Source-linked AI summary
Jev-DLC: Decision-Level Continual Learning from Typed Feedback
Yangyang Feng, Jianle Gan, Zhuoyan Feng, Chengwei Qin
TL;DR
Personal preferences often become visible only after deployment, while history-based personalization retains interactions or adapts user-specific representations. Jev-DLC carries typed corrections as bounded per-user action evidence and directed repair above a frozen scorer; on TCPA-Repeat it improves user-macro accuracy over Recent-4 history replay.
Problem
Method
Jev-DLC preserves recurring-action evidence and preferred–rejected correction direction in complementary bounded channels above a frozen scorer.
Source: §1 · p. 2Results
44.83% user-macro accuracy on TCPA-Repeat’s full stream is 2.43 points above Recent-4 history replay; after at least two earlier feedback events, accuracy reaches 56.07%, 7.54 points above the strongest comparison at that depth.
Source: §1 · p. 2Takeaways & Limitations
On repeated actions, compact correction state improves later decision accuracy without retaining raw interactions or training the language model at runtime.
Source: §8 · p. 8The formulation uses immediate accept and revoke feedback over a finite candidate set, so each correction maps directly to a decision relation.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Individual preferences often become visible only after deployment, when a user accepts one decision or revokes it in favor of another. Jev-style typed interfaces make this feedback explicit by identifying the selected action and, under revocation, the decision it replaces. We study how these consequences persist across requests as Decision-Level Continual Learning from Typed Feedback (Jev-DLC). Jev-DLC treats each correction as both evidence about a recurring action identity and a directed relation between the preferred and rejected choices. It preserves these signals in complementary bounded channels above a frozen scorer and combines their scores to shape the next decision. The method retains no raw interaction history, adds no scorer call, and performs no runtime language-model backward pass. We also introduce TCPA-Bench, a chronological suite that credits a correction only through its effect on later decisions. On TCPA-Repeat, Jev-DLC reaches 44.83% user-macro accuracy overall, 2.43 points above Recent-4 history replay. After at least two earlier feedback events, it reaches 56.07%, a 7.54-point margin over the strongest reported comparison at this depth. These results show how explicit typed corrections can become compact causal state for continual personalization.
1 Introduction
Jev-DLC turns explicit accept/revoke feedback into bounded per-user state that carries recurring-action evidence and correction direction to later decisions. It evaluates this persistence chronologically, crediting feedback only through its effect on subsequent decisions.
1 Introduction
- Personalization often becomes possible only after deployment, when users accept, replace, or correct a system’s choices.Prior approaches retrieve or replay interactions or adapt user representations or models, preserving richer evidence that later predictions must reinterpret or consult.Source: §1 · p. 1
- Jev-DLC preserves recurring-action evidence and preferred–rejected correction direction in complementary bounded channels above a frozen scorer.Earlier feedback influences the next decision while the scorer interprets its current context and candidates.Source: §1 · p. 2
- The method updates per-user channels and their weights after a committed prediction, without retaining raw interactions, adding a scorer call, or running language-model backward passes.Directional updates make the smallest capped change needed to repair the exposed ordering; revealed channel losses adjust their relative contribution.Source: §1 · p. 2
- 44.83% user-macro accuracy on TCPA-Repeat’s full stream is 2.43 points above Recent-4 history replay; after at least two earlier feedback events, accuracy reaches 56.07%, 7.54 points above the strongest comparison at that depth.The benchmark credits corrections only through their effects on later decisions.Source: Figure 1 · p. 1, §1 · p. 2
- TCPA-BENCH frames typed accept/revoke feedback as a chronological decision problem in which it can affect only later decisions for the same user.The interface identifies the accepted action and, under revocation, the decision it replaces.Source: §1 · pp. 1–2
2 Related Work
Prior personalization work commonly infers preferences from interaction histories or adapts mutable user/model representations. Jev-DLC instead links typed decisions and minimum-change online updates, then measures their effect on later decisions.
2.1 Personalization from User History
- History-based personalization retrieves or replays past interactions or adapts user representations and parameters; Recent-K preserves recent evidence for a frozen scorer to reinterpret at each event.Related settings include LaMP across classification and generation tasks (Salemi et al., 2024) and MIND’s sequential recommendation from news-click histories.Source: §2.1 · p. 2
2.2 Continual and Online Adaptation
- Continual-learning approaches make a model, adapter, or shared memory mutable, whereas Jev-DLC confines mutability to a per-user decision state above a frozen scorer.The paper places the stability–plasticity trade-off at that state level and uses PA-I as its minimum-change update.Source: §2.2 · pp. 2–3
2.3 Typed Decisions and Online Decision Updates
- Jev-DLC connects a typed accept/revoke interface to a minimum-change online update and evaluates whether that state changes later decisions.Prior work separately supplies typed readouts or update methods, while conversation benchmarks and preference arenas assess broader response quality (Zheng et al., 2023; Chiang et al., 2024).Source: §2.3 · p. 3
3 Problem Setup
The setup asks whether typed feedback improves later decisions for the same user while keeping the scorer fixed and the persistent state bounded. Its chronological protocol evaluates each committed prediction before the feedback it exposes can affect future events.
3 Problem Setup
- The problem is whether a user’s typed correction can improve later decisions without changing the scorer or replaying the interaction.The information boundary is established before any particular update rule or evaluation set is specified.Source: §3 · p. 3
3.1 Ordered Typed Interaction
- For each chronological event, a frozen scorer evaluates a finite candidate slate and a fixed mask removes actions that violate hard constraints.The learner commits a prediction from the safe slate before feedback is available.Source: §3.1 · p. 3
3.2 Typed Feedback
- After prediction, feedback either accepts the committed action or revokes it while identifying its replacement, providing decision-level rather than free-text supervision.The revoke specifies the replacement relation directly.Source: §3.2 · p. 3
3.3 Persistent State and Causality
- Per-user state is updated with the current event and revealed feedback, but a prediction may depend only on earlier events’ feedback and cannot be changed by the feedback it elicited.The state is bounded, reset between users, and retains neither raw previous interactions nor candidate text; scorer parameters stay fixed at runtime.Source: §3.3 · p. 3
3.4 Evaluation Objective
- User-macro accuracy is the primary metric because the unit of personalization is the user.The objective tests whether past typed feedback improves later decisions for that same user.Source: §3.4 · p. 3
4 Method
Jev-DLC adds bounded per-user preference state above a frozen typed scorer. It combines recurring-action evidence, directed corrections, and reliability-weighted channel scores to influence later decisions under a compact serving contract.
4 Method
- A frozen typed scorer supplies current-slate judgments while two bounded user-state channels carry recurring-action evidence and correction direction into the next decision.Their combined scores shape decisions without revisiting original interactions.Source: §4 · pp. 3–4
4.1 Frozen Typed Decision Layer
- The scorer assigns candidate logits and hidden representations, while a shared projection is fitted on training users and frozen before development and test replay.Scorer parameters, projection, candidate order, and hard mask remain fixed at runtime.Source: §4.1 · p. 4
4.2 A Typed Correction Provides a Prior and a Direction
- A correction supplies both a preferred-action semantic identity for recurring-action evidence and a directed relation between preferred and rejected choices.The prior records what the user repeatedly chooses; the contrast records how a committed decision should be replaced.Source: §4.2 · p. 4
4.3 Conservative and Responsive Channels
- The conservative channel combines a semantic prior with a local correction vector, while the responsive channel stores correction geometry without the count prior.A channel already respecting the correction remains unchanged; otherwise its minimum-change move repairs the ordering subject to a cap.Source: §4.3 · p. 4
4.4 Combining the Two Channels
- Jev-DLC geometrically combines the channel distributions using per-user reliability weights, which update after the target is revealed and retain a fixed share toward uniformity.The fixed share prevents either channel from becoming permanently unreachable as evidence accumulates.Source: §4.4 · pp. 4–5
4.5 Serving Contract and State Budget
- Each event uses one frozen-scorer call, and feedback from the committed mixture updates channel vectors, semantic counts, and reliability weights for that user’s next event.The state peaks at 324 bytes on TCPA-Repeat and 320 bytes on TCPA-Open; the system retains no interaction history and runs no runtime language-model backward pass.Source: §4.5 · p. 5
5 Theory
The theory explains how typed corrections preserve preference information and how action identity and correction direction can shape later decisions. It combines these signals through persistent channels above a frozen scorer.
5.1 What Typed Feedback Preserves
- Typed action A is sufficient for preference information when the realization distribution factors through A; conditioning on A removes variation among realizations of the same action.The covariance decomposition attributes that removed variation to a positive-semidefinite within-action term.Source: §5.1 · p. 5
5.2 Two Persistence Geometries
- The semantic count score carries repeated categorical tendency, while the fixed-dimensional state also preserves which candidate relation was corrected.The prior tracks recurring action identity; the pairwise direction carries correction geometry.Source: §5.2 · pp. 5–6
- The local minimum-change repair leaves a state unchanged when it already respects the correction; otherwise, it projects toward the ordering boundary, with the cap limiting the move.The correction changes no state direction orthogonal to its pairwise direction.Source: §5.2 · pp. 5–6
5.3 From Local Repair to Persistent Decisions
- Later pairs inherit a repair according to candidate-space alignment, and the two-channel mixture combines recurring-action evidence with transferred correction direction.The channels can update differently because only the conservative margin includes the prior; pre-decision weights determine their contribution.Source: §5.2 · p. 6, §5.3 · p. 6
6 TCPA: Chronological Evaluation of Typed Adaptation
TCPA evaluates typed adaptation through chronological, predict-then-update decisions, crediting feedback only on later events. Its tracks cover repeated actions, changing-item transfer, and controlled temporal diagnostics.
6.1 Predict-Then-Update Protocol
- Each event is committed before feedback is revealed; the resulting state update can affect only subsequent decisions in that user’s stream.Users are disjoint across splits, and the state resets when a new user stream begins.Source: §6 · p. 6, §6.1 · pp. 6–7
6.2 TCPA-Repeat
- TCPA-Repeat preserves chronological user streams and repeated decisions under a stable finite ontology to test whether feedback improves later decisions.The track is derived from LaMP-2.Source: §6.2 · p. 7
6.3 Additional Diagnostic Tracks
- TCPA-Open probes changing-item weak transfer, while TCPA-Dynamics diagnoses stability, shift, recurrence, and noise.The benchmark reports user-macro accuracy as its primary metric and accounts for persistent state, retained history, and extra model computation.Source: §6.3 · p. 7
7 Experiments
The experiments evaluate whether typed state improves later decisions under a chronological feedback protocol, how the two channels contribute, and what serving state and computation require.
7.1 Later-Decision Accuracy
- 44.83% user-macro accuracy on TCPA-Repeat Full is 2.43 points above Recent-4; after at least two earlier feedback events, Jev-DLC reaches 56.07%, 7.54 points above the strongest comparison at that depth.At d ≥4 and d ≥8, Jev-DLC reaches 50.47% and 53.01%, versus 47.18% and 51.88% for MRM-style. Each depth evaluates decisions after at least the specified number of earlier feedback events.Source: Table 1 · p. 7, §7.1 · p. 7
7 Experiments
- Comparisons use the same safe slates, targets, chronological order, and predict-then-update protocol, with shared components fitted on training users and configurations selected on development users.Recent-4 replays four typed interactions; PPlug encodes behavior, TAP-PER queries bounded temporal memory, and MRM adapts a low-dimensional head.Source: §7.1 · p. 7, §7 · p. 7
7.2 Contribution of the Two Channels
- 44.83% on Full is 1.71 points above the conservative channel and 3.33 points above the responsive channel; Jev-DLC has the highest accuracy in every reported column.The conservative channel reaches 54.85% after at least two earlier feedback events, while the responsive channel reaches 50.45% in the deepest stratum, versus 44.30% for the conservative channel; Jev-DLC reaches 56.07% and 53.01%, respectively. Adaptive weighting gains 0.68 points on Full and 4.86 points at d ≥8 relative to development-fixed weighting.Source: §7.2 · p. 8, Table 2 · p. 8
7.3 Serving State and Computation
- Jev-DLC's maximum per-user state is 324 bytes on TCPA-Repeat, independent of stream length, and its accuracy gain over Frozen requires neither raw-history replay nor an additional scorer forward or runtime language-model training.State computation takes 0.1781 ms for prediction and 0.0810 ms for feedback updates after the frozen scorer output is available; these are CPU state-operation measurements.Source: §7.3 · p. 8, Table 3 · p. 8
8 Conclusion
Jev-DLC carries typed correction evidence into later decisions through bounded state above a frozen scorer. On repeated actions, this improves accuracy without raw interaction retention or runtime language-model training.
8 Conclusion
- Jev-DLC preserves recurring-action evidence and the preferred–rejected relation in complementary bounded channels above a frozen scorer.TCPA-BENCH credits feedback only through its effect on later decisions.Source: §8 · p. 8
- On repeated actions, the retained state improves decision accuracy without retaining raw interactions or training the language model at runtime.The paper presents this as a decision-level path to continual personalization in which corrections remain useful after their original requests have passed.Source: §8 · p. 8
- TCPA-BENCH makes correction persistence measurable by counting a correction only through its effect on later decisions.Source: §8 · p. 8
Limitations
The formulation assumes immediate accept-or-revoke feedback over a finite candidate set, and the clearest evidence for the state comes from repeated-action settings. The storage bound also depends on the evaluated semantic vocabularies.
Limitations
- The formulation assumes immediate accept and revoke feedback over a finite candidate set, where each correction maps directly to a decision relation.The stated future directions include recovering relations from delayed or free-form feedback and representing ranked alternatives or partial acceptance.Source: Limitations · p. 9
- The repeated-action setting provides the clearest evidence for Jev-DLC's state; changing-item recommendation favors PPlug overall.The storage bound depends on the evaluated semantic vocabularies, so an expanding action ontology requires a count-storage policy.Source: Limitations · p. 9
Contents of Appendix
- Table 4 collects notation shared by the frozen scorer, persistent state, and proofs.Source: Table 4 · p. 12
A.1 Causal Serving Contract
- The runtime predicts before receiving evaluator-owned typed feedback; feedback may update state for later events but cannot alter the decision that elicited it.State resets at user boundaries.Source: §A.1 · p. 12
A.2 Runtime and Evaluator Boundary
- The implementation separates model-visible events from evaluator-owned targets, allowing a runtime adapter to be audited without access to target files.Source: §A.2 · p. 12
A.3 TCPA-Repeat Construction
- TCPA-Repeat converts LaMP-2 into 15-way movietag decisions while preserving each user's source-record order and separating candidate context from selected-tag targets.Its stable action vocabulary creates repeated identities; feedback depth counts only corrections revealed before the current prediction.Source: §A.3 · p. 12
A.4 Additional Track Construction
- TCPA-Open uses four-item slates from MIND impressions with new items at each later impression, while TCPA-Dynamics tests stable, shifting, recurring, and 20%-label-noise streams.TCPA-Open is intended to test transfer through representation rather than recurring action identity; TCPA-Dynamics has eight users and 512 evaluated events per regime.Source: §A.4 · p. 12
A.5 Validation and Leakage Audits
- Validation rejects duplicate event IDs, out-of-slate targets, runtime/evaluator mismatches, and target-like runtime fields; the runner scores predictions before reading feedback.Recorded hashes distinguish leakage, row mismatch, and accidental result reuse.Source: §A.5 · p. 12
A.6 Metrics and Statistical Inference
The appendices specify a chronological, user-level evaluation protocol and the implementation choices behind Jev-DLC’s frozen scorer and bounded per-user state. They also state conditions and limits for its local updates, typed-feedback sufficiency, and channel-weighting analysis.
A.6 Metrics and Statistical Inference
- User-macro accuracy averages within user before averaging across users; feedback-depth eligibility uses only the number of earlier events available before each prediction.Confidence intervals use 10,000 paired bootstrap draws over users, while event-micro accuracy pools events.Source: §A.6 · pp. 12–13
B.1 Typed Interface and Scorer Implementation
- Jev supplies a typed choice and replacement identifier, while local JEV-SFT provides candidate logits and representations; revealed feedback updates per-user state above the frozen scorer.The scorer remains frozen as feedback changes the state for later decisions.Source: §B.1 · p. 13
B.3 Learned Candidate Projection
- The shared projection is fitted on training users through chronological replay, scoring each event from prior state before updating with its revealed feedback.Development and test users never update the projection.Source: §B.3 · p. 13
B.4 Jev-DLC Configuration
- Jev-DLC initializes bounded correction channels and semantic counts, then updates reliability weights and state only after committing each prediction.The evaluated maximum state is 324 bytes on Repeat and 320 bytes on Open; the count-state bound depends on the evaluated semantic vocabulary.Source: §B.4 · p. 14
B.5 Matched Baseline Contracts
- Matched baselines share safe slates, targets, chronological order, and predict-then-update evaluation, but Recent-4 reruns the scorer with history rather than using identical cached representations.The archived AdaGrad and online-ladder controls have representation provenance that requires reconciliation before calling them matched controls.Source: §B.5 · p. 14
B.6 Timing Protocol
- Jev-DLC’s state computation and feedback update are timed separately from full scorer forwards, so serving comparisons report latency alongside storage, input tokens, scorer calls, and backward passes.On TCPA-Repeat, prediction-side state computation is 0.1781 ms and the feedback update is 0.0810 ms in the current CPU replay.Source: §B.6 · pp. 14–15
C.1 Closed-Form PA-I Solution
- The PA-I repair moves the correction state along the exposed pair direction by a clipped ratio, leaving it unchanged when the direction is zero.With an inactive cap, it is the minimum-distance projection onto the margin half-space; an active cap instead solves the slack-penalized problem.Source: §C.1 · p. 15, §C.3 · p. 15, §C.2 · p. 15
C.5 Action Sufficiency and Realization Nuisance
- Typed action sequences are sufficient for preference under the stated factorization, but not when realization mechanisms depend on preference or surface forms carry distinct information.The channel-weighting bound concerns weighted channel loss relative to a switching path, not deterministic mixed-distribution accuracy or an advantage over another fusion.Source: §C.5 · p. 16, §C.6 · p. 17
D.2 Feedback-Depth Results
The results compare Jev-DLC’s feedback-depth behavior with frequency controls, alternative representations, and fusion configurations. Across these analyses, the reported gains vary by depth and configuration, while some diagnostic comparisons do not establish an overall advantage.
D.2 Feedback-Depth Results
- At depth ≥8, the reported result uses 339 decisions from 15 users; eligible-user counts decrease with depth while paired intervals preserve the same causal unit.The complete depth matrix reports absolute user-macro accuracy on common eligible events.Source: §D.2 · p. 17, §D.4 · p. 17
D.3 Fixed-Cohort Depth Check
- The archived directional PA-I diagnostic uses the same 15 users at every depth threshold, with a within-user slope of +0.71 points per prior feedback and a 95% interval of [−0.34, +2.04].The main depth tables instead use cohorts whose eligibility changes with the threshold.Source: §D.3 · p. 17
D.5 Action-Frequency Variants
- Jev-DLC gains +0.96 points on Full and +6.58 at d ≥8 relative to the recency prior in the local frequency-control replay.The frequency variants retain action counts, whereas Jev-DLC also retains preferred–rejected relations.Source: §D.5 · p. 18
D.6 History Replay and Serving Cost
- Recent-4 replays up to four earlier typed interactions by reconstructing the history block and rerunning JEV-SFT on the enlarged prompt.The empty-history forward is a latency anchor; frozen-cache accuracy remains the common reference for online methods.Source: §D.6 · p. 18
D.7 Chronological Linear Reference
- The chronological linear reference fits a per-user head only on the prefix before each event and is diagnostic rather than a deployable baseline.It repeatedly fits a user-specific model.Source: §D.7 · p. 18
E.1 Directional Representation and Accumulation Controls
- MLP64 PA differs from learned linear PA by +0.55 points on Repeat and +0.08 on Open, with both paired 95% intervals including zero.The two representations have different depth profiles without an established overall advantage.Source: §E.1 · p. 19, §E.2 · p. 19
E.4 Jev-DLC Channel and Fusion Configurations
- 44.83% on Full and 56.07% at d ≥8 are reported for the combined state, which has the highest accuracy among the tested channel configurations at each reported depth.Adaptive weighting improves on the development-fixed combination by 0.68 points on Full and 4.86 points at d ≥8.Source: §E.4 · p. 19
E.3 Future-Predictive Geometry
- The cosine diagnostics characterize alignment between past state and future correction directions; Repeat exhibits alignment-conditioned reuse, unlike Open.The diagnostics are framed as comparisons of past state with future directions, not as a global user direction.Source: §E.3 · p. 19, Table 19 · p. 20
F.1 TCPA-Open
TCPA-Open tests feedback transfer across changing news slates, while TCPA-Dynamics isolates update behavior under controlled preference dynamics. The appendix also specifies the prompt, runtime, artifact, and data-use protocols that bound these evaluations.
F.1 TCPA-Open
- Full-through-depth-four gains are positive, but the depth-eight stratum does not preserve that pattern.Table 19 is cited as the corresponding representation-geometry diagnostic.Source: §F.1 · p. 20
- 29.74% overall Jev-DLC accuracy on TCPA-Open trails PPlug’s 31.30%, with the paired Jev-DLC-minus-PPlug interval at [−2.17, −0.92] points.The reported comparison is for the full stream.Source: Table 23 · p. 20, §F.1 · p. 20
F.2 TCPA-Open Chronological Linear Reference
- The chronological linear reference shows no substantial additional gain in this representation, positioning TCPA-Open as a transfer diagnostic rather than a repeated-action case.The comparison uses each user’s available prefix.Source: §F.2 · p. 20, Table 24 · p. 20
F.3 TCPA-Dynamics; F.4 Full-Method Fusion Diagnostic on Existing Dynamics Caches
- Typed PA and Directional Drift tie on five rounded top-1 columns, yet their event-level states and probabilities differ.The mean maximum probability difference is .043–.046 across regimes; a separate full-method replay uses fixed channel configuration on existing synthetic streams.Source: §F.3 · pp. 20–21, Table 26 · p. 21, §F.4 · p. 21
G.1 Frozen Local Prompt; G.2 Recent-K History Block
- The frozen local prompt scores only current safe-slate options, while Recent-K appends revealed feedback records with event-local candidate mappings.The local prompt uses first-position logits and does not parse generated rationale; history records distinguish ACCEPT from REVOKE.Source: §G.1 · p. 21, §G.2 · p. 21
G.3 Hosted Jev Frozen Anchor
- 41.22% user-macro accuracy on TCPA-Repeat and 24.05% on TCPA-Open are reported for hosted Jev, which receives no replayed interactions or continual adaptation state.All completed natural-track calls returned jev-1.13.0.Source: Table 28 · p. 22, §G.3 · p. 22
G.4 Runtime State API; H.1 Artifact Chain; H.2 Compute
- Runtime updates validate state revisions and support reset, while the reproducible artifact chain retains canonical results, hashes, traces, and prediction-level files.Compute accounting separates scheduled GPU jobs from CPU replay and analysis; matched latency measurements keep post-cache CPU update time distinct from end-to-end model latency.Source: §G.4 · p. 22, §H.1 · p. 22, §H.2 · p. 22
H.3 Privacy and Data Use; H.4 Scope
- The evidence is limited to finite-slate immediate typed correction with a frozen scorer and train-only projection; derived per-user state may still encode behavior and requires personal-data governance.Evaluation uses disjoint users and predict-then-update replay; the evaluator sidecar is not a privacy boundary.Source: §H.4 · p. 23, §H.3 · p. 23