Jev-DLC: Decision-Level Continual Learning from Typed Feedback

Yangyang Feng; Jianle Gan; Zhuoyan Feng; Chengwei Qin

cs.AIOct 22, 2026

Jev-DLC: Decision-Level Continual Learning from Typed Feedback

Authors: Yangyang Feng, Jianle Gan, Zhuoyan Feng, Chengwei Qin

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · Sun Yat-sen University

Abstract

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.

Figures & tables

Figure 1: A typed correction exposes two forms of persistence. The selected action contributes evidence about what may recur, while the preferred–rejected pair gives a direction for repairing later decisions. Jev-DLC combines both through complementary bounded channels above a frozen scorer.

Figure 2: Typed feedback becomes persistent state above a frozen scorer. Conservative and responsive channels retain recurring action evidence and correction geometry. Causal per-user weights combine their distributions before commitment, and revealed feedback updates the channels and weights for later requests.

Method	Full	$d\geq 2$	$d\geq 4$	$d\geq 8$
Frozen	39.44	41.58	36.48	40.38
Recent-4	42.40	46.40	37.56	42.17
AdaGrad	41.99	48.53	39.33	41.68
PPlug	40.37	41.77	37.86	43.38
TAP-PER	41.99	46.48	46.45	44.25
MRM-style	40.67	42.14	47.18	51.88

Table 1: TCPA-Repeat user-macro accuracy (%). Every depth column evaluates decisions made after at least

d

earlier feedback events. Counts, paired bootstrap intervals, and the complete baseline ladder appear in Appendix D .

State	Full	$d\geq 2$	$d\geq 4$	$d\geq 8$
C only	43.12	54.85	48.60	44.30
R only	41.50	44.20	43.85	50.45
Jev-DLC	44.83	56.07	50.47	53.01

Table 2: Channel retention on TCPA-Repeat, in user-macro accuracy (%). Both channels contribute to the full Jev-DLC decision state.

Method	State	Retained	Input	Runtime
Recent-4	0 B	1,166.5 B raw	+262.8 tok.	Longer fwd.
PPlug	132 B	0 B	0	Projection
TAP-PER	128 B	1,040 B vec.	0	Query
MRM	16 B	0 B	0	3 low-D steps
Jev-DLC	324 B	0 B	0	Vector update

Table 3: Maximum per-user serving objects on TCPA-Repeat. State excludes the shared frozen scorer and train-time shared parameters.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$x_{t}$	current context
$\mathcal{A}_{t}$	current candidate slate
$M_{t}(a)$	immutable hard mask
$b_{t}(a)$	frozen option logit
$h_{t}(a)$	frozen candidate representation
$P$	shared train-only linear projection

Appendix

Table 4: Notation shared by the frozen scorer, persistent state, and proofs.

Object	Fields	Visibility and role
Candidate	candidate ID, type ID, text, metadata	visible before prediction
Runtime event	user ID, event ID, time, context, slate, track	visible before prediction
Prediction	event ID, choice, probability map, confidence, state version	committed before reveal
Target	event ID, target candidate/type, time, diagnostic labels	evaluator sidecar only
Feedback	event ID, kind, predicted ID, target ID, semantics	revealed after commit
State	counts, two correction vectors, reliability, version	persistent per user

Appendix

Table 5: Serializable objects in the chronological protocol. Target and diagnostic fields never enter the model-visible runtime event.

Split	Users	Events	Slate size
Train	747	5,216	15
Development	90	556	15
Test	92	711	15

Appendix

Table 6: TCPA-Repeat partitions after chronological user-disjoint compilation.

Method	Per-user state	Retained history	Extra input	Runtime update	Bytes
Frozen	none	none	0	none	0
Recent-4	none	4 typed interactions	+262.8 tok./event	scorer forward	1,166.5 mean
AdaGrad	state + diagonal accumulator	none	0	vector update	256
PPlug	sum + count	none	0	projection	132
TAP-PER	global state	8 vector pairs	0	memory query	1,168
MRM	4 coefficients	none	0	3 low-D steps	16

Appendix

Table 7: Matched serving contracts on TCPA-Repeat. Byte counts exclude the shared frozen scorer, projection, and train-time shared parameters. TAP-PER’s total includes its 128-byte state and at most 1,040 bytes of vector memory.

Frozen scorer	User-macro accuracy (%)
Raw Qwen option logprob	26.93
JEV-SFT	39.44
Hosted Jev	41.22

Appendix

Table 8: Current-event scorer anchors on the TCPA-Repeat test stream.

Prior depth	Users	Events	Frozen	Jev-DLC	Gain	95% CI
$\geq 0$	92	711	39.44	44.83	+5.39	$[+2.04,+8.74]$
$\geq 1$	85	619	39.21	46.67	+7.46	$[+2.12,+12.76]$
$\geq 2$	63	534	41.58	56.07	+14.49	$[+7.42,+22.63]$
$\geq 4$	29	434	36.48	50.47	+13.99	$[+2.01,+27.42]$
$\geq 8$	15	339	40.38	53.01	+12.63	$[+2.31,+27.23]$

Appendix

Table 9: TCPA-Repeat user-macro accuracy for Jev-DLC. Gains and paired user-bootstrap intervals are in percentage points.

Prior depth	Events	Gain	95% CI
$\geq 0$	459	+7.08	$[+3.18,+12.18]$
$\geq 1$	444	+7.57	$[+3.35,+13.14]$
$\geq 2$	429	+7.65	$[+2.92,+13.90]$
$\geq 4$	399	+8.13	$[+2.31,+16.24]$
$\geq 8$	339	+12.20	$[+2.54,+26.75]$

Appendix

Table 10: Fixed-cohort depth diagnostic for the archived directional PA-I replay. All rows use the same 15 users; gains are in percentage points over Frozen.

Method	Full	$d\geq 2$	$d\geq 4$	$d\geq 8$
Frozen	39.44	41.58	36.48	40.38
Recent-4	42.40	46.40	37.56	42.17
AdaGrad	41.99	48.53	39.33	41.68
DER++-8	41.94	48.45	39.14	41.20
Directional Drift	41.76	43.26	38.61	44.71
ONS	40.70	42.14	37.85	42.01

Appendix

Table 11: TCPA-Repeat user-macro accuracy across prior feedback depth. The table reports absolute user-macro accuracy at each prior-feedback depth. Other fusion configurations are reported separately in Table 20 .

Method	Full	$d\geq 1$	$d\geq 2$	$d\geq 4$	$d\geq 8$
Frozen	39.44	0.00	0.00	0.00	0.00
Action count prior	43.74	+5.69	+12.76	+13.23	+6.08
Recency action prior	43.87	+5.86	+13.01	+14.07	+6.05
Jev-DLC	44.83	+7.46	+14.49	+13.99	+12.63

Appendix

Table 12: Causal action-frequency variants on TCPA-Repeat. Full reports user-macro accuracy in percent; depth columns report gain over Frozen in percentage points.

Split/ $K$	Acc.	NLL	Extra tok.	Hist. B	Forward ms
Dev/4	48.20	1.549	241.5	1,082.2	453.10
Dev/8	47.89	1.547	392.8	1,759.7	296.75
Test/0	38.98	1.963	0	0	300.54
Test/4	42.40	1.791	262.8	1,166.5	449.90

Appendix

Table 13: Prompt reconstruction with the frozen scorer. The matched test runs provide the serving comparison; development latency is descriptive because

K=4

and

K=8

ran sequentially. Test/0 reports a separate empty-history forward run of the same checkpoint for matched latency measurement; Frozen reports predictions from the scorer cache used for online adaptation.

Method	Acc.	Retained	Extra tokens	Backward
Frozen	39.44	0 B	0	No
Jev-DLC	44.83	324 B state	0	No
Recent-4	42.40	1,166.5 B history	262.8/event	No

Appendix

Table 14: TCPA-Repeat serving footprint. Jev-DLC uses 0.1781 ms for prediction-side state computation and 0.0810 ms for the feedback update; Recent-4 reruns the scorer over its enlarged prompt.

TCPA-Repeat stratum	Frozen	Prefix-fit ref.	Gain
Full stream	39.44	42.56	+3.11
Length $\geq 4$	43.56	49.51	+5.94
Length $\geq 8$	48.56	54.86	+6.30

Appendix

Table 15: Chronological personalization headroom in user-macro accuracy percentage points. Stream length is distinct from prior feedback depth.

Comparison	Difference	95% CI
Recent-4	+2.43	$[-1.33,+6.30]$
Archived AdaGrad	+2.84	$[-0.60,+6.20]$
PPlug	+4.47	$[+1.15,+7.71]$
TAP-PER	+2.84	$[-0.39,+6.13]$
MRM	+4.16	$[+0.64,+7.71]$

Appendix

Table 16: Main Jev-DLC minus each archived external comparison on TCPA-Repeat Full, in percentage points.

Representation	Update	TCPA-Repeat	TCPA-Open
Frozen scorer	none	39.44	28.04
Learned linear	Mean	40.50	28.63
Learned linear	PA-I	41.95	28.70
MLP64	Mean	43.05	28.69
MLP64	PA-I	42.50	28.78

Appendix

Table 17: User-macro test accuracy for archived directional representation and accumulation controls. These rows vary the local correction path separately from the two-channel fusion experiment.

Projection with PA-I	Full	$d\geq 1$	$d\geq 4$	$d\geq 8$
Learned linear	+2.50	+3.76	+8.51	+12.20
MLP64	+3.06	+5.04	+10.08	+10.85

Appendix

Table 18: TCPA-Repeat gain over Frozen in percentage points.

Track	Past centroid	Matched	Shuffled
TCPA-Repeat	.340	.285	-.046
TCPA-Open	-.027	-.027	.005

Appendix

Table 19: Cosine diagnostics between past state and future correction directions. TCPA-Repeat exhibits alignment-conditioned reuse; TCPA-Open does not show the same geometry.

Configuration	Full	$d\geq 2$	$d\geq 4$	$d\geq 8$
$C$ only	43.12	54.85	48.60	44.30
$R$ only	41.50	44.20	43.85	50.45
$C+R$ , equal weights	42.65	50.70	47.90	51.20
$C+R$ , dev-fixed weights	44.15	55.42	49.85	48.15
$C+R$ , adaptive (Jev-DLC)	44.83	56.07	50.47	53.01

Appendix

Table 20: Jev-DLC channel and fusion configurations on TCPA-Repeat, reported in user-macro accuracy (%). The adaptive two-channel configuration (Jev-DLC) achieves the highest accuracy among the evaluated configurations on Full and at each reported feedback depth.

Stratum	Difference	95% CI
Full	+0.68	$[+0.16,+1.20]$
$\geq 2$	+0.65	$[-0.45,+1.75]$
$\geq 4$	+0.62	$[-1.30,+2.54]$
$\geq 8$	+4.86	$[+0.35,+9.37]$

Appendix

Table 21: Accuracy improvement of adaptive over development-fixed two-channel fusion (in percentage points). Paired confidence intervals are computed using 10,000 user-level bootstrap resamples.

Prior depth	Users	Events	Frozen	Jev-DLC	Gain	95% CI
$\geq 0$	4,209	9,845	28.04	29.74	+1.71	$[+1.14,+2.27]$
$\geq 1$	2,348	5,636	28.01	32.76	+4.75	$[+2.95,+6.52]$
$\geq 2$	1,292	3,288	26.82	31.94	+5.13	$[+2.65,+7.60]$
$\geq 4$	408	1,278	28.39	33.83	+5.44	$[+1.41,+9.45]$
$\geq 8$	84	290	28.37	27.78	-0.59	$[-9.88,+8.34]$

Appendix

Table 22: TCPA-Open user-macro accuracy for Jev-DLC. Gains and paired user-bootstrap intervals are in percentage points.

Regime	Frozen	Fixed fusion	Dynamic fusion
Stable	25.59	23.83	24.02
Partial shift	24.02	23.24	23.05
Noise	28.91	27.54	27.54
Recurrence	24.80	25.00	25.20

Appendix

Table 27: User-macro accuracy (%) on the existing Dynamics caches.

User Preference Modeling

User preference modeling infers what people value from choices, behavior, and feedback to personalize recommendations, predictions, and AI actions.

See all 271 papers

JevOut: Natural Context Can Flip Decision Models
Sep 24, 2026Zixiang Xu, Zirui Song, Chiyu Zhang +4Adversarial Attacks on LLMs LLM Decision-Making
University of Southern California
Confidence Is Not a Promise: A Reference Audit of Typed Decision Models
Oct 26, 2026Pedro Carvalho BromLLM Reliability LLM Auditing
Department of Mathematics, Federal Institute of Brasília, Brasília, Brazil
AnyJev Technical Report
Sep 30, 2026Jiamu Zhang, Tianze Yang, Yucheng Shi +6Multiple-Choice Question Answering Language Model Calibration
Nokia, Sunnyvale, CA, USA · Tencent Hunyuan · The Hong Kong Polytechnic University

Sep 24, 2026J/K move · Enter open · S save

JevOut: Natural Context Can Flip Decision Models

Zixiang Xu, Zirui Song, Chiyu Zhang, Xiuying Chen+3

University of Southern California

Adversarial Attacks on LLMs LLM Decision-Making Language Model Robustness

Open paper AskPDF arXiv

An ordinary-looking background detail can turn a correct model decision into a confident mistake. We demonstrate this fragility in four decision systems, including Jev, across seven datasets covering knowledge, reasoning, and tool routing. Within 64 accepted target evaluations per decision, we uncover short context additions that redirect 61.4%-73.2% of each system's initially correct decisions toward a wrong option fixed in advance. The additions supply background or procedural information rather than explicit answer-selection instructions, leaving the original question and choices intact. We construct them through probability-guided context optimization, which uses shifts in the option distribution to refine surrounding text under naturalness and answer-preservation constraints. Redirection affects initially confident decisions, often produces high-confidence wrong choices, and transfers across models. In blinded human evaluation, 91.6% of 250 sampled successful contexts are judged natural, answer-preserving, and free of decisive answer-changing evidence by a majority of three independent annotators. These findings expose a weakness in current decision models: context that looks entirely compatible with an input can redirect the choices that agents, routers, and evaluators rely on.

Abstract

User Preference Modeling

Explore similar work

JevOut: Natural Context Can Flip Decision Models

Confidence Is Not a Promise: A Reference Audit of Typed Decision Models

AnyJev Technical Report