph path from silently acquiring a stronger scientific meaning than its evidence permits.

---

6. Reasoning as Typed Evidence Computation

The neurosymbolic layer combines continuous representation similarity with direct symbols, bounded paths, and relation compatibility:

This is a pragmatic hybrid in the broad neurosymbolic tradition [7]. It is not a theorem prover. Its purpose is to make explicit relational evidence and continuous similarity contribute to the same candidate while preserving the source of each contribution.

6.1 The reasoning object

Adapt-1 represents a general reasoning result as five typed collections:

M_q contains retrieved memories, P_q relational paths, H_q candidate or induced hypotheses, X_q counterevidence and unresolved conflicts, and D_q a selection or abstention.

The system may later render these structures as prose. The underlying computation does not depend on prose generation. Vector retrieval, condition evaluation, graph traversal, posterior evidence, and abstention gates already have executable meaning.

6.2 The Core pass

A general query proceeds through five stages.

1. Cue construction. The query becomes normalized tokens, sparse concepts, and a hybrid vector.
2. Episodic recovery. Relational Hopfield Memory retrieves direct and associatively completed candidates.
3. Relational expansion. Hypergraph and neurosymbolic subsystems add bounded typed paths.
4. Connection discovery. Concept memory searches recurrent edges, multi-hop paths, temporal candidates, and optional learned links.
5. Evidence fusion. Graph and symbolic channels can strengthen only candidates linked back to stored evidence.

Positive channel evidence is fused monotonically:

b is base support, e_c is channel strength, and rho_c is the channel's learned reliability. Counterevidence and incomplete provenance are applied afterward as downward-only terms. Positive support cannot lower a candidate, counterevidence cannot raise it, and a channel without memory provenance cannot manufacture a result.

Channel reliability begins with a Beta prior and changes only after an execution outcome is attributed to the affected memory and retained fusion trace. The query cannot train its own fusion weights from its ranking score.

6.3 Mechanistic explanations

Core classifies retrieved records into roles such as observation, intervention, outcome, and conflict. It can order process stages, connect repeated stage failures across episodes, and propose mechanisms that account for the current evidence.

Each mechanistic hypothesis retains:

• evidence for and against it;
• predicted next observations;
• observations still missing;
• an epistemic label;
• a causal-identification flag;
• the records and paths from which it was constructed.

The output therefore separates two questions: what sequence is supported by the record, and what scientific status may be assigned to that sequence?

6.4 Replayable decisions

When Domain reasoning selects or rejects an alternative, Adapt-1 seals the decision. The snapshot includes a decision identifier, decision time, state version, state hash, response hash, context, alternatives, and supporting evidence.

Delayed feedback can consequently refer to the decision that was actually made. It does not have to reconstruct an earlier state from whichever memories and models happen to exist later.

This provides three levels of provenance:

• evidence provenance: which records and relations supported the result;
• model provenance: which learner versions and calibration states were active;
• decision provenance: which context and state identity produced the selection.

---

7. Domains Turn Memory Into a Learning Contract

Generic memory asks what stored evidence resembles a query. A Domain asks what should be estimated from a declared stream.

A Domain can define:

• event fields and validation constraints;
• competing hypotheses and their supporting or falsifying conditions;
• context features and candidate policies;
• a reward model over measured outcomes;
• transition inputs, actions, groups, and prediction targets;
• predicates eligible for structural induction;
• recurrent-set or interval fields;
• support, distance, uncertainty, and abstention thresholds.

The declaration configures a bounded set of adaptive primitives implemented by Core. It is not arbitrary executable code. That restriction gives every learner a known input contract, output schema, and reason for abstention.

7.1 Evidence-conditioned hypotheses

Declared hypotheses contain a name, supporting conditions, predictions, falsifying conditions, and an optional weight. Structured conditions can address paths in Domain events and use equality, inequality, ordered comparison, containment, existence, and missing-value operators.

During explanation, matched_conditions contains clauses matched by admissible evidence. falsifying_matches contains evidence that matched a falsifying clause. Unmatched declarations do not appear merely because they were registered.

Hypothesis ranking combines semantic applicability with execution evidence, but those meanings remain separate. A hypothesis may describe the current context while its associated policy has poor measured outcomes. A successful policy may also be irrelevant to the current evidence. Eligibility requires matched evidence or measured execution support; equal top scores produce a tie rather than an arbitrary winner.

7.2 Ownership is part of state

Domain definitions and all derived learners are session-owned. The owner boundary applies to evidence retrieval, policy samples, transition neighborhoods, latent regimes, structural rules, and interval state. Domain-scoped routes apply their path Domain automatically.

This makes isolation an architectural property rather than a metadata convention. A private record list would be insufficient if a global policy bucket or Domain registry still exposed another user's learned state.

---

8. Contextual Policy Learning

Retrieval estimates relevance. Policy learning estimates utility: which declared action or method has produced the best measured outcome in a similar context?

Adapt-1 keys policy evidence by Domain, relation, policy, and context. A relation is a caller-defined decision family. A policy is a candidate action or strategy. Context is a typed feature map selected by the Domain.

8.1 Reward is explicit

A Domain maps measured feedback into utility on [0, 1]. Each reward component can specify direction, acceptable range, hard-failure conditions, and weight. Under weighted-mean aggregation:

Geometric aggregation can penalize a policy that collapses on one dimension, while minimum aggregation implements a strict bottleneck objective.

When no declarative reward is present, the system recognizes explicit numeric reward, score, or utility fields, along with a bounded vocabulary of Boolean and categorical outcomes. An unknown label remains stored as feedback but does not update value-bearing state. Clients using a domain-specific numeric error therefore define how that measurement becomes utility instead of relying on the field name alone.

Delayed feedback can be discounted by decision distance. This reduces the effect of a remote outcome while preserving its link to the sealed decision. Temporal discounting is a credit heuristic; explicit attribution still determines which decision, policy, stage, or memory receives the update.

8.2 A local posterior over outcomes

For a candidate policy, Adapt-1 starts from a uniform Beta prior and accumulates weighted rewards:

The posterior mean and variance are:

Weights combine recency, relation match, policy match, action similarity, and context similarity. Same-relation, same-policy samples dominate. Cross-relation and cross-policy transfer are attenuated.

The result is Beta-like because rewards may be continuous and evidence weights fractional. It is a compact uncertainty summary, not an assertion that every task follows a Bernoulli process.

The raw posterior remains visible alongside prequential calibration. Once enough real outcomes exist, an online monotonic map can calibrate predicted reward against later observations. Diagnostics include Brier score, log loss, expected calibration error, reliability bins, Wilson intervals, risk-coverage points, and area under the risk-coverage curve.

8.3 Selection and exploration

The default selection mode exploits the highest supported expected reward. When an application explicitly permits exploration, the decision score can use an upper-confidence value:

or seeded Thompson sampling from the policy posterior. These methods make uncertainty decision-active in the contextual-bandit sense [8]. Exploration remains explicit so a production caller expecting the current best policy does not silently receive an experimental action.

8.4 Regime-aware learning

A global policy mean can fail when the same action behaves differently across operating regimes. Adapt-1 monitors context distance and reward residuals. Persistent residual change can open a new relation-specific regime; later evidence forms a local belief instead of immediately overwriting the established one.

Transition counts preserve recurring regimes. Missing configured features cannot trigger drift, change detection has a cooldown, and replay after a change point is bounded. The mechanism is related in motivation to Bayesian online change-point detection [9, 10], while using a compact engineering rule rather than a full posterior over run length.

When enough feedback accumulates, an offline worker can compare nonlinear reward models on a temporal validation slice. A candidate is installed only if it improves over a simple baseline. At inference, learned predictions are blended with the local posterior only under sufficient support, positive validation skill, and acceptable out-of-distribution distance.

Sequential training can also use multi-step Double Q-style targets and conservative penalties for unsupported actions, drawing on established methods for overestimation control and offline policy learning [11, 12]. The deterministic posterior remains available when no validated model is installed.

---

9. Structured Transition Learning

Policy learning asks which action has worked. Transition learning asks what follows from a state and action. Keeping those questions separate makes model-based reasoning possible without forcing every Domain into reinforcement learning.

A transition declaration specifies input paths, an optional action path, grouping paths, typed prediction targets, minimum support, neighbor count, and maximum distance. Each eligible event becomes a transition sample.

At prediction time, Adapt-1 restricts candidates to the declared group and computes a mixed-type distance:

The maximum term prevents one severely mismatched feature from vanishing inside an average. Numeric targets use inverse-distance weighting. Categorical and Boolean targets use weighted votes.

After support and distance gates pass, confidence is:

d_bar is mean neighbor distance, a target agreement, and s the support factor.

Transition predictions are evaluated prequentially. A prediction is retained until a later labeled event reveals the target; only then can it update calibration. This separates local neighbor agreement from observed predictive reliability.

Composition remains explicit

A transition result can inform a later policy decision or hypothesis query, but Core does not silently feed its own output back as evidence. The application can store a versioned transition result as a derived observation with provenance, then include selected values in a later context.

The extra step prevents a hidden loop in which a model's prediction reinforces the policy that produced it. It also makes composition replayable: the next decision identifies the predicted value, the transition model version, and the evidence used to produce it.

---

Adapt-1's reasoning path is therefore a sequence of typed transformations. Episodic memory supplies evidence. Associative and relational state expands the cue. Domain conditions establish applicability. Policy state summarizes measured utility. Transition state estimates consequences. Uncertainty gates decide whether the available support warrants a selection.

Learning travels in the opposite direction. A later observation or outcome returns through a sealed decision and updates only the mechanisms to which it is explicitly attributable.

The mechanism sequence now continues from declared policy and transition surfaces into the structures that can be induced, retained, revised, and measured over longer operation.

---

10. Autonomous Structural Memory

Declared hypotheses work when candidate explanations are known in advance. Autonomous structural memory addresses the complementary case: the observations and targets are typed, but the useful rule has not been declared.

10.1 Predicate induction

The learner converts configured event fields into candidate predicates. Supported forms include equality, inequality, numeric ranges, category membership, presence, and absence. It then searches bounded conjunctions with beam search.

A candidate rule receives predictive gain minus a complexity penalty:

For categorical targets, Delta I is information gain. For numeric targets, it is variance reduction. |r| counts literals and lambda discourages rules that explain the stream only by becoming unnecessarily specific.

The bounded beam is important. Exhaustive conjunction search grows combinatorially with the number of available predicates. Adapt-1 retains a small frontier of promising candidates and requires minimum support before a rule can become active. Its information-gain objective belongs to the established family of decision-rule induction methods [13].

10.2 Temporal validation and conservative support

Candidate rules are fit on an earlier temporal partition and calibrated on later observations. Binary support uses Beta evidence and a Wilson-style lower confidence bound:

Ranking by a lower bound protects the learner from tiny, perfect-looking samples. A rule needs both apparent quality and enough evidence to survive uncertainty. The interval is derived from Wilson's score construction for binomial proportions [14].

Active rules remain revisable. New evidence can specialize, generalize, replace, or retire a rule. The learner stores parent identifiers, model versions, revision reasons, and a bounded archive of retired structures. A response can therefore identify both the current rule and the lineage through which it emerged.

10.3 Drift and recurring latent states

A lightweight prequential residual monitor runs per target. Sustained residual change opens a suspected regime window before the next full induction pass. Post-change evidence enters a bounded buffer, incomplete feature vectors are excluded from change decisions, and a cooldown prevents one unstable period from producing many adjacent regimes.

Rules that fire together form a content-addressed latent state. Its identifier depends on active predicates and target structure rather than insertion order. When a previously observed configuration returns, the same structural signature can be recognized and its transition history reused.

This is a discrete belief-state approximation. It turns recurrent configurations into stable symbols and learns transitions between them. The representation remains grounded in fired rules and supporting events, so a latent state can be inspected rather than existing only as an unnamed vector.

10.4 Transfer through sufficient statistics

Compatible Domains owned by the same session can exchange bounded sufficient statistics. Raw records are not copied. A transferred rule enters the receiving Domain as a provisional structural prior and requires local confirmation before becoming authoritative.

This separates reusable structure from local evidence. It permits one stream to suggest what another should test while retaining tenant boundaries and target-specific validation.

---

11. Recurrent Sets and Adaptive Intervals

Some streams reveal sets rather than a single label: active devices, occupied bands, recurring incident windows, spatial regions, or objects intermittently visible through occlusion. A scan may omit a persistent member without proving that the member disappeared.

Adapt-1 uses two related mechanisms for this setting.

11.1 Recurrent set memory

The set learner matches predicted and observed members one-to-one. This prevents one broad prediction from receiving credit for several nearby targets. For a track with s confirmed appearances and f failures, recurrence belief begins from a uniform prior:

Missing members are handled through delayed negatives. One absent scan does not immediately erase a track. Repeated misses eventually accumulate counterevidence, lower the persistence bound, and suppress stale predictions.

Track competition prevents several candidates from claiming the same observation. Recent observed cardinality constrains output size. These mechanisms turn recurrence into a precision problem as well as a recall problem: reconstructing every plausible member is not useful if the predicted set grows without control.

11.2 Adaptive interval state

The interval learner consumes bounded numeric observations such as lower and upper positions, extent, strength, and optional scalar signals. It maintains persistent tracks, matches nearby observations, estimates robust extents, learns repeated spacing, and can infer latent gaps when independently recurring structure supports them.

Observed track parameters use a bounded running rate:

Early observations move an estimate quickly, while mature tracks remain stable. Repeated sightings raise belief approximately as:

Additional persistence and reporting gates still apply. Period candidates are inferred from repeated position differences; dense patterns are penalized; secondary periodic structures need independent support; and inferred gaps remain labeled as inferred rather than observed.

The interval primitive is domain-independent in vocabulary and intentionally specific in geometry. Relationship structure belongs to hypergraph and autonomous structural memory. Sequential values belong to transition learning. The Domain selects the primitive whose representation matches the observable process.

---

12. How Learning Changes the System

Adapt-1 distinguishes storing, associating, estimating, and training. They are different updates with different evidence requirements.

12.1 Route-specific updates

Let the composite state remain S = (E, A, G, R, P, T, L). An accepted item changes only eligible components:

Each indicator records an eligibility decision. The major routes are:

A successful query is not an outcome. Retrieval rank, symbolic path count, graph density, and model confidence are diagnostic signals. They do not tell the system whether an external task succeeded.

12.2 Credit follows provenance

Execution feedback can identify a decision, target memory, relation-policy pair, stage-specific evidence, or several of these scopes together.

• A decision identifier restores the context and alternatives visible when an action was chosen.
• A memory identifier attributes measured value to a specific retrieved record.
• Stage evidence distributes stage rewards to the records that supported each stage of a plan.
• A relation-policy pair updates aggregate policy evidence when a complete decision snapshot is unavailable.

For measured reward r_d, only the attributed subset receives memory-value evidence:

The same principle applies to fusion channels. Only channels with positive strength in the sealed trace of an attributed memory receive outcome evidence. An active subsystem is not globally reinforced merely because it appeared in an explanation.

Unknown outcomes remain stored as events but do not cross the reward gate. This preserves audit history without training value-bearing state on administrative or semantically unmapped labels.

12.3 Fast and slow learning timescales

Most online adaptation is incremental:

A policy outcome updates weighted posterior evidence. A transition event joins a bounded neighborhood. A structural event updates residuals and rule support. A recurrent observation changes track evidence.

Trainable models follow a slower lifecycle:

Successful stores mark only the owning session as dirty. Sample count, target diversity, version interval, cooldown, and concurrency determine whether training is justified. A worker snapshots that session's data, trains outside the request lock, evaluates on a temporal validation slice, and installs the candidate atomically only if the acceptance criterion passes.

This allows reads and writes to continue during training. A failed candidate leaves the previous inference state intact. If newer evidence arrives during training, version checks prevent a stale model from silently replacing current state.

PyTorch can train small evidence rerankers, link predictors, and nonlinear policy models. Tree ensembles are also available for policy prediction. Deterministic numerical fallbacks remain part of query-time Core. Hardware acceleration changes throughput, while the memory and Domain contracts remain stable.

12.4 Determinism and explicit exploration

For a fixed event order, seed, backend, and exploitative selection mode, most online paths are deterministic. Learning does not require randomness: a deterministic algorithm can update its state from new evidence and produce a different future answer.

Stochasticity enters only through declared mechanisms such as Thompson sampling or model-training initialization. It is not inferred from the mere presence of machine learning.

12.5 Observable learner state

Domain responses report learner state by subsystem. Policy, transition, and structure components independently report whether they are disabled, accumulating evidence, or ready, along with sample counts and model versions.

This avoids a misleading global trained flag. One response may contain an active policy posterior, an abstaining transition learner, and a structural learner waiting for minimum support. Their states describe different evidence contracts and mature at different rates.

12.6 Outcome-trained retrieval adaptation

Adapt-1 can adapt query-plan parameters and optionally rerank retrieved evidence. Query-plan units contain bounded traversal parameters for depth, exploration, backtracking, hyperedge bias, and neural-symbolic balance. A unit changes only after measured feedback identifies the affected evidence. Retrieval diagnostics cannot award it fitness.

The optional neural evidence reranker operates after deterministic candidate generation. It trains only when both positive and negative examples exist and retains the base scoring path whenever accepted weights are unavailable. The learned model can therefore reorder evidence without becoming the sole path by which evidence is found.

---

13. External Perception as Typed Evidence

For visual tasks, an application-side adapter turns an image and a spatial question into grounded objects, masks, geometric relations, or pointing candidates for Core.

The adapter may call an external open-vocabulary detector compatible with OWLv2-style grounding [15], an external segmentation model compatible with Segment Anything [16], or other 2D and 3D vision capabilities. Classical saliency, connected components, pose geometry, free-space geometry, and relation-specific operations can likewise remain in the surrounding toolchain. Their outputs enter Core as typed, provenance-bearing evidence. Core performs the task-level composition and retains ownership of decision state, outcome routing, and learning.

The external perception worker remains separate from Core's general memory traffic. A deployment can route vision requests to GPU-backed workers while keeping the primary state service on CPU infrastructure. Adapter inputs are downscaled to bounded model dimensions so attention maps, segmentation tensors, and intermediate feature maps have predictable memory requirements outside Core.

Adapter output is not written automatically. The application decides whether a grounded object, relation, or result becomes episodic evidence or a Domain event. This boundary prevents every external capability call from silently changing persistent memory.

---

14. Direct and Model-Augmented Execution

Adapt-1 exposes the same adaptive state through two execution patterns.

In direct execution, an application requests retrieval, explanation, hypothesis support, transition prediction, interval state, or policy selection. Core computes the result from stored evidence and installed learners.

In model-augmented execution, Adapt-1 retrieves evidence and supplies it to an external OpenAI-compatible completion transport. The language model synthesizes natural language, and the exchange may be stored as another episode.

External capability use is orthogonal to these two execution patterns. Direct execution may reason over typed evidence from an external image model, classical 3D-vision pose solver, or other adapter; those tools do not become part of Core or take ownership of its learning state.

Generated content remains distinguishable from measured observation. If it is stored, source model and version metadata preserve that boundary so later learning can apply the appropriate evidence contract.

---

15. Persistence, Isolation, and Operational Semantics

Adaptation matters only when state survives, remains attributable, and can be removed correctly.

15.1 Durable evidence and replaceable models

PostgreSQL persistence stores primary records, session-derived state, Domain definitions, and learner snapshots. A buffered writer reduces synchronous database overhead. Startup import restores state and validates the embedding backend.

Primary evidence and derived models have different roles. Evidence is the rebuild source. Derived snapshots are versioned acceleration and can be reconstructed from surviving records when necessary.

Each record carries a canonical SHA-256 content digest and a lineage digest over source kind, content, parent lineage, and parent memory identifiers. The integrity label is tamper_evident_not_authenticated: the digest can reveal unexpected modification, while source authentication remains a deployment responsibility.

15.2 Session ownership

Records, vector candidates, graphs, policy samples, transitions, latent states, and Domain definitions are keyed by session. A production application layer derives that identity from authentication and removes caller-controlled authorization metadata before Core receives the operation.

Ownership applies to definitions and learned state together. Isolation would be incomplete if records were private while policy buckets or Domain names remained global.

15.3 Clearing the complete state

Deleting a memory removes its primary record and dependent vector, concept, graph, reconsolidation, dual-store, and Hebbian entries. Clearing memory performs the corresponding operation across the owner's Core records. Domain evidence, policy scores, adaptive learner state, and the Domain definition are separately addressable so an application can reset the intended layer. Other sessions remain untouched.

This makes deletion a graph operation rather than a row operation. A zero record count would not represent a clean state if embeddings, edges, or policy samples still contained the removed evidence.

15.4 Adaptive and frozen reads

Adaptive reads may update bounded access counts, consolidation, lability, and Hebbian co-retrieval. Frozen reads disable those mutations. Public traces identify which mode was used and report embedding, vector-index, and optional-component status.

Access accounting uses totals, the most recent timestamp, and a bounded recent-time window. Large derived snapshots can use versioned lossless compression. These controls prevent metadata growth even when the primary record count remains stable.

15.5 One Core, several transports

HTTP, Python, compatibility, and Model Context Protocol clients invoke the same stateful Core. Transport adapters may reshape requests and responses, but they do not create separate learners. The scientific semantics belong to the operation and session state rather than the interface used to reach them.

---

16. Controlled Evaluation

Retrieval accuracy alone cannot characterize an adaptive memory system. The retained evaluation measures discrimination, calibration, selective prediction, adaptation delay, recurrence, state growth, latency, and trace completeness.

The protocol is prequential: each system predicts first, then observes the event and updates. No event is used to train the prediction scored at that same step.

16.1 Protocol

The predictive-structure stream uses:

• 10 random seeds;
• 900 events per seed;
• three 300-event phases: concept A, concept B, then A recurring;
• 5% label noise;
• 10% missingness in predictive features;
• scoring before every update.

The compared systems are Adapt-1 without induction, Adapt-1 with autonomous structural memory, a Hoeffding Adaptive Tree, and an Adaptive Random Forest [17].

The recurrent-set stream uses 450 events per seed under the same A -> B -> A pattern. It compares an interval-only Adapt-1 ablation, composed Adapt-1 recurrent-set memory, and DBSTREAM [18].

The generator and Domain declarations belong to the evaluator. Core contains no task names, benchmark vocabulary, or task-specific solver. This isolates the mechanisms being tested from the stream that tests them.

16.2 Predictive discrimination

Autonomous structural memory therefore produced a 26.0-point balanced-accuracy gain over Adapt-1's no-induction ablation. It exceeded the included Adaptive Random Forest by 3.5 points under the same stream.

Balanced accuracy averages recall across classes. Cohen's kappa measures improvement over chance agreement. Together they distinguish useful discrimination from a superficially strong majority-class result.

16.3 Calibration and proper losses

The structural model's Brier decomposition produced reliability 0.0089, resolution 0.1322, and uncertainty 0.2499. Resolution captures the model's ability to separate outcomes. The prequential ledger also reports area under the risk-coverage curve, 0.0806.

These probability metrics apply to outputs with later labeled outcomes. Retrieval scores and other mechanism-specific support values remain ranking quantities rather than being relabeled as probabilities.

16.4 Selective prediction

Adapt-1 can abstain when structural support is insufficient. Coverage and selective accuracy are:

The two systems occupy different operating points. Adapt-1 concentrates predictions where its learned rules have support; the forest almost always answers. Selective classification studies precisely this accuracy-coverage trade-off [19]. The retained risk-coverage sequence and AURC permit an application to choose an operating point from measured risk.

16.5 Drift and recurrence

Adapt-1's recurrence-retention ratio was 0.888, compared with 0.827 for the forest. The Hoeffding tree's high nominal ratio accompanies near-chance performance, illustrating why retention should be read beside absolute predictive quality.

Across seeds, Adapt-1's standard deviations were approximately 63 samples for drift delay and 80 for recurrence delay. The variability is retained in the chart rather than hidden behind the means.

16.6 State size and latency

Adapt-1's p50 latency was approximately 95 microseconds and p99 approximately 43.6 ms. The distribution identifies occasional induction work rather than uniformly slow operation.

Second-half state growth was 1.33x for structural Adapt-1, 1.28x for the forest, and 1.01x for the tree. These measurements connect predictive performance to the retention and bounded-state controls described earlier.

16.7 Recurrent-set quality

Adapt-1 predicted 3.00 members per event against a target cardinality of 3.00, with 0.17 false and 0.18 missed members on average. DBSTREAM predicted 6.66 members with 3.66 false members. Delayed negatives, one-to-one matching, track competition, persistence bounds, and cardinality controls account for the precision-recall balance.

Localization mean absolute error was 0.112, compared with 0.105 for the interval-only ablation and 0.336 for DBSTREAM. The composed learner improves membership quality while retaining competitive localization.

16.8 Evaluation profile

Mean structural trace completeness was 0.90 across seeds. Emitted rules retain conditions, model version, revision state, and supporting memory identifiers. Trace completeness measures whether the evaluator can reconstruct the evidence and rule behind a result. It remains separate from predictive quality, giving the architecture two independently measurable properties: what it predicted and whether the producing state is traceable.

The intrinsic study remains the controlled numerical center of this article because it retains complete per-seed outputs, internal ablations, and comparable stream baselines under one protocol. The external tasks serve a complementary purpose: they test whether the same Core and Domain abstractions can consume independently defined observations, actions, and outcomes.

---

17. External Evaluations

The external evaluations probe online state use under partial observability, frozen spatial composition, and grounded action across retained feedback. Each result is reported with the task exposure, mutation state, comparator training, and evaluated denominator that produced it.

A valid adapter maps source data and external capability outputs into declared evidence and returns measured outcomes. It may invoke general-purpose models or deterministic tools, but it does not contain a benchmark-specific answer oracle or replace Core's task-level reasoning. Runs begin from clean session state, identify the backend and hardware, preserve benchmark repositories, report trivial policies, and retain enough trace data to audit the decision path.

17.1 Evaluation scope

POPGym HigherLowerHard asks whether persistent state can improve decisions when the full deck state is hidden and the useful action changes with the cards already observed. It is a compact test of rapid online learning under partial observability.

RoboSpatial-Home asks whether Core can compose typed outputs from perception, depth, masks, reference frames, and geometry into spatial answers on first exposure. Mutation is disabled throughout, so the result measures frozen perception-to-geometry composition rather than online adaptation.

BOP-Ask-core asks what Core can produce from a clean first encounter, then how grounded action outputs change as measured feedback is retained across complete passes while visual detections remain fixed. Pass 1 and pass 3 are reported separately so initial behavior and accumulated adaptation remain visible.

Across the evaluations, the same Core operates without offline task training, but the interaction available at the result differs. One run has no online updates, one contains 1,550 environment timesteps, and one reaches 2,064 decisions across retained passes. These quantities remain in their native units because questions, decisions, and environment timesteps are not interchangeable.

17.2 Task exposure and comparison conditions

The Adapt-1 values in this section are read from the retained run artifacts for these evaluations. Benchmark definitions, published baselines, and competitor training descriptions come from the official papers and repositories. Replication guides for POPGym HigherLowerHard, RoboSpatial-Home, and BOP-Ask-core will be published at github.com/0xReisearch.

Model pretraining, task fine-tuning, online interaction, episodes, questions, and environment timesteps remain separate quantities. Ratios are used only where the count units are compatible. Architecture and model-size disclosures remain in the detailed cases below.

17.3 POPGym HigherLowerHard: learning under partial observability

HigherLowerHard asks an agent to predict whether the next card will be higher or lower while the complete deck state remains hidden. The current card is not enough. Useful behavior depends on retaining the ranks already observed and changing the decision rule as the remaining-card distribution changes. The benchmark therefore tests whether persistent state improves action selection inside a partially observed stream [20].

Adapt-1 began from a clean state after a verified global state clear and completed ten episodes while retaining its learned state between them. HigherLowerHard uses three 52-card decks, deals one card at reset, and requires 155 decisions to finish. The retained run therefore contains 1,550 task timesteps [20, 29]. All 1,550 environment actions, score events, and confirmed feedback operations match in the retained ledger. Episode 1 began at learner version 0; the last confirmed feedback advanced the same Domain to learner version 1,550.

POPGym's native return sums +1/156 for a correct guess, -1/156 for an incorrect guess, and 0 for a tie. Across all 1,550 decisions, the ten-episode return was 4.961538, or 0.496154 per episode. Episodes 2 through 10 averaged 0.507835, and the last five averaged 0.543590 [29].

The frozen uniform policy evaluated on the same ten seeds averaged 0.020513. Adapt-1's observed mean was 0.475641 higher. This same-seed control is the direct controlled comparison in the retained artifact; the published PPO values below remain a separate numerical reference with a different estimator and protocol.

The action ledger shows the policy that produced those returns. The Domain exposed the current public rank, the ordered rank vocabulary, the two legal actions, and measured outcomes. It did not supply a midpoint, cutoff, preferred action, next card, remaining-card counts, probabilities, hidden state, future outcome, or solver output.

Without a supplied midpoint, Adapt-1 selected higher on more than 99% of visits at ranks 0 through 4 and lower on more than 96% of visits at ranks 8 through 12. The transition region emerged around ranks 5 through 7. Rank 6 remained higher-biased in this retained run, while ranks 5 and 7 carried the mixed behavior around the crossover.

The published comparison is a memory-architecture study, not an LLM comparison. Every row below used the same PPO task-training protocol and outer policy network. The inner memory model changed.

The paper fixes recurrent state to 256 dimensions, projects observations and memory outputs through width-128 layers, and uses two-layer width-128 actor and critic heads. Exact total parameter counts vary by memory architecture and are plotted rather than tabulated in the paper. The shared training uses PPO from scratch on HigherLowerHard: 15M task timesteps for each independently trained policy, repeated across three trials. The table above is the complete published HigherLowerHard field. The reported statistic is MMER, the maximum over training epochs of each epoch's mean episodic reward, summarized across trials [20].

The direct task-exposure contrast is 15,000,000 versus 1,550 timesteps, or approximately 9,677× more HigherLowerHard interaction for each published baseline policy. The three reporting trials consume 45M steps in total, but 15M is the correct per-policy comparison. There is no fixed held-out corpus: the environment continually reshuffles the deck, and the published score is selected from training performance.

The retained Adapt-1 artifact is one descriptive ten-episode sequence, not a variance-matched superiority estimate. POPGym's baseline statistic is maximum-over-training MMER across three PPO trials. The observed result is direct: after entering cold at 0.391026, the same Core averaged 0.507835 over episodes 2 through 10 and 0.543590 over the last five after 1,550 total task timesteps, compared with 15M timesteps for each trained baseline policy and trial. The rank-conditioned ledger makes the learned action structure visible rather than reducing the run to one aggregate score.

17.4 RoboSpatial-Home: frozen spatial composition

RoboSpatial-Home separates spatial reasoning into configuration, compatibility, and context. Configuration asks for relative object relationships. Compatibility asks whether an object can fit in a specified region. Context asks for points in usable free space under a stated reference frame [21, 22].

The broad goal here is different from the online runs. A stateless Neuroadapt vision route processed all 350 examples once. It used OWLv2-base as a fixed open-vocabulary detector, the supplied depth image, ten classical scene masks, explicit relation parsing, and deterministic spatial geometry. No Domain, memory store, retained experience, or task-time weight update was used. This is therefore a frozen perception-to-geometry result, not a Core-learning result. The complete 350-row trace records successful detector and segmenter status for every request.

The public repository defines its Total as the unweighted average of configuration, compatibility, and context. Its published-works table currently tops out at 73.9% for Qwen3-VL 235B-A22B Thinking, with SpaceTools-3B at 70.4% and RoboBrain2.5-8B at 73.0% [22]. Adapt-1's frozen 74.65% three-category mean is therefore the like-for-like repository comparison. A later CVPR 2026 challenge report uses a task-distribution-weighted total and reaches 80.9% with RoboSpatialBrain. The matched Adapt-1 value for that comparison is its 74.29% weighted aggregate, 6.61 points lower [24]. RoboSpatialBrain combines RoboBrain2.5-8B-NV with a separate Qwen3.5-2B helper, forced reasoning, task-specific prompts, and explicit reference-frame redirection.

The score table contains very different exposure histories:

These counts provide data-exposure context, not compute normalization. Tokens, question-answer pairs, trajectories, mixed post-training samples, and separate tool parameters are different units. Qwen3-VL, RoboBrain2.5, and the SpaceTools stack contain substantial spatial or embodied training even when no RoboSpatial-specific fine-tune is named.

Adapt-1 received no RoboSpatial training and no online updates. The stateless vision route combined generic external detection with supplied depth, classical masks, and explicit spatial computation. On the repository's macro aggregation it exceeded the published-works leader; on the later challenge's weighted aggregation it remained below RoboSpatialBrain.

With no Domain or state mutation, this is not an online-learning result and does not evidence Core memory or adaptation. It is a cold, stateless composition result from perception, depth, masks, geometry, and output formatting.

17.5 BOP-Ask-core: grounded action from zero-shot entry to retained adaptation

BOP-Ask-core contains 688 held-out questions spanning 3D pose, grasp prediction, trajectory generation, spatial relations, relative depth, and object rearrangement. Its companion training set contains approximately 150,000 images and 33.8 million question-answer pairs. The paper fine-tunes Qwen2.5-VL and NVILA on that corpus using their official training codebases. The fine-tuning runs use eight NVIDIA A100 GPUs with default hyperparameters [25].

Core itself received no BOP-Ask offline training or fine-tuning. The run used OWLv2-base as a separate fixed external visual front end across all passes and cached its detections. Core remained the task-level learning and reasoning substrate, composing the available evidence, maintaining decision state, routing measured outcomes, and adapting future behavior. Three complete passes over the 688-item set produced 2,064 benchmark decisions. The retained trace contains all 2,064 rows and matches the SHA-256 recorded in the run manifest. The paper's closed API baseline does not disclose parameter count or underlying training exposure. The most explicit task-training comparison is therefore against the BOP-Ask-supervised open models.

Pose is scored as the benchmark's true volumetric 3D cuboid IoU. After Core emits eight image points, a calibrated PnP reconstruction and convex-polyhedron evaluator recover the predicted cuboid and measure its volumetric overlap with the reference pose. Camera matrices, object dimensions, and ground-truth pose metadata are confined to this post-prediction scorer; they are never exposed to Core or the prediction path. The run manifest accordingly records no external task solver.

Pass 1 and pass 3 use different exposure states:

• Pass 1 begins from clean state and is the system's first interaction with the benchmark tasks. It is zero-shot at entry. Because feedback is online, later items in pass 1 may use evidence admitted earlier in that same traversal. The pass is therefore online rather than frozen or static.
• Pass 3 begins after two complete prior traversals and continues learning within the third. Its scores measure behavior after accumulated benchmark interaction, not initial capability.

First-pass metric profile

The BOP-Ask paper reports Gemini Robotics-ER 1.5 on all seven metrics compatible with Adapt-1 pass 1. The table below reproduces that API reference. It is separate from the full published model field used for pass 3.

Against the Gemini Robotics-ER 1.5 API row, Adapt-1 pass 1 records four wins and three losses: it is ahead on trajectory success, trajectory error, grasp NCE, and spatial reasoning, while trailing on pose, relative depth, and rearrangement. Before two complete passes of accumulated feedback, the cold-entry system composed detections and geometry into paths, cuboids, grasp locations, and spatial relations.

Third-pass published-field comparison

The fully traced third pass reports all seven metric-compatible benchmark outputs. Adapt-1 is above the published model field on four and below it on three:

Figure 24 plots pass 1 and pass 3 against the strongest published value in every compatible metric column. Pass 1 is below the column-wise field best on all seven metrics. Pass 3 records four wins and three losses against those column-wise bests: it is above them on trajectory success, trajectory error, true 3D pose IoU, and rearrangement recall, while remaining below them on grasp, spatial reasoning, and relative depth. Because detections were fixed and cached across passes, the change occurred downstream of detection.

NVILA-15B SFT supplies the strongest published model values for trajectory, grasp, spatial, depth, and rearrangement, with Molmo-72B tying its displayed 1.40 grasp NCE. NVILA-2B SFT supplies the strongest published pose value at 77.4%. The supervised NVILA models were trained on the 33.8-million-pair task corpus [25]. Human grasp NCE is lower still. By pass 3, Adapt-1 exceeded every reported model baseline on four of seven compatible metrics despite receiving no offline BOP-Ask training.

The pose result is the benchmark-compatible volumetric 3D IoU, not the older run's projected 2D-hull proxy. The archive rescored that older pass-3 pose output at 0.00% true 3D IoU; it is not used here. The values above all come from the single retained full-3d-v1 run whose complete trace ships with the evidence package.

The benchmark reports the resulting task-level behavior rather than a component ablation. Fixed cached detections keep the visual front end constant, so the measured progression belongs to the decision and adaptation stack downstream of detection. Pass 1 measures behavior from zero-shot entry. Pass 3 measures behavior after persistent evidence and outcome routing across prior passes.

Training and exposure

Using the paper's corpus counts, 33.8M / 688 is approximately 49,128 task-training question-answer pairs per held-out Core question. This is a corpus-size ratio. The paper does not disclose epochs or cumulative sample consumption. The current public dataset card reports a different release snapshot: 32,680,582 training records and 172,052 unique RGBs, while the comparison above uses the paper's counts for the paper's model table [25, 28].

Pass 1 to pass 3

Across the three passes, Core admitted grounded feedback online, retained the resulting state, and changed later outputs while perception remained fixed. The sequence reports both initial capability and data-efficient adaptation: no BOP-Ask offline training, 688 online decisions for the first-pass result, and 2,064 decisions for the third-pass result, compared with a 33.8-million-pair supervised task corpus for the strongest open action reference.

---

18. Scientific Position

Adapt-1 draws on mature research in associative memory, continual learning, hypergraphs, neurosymbolic reasoning, contextual bandits, transition models, rule induction, selective prediction, and visual grounding.

Its contribution is the integration of these mechanisms around four commitments:

1. Adaptive state persists outside an inference call. Episodes, relations, rules, policy evidence, transition samples, and latent state survive as versioned objects.
2. Mechanisms remain typed. Retrieval support, hypothesis applicability, transition confidence, and expected reward retain separate semantics.
3. Learning is evidence-gated. Support, distance, validation, out-of-distribution, attribution, and abstention gates determine when state may change or a prediction may be emitted.
4. Provenance is part of the result. Learned outputs retain links to evidence, learner versions, decision snapshots, and revision history.

This combination creates a middle layer between storage and a general model. It is more adaptive than a vector index, more structured than generic retrieval, and narrower than an end-to-end foundation model.

---

Try Adapt-1 for free: https://app.reilabs.org/adapt-1

Guide to replicating the evaluations: https://github.com/0xReisearch/adapt1-behavioral-replication

---

References

1. McClelland, J. L., McNaughton, B. L. & O'Reilly, R. C. "Why There Are Complementary Learning Systems in the Hippocampus and Neocortex: Insights From the Successes and Failures of Connectionist Models of Learning and Memory." Psychological Review 102, 419-457 (1995). Author-hosted paper
2. Kumaran, D., Hassabis, D. & McClelland, J. L. "What Learning Systems Do Intelligent Agents Need? Complementary Learning Systems Theory Updated." Trends in Cognitive Sciences 20, 512-534 (2016). PubMed Central
3. Tronson, N. C. & Taylor, J. R. "Molecular Mechanisms of Memory Reconsolidation." Nature Reviews Neuroscience 8, 262-275 (2007). Nature
4. Nader, K. & Hardt, O. "A Single Standard for Memory: The Case for Reconsolidation." Nature Reviews Neuroscience 10, 224-234 (2009). Nature
5. Zhou, D., Huang, J. & Scholkopf, B. "Learning with Hypergraphs: Clustering, Classification, and Embedding." NeurIPS 19 (2006). NeurIPS
6. Yadati, N. et al. "HyperGCN: A New Method for Training Graph Convolutional Networks on Hypergraphs." NeurIPS 32 (2019). NeurIPS
7. De Raedt, L., Dumancic, S., Manhaeve, R. & Marra, G. "From Statistical Relational to Neuro-Symbolic Artificial Intelligence." (2020). arXiv
8. May, B. C., Korda, N., Lee, A. & Leslie, D. S. "Optimistic Bayesian Sampling in Contextual-Bandit Problems." JMLR 13, 2069-2106 (2012). JMLR
9. Altamirano, M., Briol, F.-X. & Knoblauch, J. "Robust and Scalable Bayesian Online Changepoint Detection." ICML (2023). PMLR
10. Alami, R., Maillard, O.-A. & Feraud, R. "Restarted Bayesian Online Change-point Detector Achieves Optimal Detection Delay." ICML (2020). PMLR
11. van Hasselt, H. "Double Q-learning." NeurIPS 23 (2010). NeurIPS
12. Kumar, A., Zhou, A., Tucker, G. & Levine, S. "Conservative Q-Learning for Offline Reinforcement Learning." NeurIPS 33 (2020). NeurIPS
13. Quinlan, J. R. "Induction of Decision Trees." Machine Learning 1, 81-106 (1986). DOI
14. Wilson, E. B. "Probable Inference, the Law of Succession, and Statistical Inference." Journal of the American Statistical Association 22, 209-212 (1927). DOI
15. Minderer, M., Gritsenko, A. & Houlsby, N. "Scaling Open-Vocabulary Object Detection." (2023). arXiv
16. Kirillov, A. et al. "Segment Anything." (2023). arXiv
17. Gomes, H. M. et al. "Adaptive Random Forests for Evolving Data Stream Classification." Machine Learning 106, 1469-1495 (2017). Author manuscript
18. Hahsler, M. & Bolanos, M. "Clustering Data Streams Based on Shared Density Between Micro-Clusters." IEEE Transactions on Knowledge and Data Engineering 28, 1449-1461 (2016). DOI
19. Gangrade, A., Kag, A. & Saligrama, V. "Selective Classification via One-Sided Prediction." AISTATS (2021). PMLR
20. Morad, S. et al. "POPGym: Benchmarking Partially Observable Reinforcement Learning." ICLR (2023). arXiv
21. Song, C. H. et al. "RoboSpatial: Teaching Spatial Understanding to 2D and 3D Vision-Language Models for Robotics." CVPR (2025). arXiv
22. Song, C. H. et al. "RoboSpatial-Eval: Evaluation Script and Published-Works Leaderboard for RoboSpatial-Home." GitHub
23. Chen, B. et al. "SpaceTools: Tool-Augmented Spatial Reasoning via Double Interactive RL." (2025). arXiv
24. Xie, Y. et al. "Technical Report of RoboSpatial Challenge at CVPR 2026: Selective Reasoning Activation and Reference-Frame Disambiguation for Embodied Spatial Reasoning." (2026). arXiv
25. Bhat, V. et al. "BOP-ASK: Object-Interaction Reasoning for Vision-Language Models." (2025). arXiv
26. Bai, S. et al. "Qwen3-VL Technical Report." (2025), with the Qwen3 backbone report. Qwen3-VL · Qwen3
27. Ji, Q. et al. "RoboBrain 2.5 Technical Report." (2026). arXiv
28. BOP-Ask official release cards. Training release · Test release
29. Morad, S. et al. "POPGym HigherLower environment implementation." Official source