PaperlayerView exact revision on arXiv

Source-linked AI summary

SimInter: Evidence-Grounded User Simulation with Interview-Based Profile Enrichment for Conversational Recommender Systems

Erfan Shahabi, Amirhossein Roshandel, Mohammad Sadegh Sadeghi, Azadeh Shakery

arXiv:2610.18308v1cs.CLcs.AIcs.IR

TL;DR

LLM-based user profiles often omit behavioral dispositions that shape conversation. SimInter enriches profiles through evidence-grounded, audited synthetic interviews; in matched ReDial comparisons, it increases expressed aspect richness over a statistics-only baseline.

  • Problem

    Profiles based on observable signals capture consumption but leave behavioral dispositions that shape conversational behavior largely unspecified.

    Source: §1 · p. 1

  • Method

    SimInter grounds synthetic interviews in user statistics and historical conversations, audits turns for fabrication or stereotyping, and enriches profiles with evidence-linked attributes.

    Source: §1 · pp. 1–2

  • Results

    T3 aspect entropy was 1.619 vs. 0.859 for STATISTICALONLY (∆= +0.760, dz = 0.955, padj < 0.001).

    Source: §4.3 · p. 6

  • Takeaways & Limitations

    User-simulation quality is multidimensional: richer aspect expression coexists with concentrated item mentions and reduced lexical variety.

    Source: §5 · p. 8

    Generalization beyond English movie recommendation remains to be investigated.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Conversational Recommender Systems (CRSs) require reliable user simulators for cost-effective evaluation. LLM-based simulators have made realistic simulation possible but typically condition on profiles derived from explicit ratings, demographics, and static sources capturing surface preferences while omitting implicit behavioral characteristics such as decision-making style, aspect-level preferences, and domain expertise. We propose SimInter, a pipeline that elicits these attributes via synthetic interviews between two LLM agents, grounded in the user's statistical profile and historical conversational evidence. Before profile enrichment, an audit agent explicitly flags potential fabrication and stereotyping at the turn level. We instantiate and evaluate SimInter on movie recommendation. We simulate 300 ReDial users and, on a matched subset of 100 users, compare SimInter with a statistics-only baseline: SimInter significantly increases the diversity of expressed preference aspects (filtered aspect entropy: 1.619 vs. 0.859; paired d = 0.96, Holm-adjusted p < 0.001). This gain persists when the simulator converses with two trained recommenders, KBRD and UniCRS, under two independent evaluation judges (+0.28 to +0.56 nats; all p < 0.001). Human evaluation finds 94.3% of enriched-attribute claims supported or partially supported, while human judgments agree with 95.2% of the auditor's binary filtering decisions on a stratified sample.

1 Introduction

LLM-based simulators are limited by profiles that capture consumption but omit behavioral dispositions, while direct inference and self-play risk stereotyped or fabricated traits. SimInter grounds synthetic interviews in user evidence, audits them, and enriches profiles with supported attributes.

1 Introduction

  • Real-user CRS evaluation is expensive, slow, and rarely reproducible, while LLM-based simulators can produce repetitive responses, narrow choice ranges, binary reactions, and generic-user behavior.The paper identifies the conditioning profile as an important source of these simulator limitations.Source: §1 · p. 1
  • Ratings, interaction histories, and demographics describe consumption but leave decision-making style, mood or social-context influences, and openness to unfamiliar recommendations largely unspecified.Inferring those attributes from static profiles or self-play can produce stereotyped or fabricated claims that reduce individual fidelity.Source: §1 · p. 1
  • SimInter combines statistical grounding, asymmetric interview dialogue, turn-level auditing, and evidence-linked profile enrichment.The audit distinguishes unsupported claims from invented rationalizations, and enriched claims are paired with supporting evidence quotes.Source: §1 · pp. 1–2
  • On 300 ReDial users, SimInter increased aspect richness over statistics-only profiles (T3, d = +0.96, Holm-adjusted p < 0.001).The evaluation used matched 100-user comparisons across four tracks.Source: §1 · p. 2

2 Related Work

Prior user simulators span rule-based, neural, and LLM-based approaches, with newer systems adding richer user models. SimInter instead elicits user-specific behavioral attributes through historically grounded interviews and audits claims before enrichment.

2 Related Work

  • Earlier simulators used agenda-based control, while LLM-based frameworks enable flexible interaction but raise concerns about realism, data leakage, and controllability.GenTUS is cited as an agenda-based framework extended with neural generation; iEvaLM is an LLM-based example.Source: §2 · p. 2
  • Recent systems model users through bounded rationality, assigned or inferred personality traits, review-derived personas, or controllable target-conditioned simulation.Examples include RecUserSim, PerCRS, PUB, PEARL, and CSHI.Source: §2 · p. 2
  • SimInter elicits implicit attributes from user-specific dialogue grounded in historical conversations and audits potentially fabricated or stereotyped content before profile enrichment.The conceptual comparison covers profile sources, ground truth, reliability mechanisms, evaluation objectives, and data requirements.Source: §2 · p. 2

3 Method

SimInter builds behavioral profiles from ReDial histories, uses them to ground and audit synthetic interviews, then enriches profiles for CRS user simulation.

3 Method

  • The six-stage pipeline constructs interaction-based profiles, extracts behavioral statistics, generates and audits interviews, enriches profiles, and simulates CRS interactions.The described components include interview generation, auditing, enrichment, and user-simulation models.Source: §3 · p. 2, Figure 1 · p. 3

3.1 Basic User Profiles and Behavioral Statistics

  • Profiles combine observed movie preferences and conversation histories with rule-based statistics such as acceptance rates, reference style, genre distributions, exploration ratio, and popularity score.From 543 eligible users, 300 were sampled for the full pipeline and 100 for matched comparisons.Source: §3.1 · p. 3

3.2 Synthetic Interview Generation

  • The interviewee receives the user’s statistical profile and a historical seeker conversation, while the interviewer elicits information without access to that profile or history.Interviews use an initial six-turn elicitation phase followed by four clarification turns.Source: §3.2 · pp. 3–4

3.3 Conversation Auditing

  • An Audit Agent checks interviewee turns for fabrication, caricature, and contradictions with profile or conversation evidence, alongside deterministic structural checks.Structural errors or at least three high-severity flags trigger regeneration; other severity thresholds determine whether flagged turns are excluded or the interview accepted.Source: §3.3 · p. 4

3.4 Profile Enrichment

  • Audited interviews provide six implicit attributes, including aspect preferences, pickiness, decision-making style, viewing motivation, novelty preference, and domain expertise.The inferred attributes and pickiness are combined with the statistical profile to condition later simulation.Source: §3.4 · pp. 4–5

3.5 User Simulation

  • Enriched profiles condition user simulations across four opening strategies: aspect, movie, mood, and expertise prompts.A termination judge evaluates exchanges from the fourth CRS turn onward, and conversations are capped at eight CRS turns.Source: §3.5 · p. 5

4 Evaluation

The evaluation compares SimInter with statistics-only and CSHI baselines on ReDial, measuring item diversity, population-level rating alignment, aspect richness, and linguistic variety. Interview-based enrichment substantially increases aspect richness, with human evaluations and cross-recommender tests providing additional evidence about the pipeline.

4.1 Experimental Setup

  • The study ran FULL on 300 ReDial users and matched 100 users for baseline comparisons, assessing item diversity, rating alignment, aspect richness, and linguistic variety.The matched comparison includes STATISTICALONLY and CSHI; T2 correlations are descriptive.Source: §4 · p. 5, §4.1 · p. 5

4.3 Results

  • FULL had lower T1 item diversity than STATISTICALONLY, while its JSD from ReDial was 0.681 versus 0.676; T2 rating correlations were descriptive, not paired tests.The reported correlations were calculated over condition-specific movie sets and do not establish individual-level preference fidelity.Source: §4.3 · p. 6
  • Filtered T3 aspect entropy was 1.619 for FULL versus 0.859 for STATISTICALONLY (∆= +0.760, dz = 0.955, padj < 0.001); FULL was higher for 81 of 100 users.Zero aspect entropy occurred for 5% under FULL versus 41% under STATISTICALONLY.Source: §4.3 · p. 6, Figure 4 · p. 6
  • FULL's slightly higher SBERT diversity was nonsignificant, while TTR fell and n-gram monotony rose; versus CSHI, FULL exceeded it on T1, T3, and T4, but protocol differences confound that comparison.The TTR and monotony differences were uncorrected; CSHI steers conversations toward predefined target items.Source: §4.3 · pp. 6–7

4.4 Auditor Behavior

  • The auditor raised 676 flags, chiefly caricature (372), and excluded 196 interviewee turns; only one interview reached the REGENERATE threshold.Other corrective action was turn-level exclusion under PARTIAL; the flags also included 173 history contradictions and 131 fabrications.Source: §4.4 · p. 7

4.5 Human Evaluation of Audit Decisions

  • On a stratified 105-turn sample, the auditor achieved 95.2% binary filtering accuracy against majority human judgments.This is human-referenced accuracy on a balanced sample, not a prevalence-adjusted estimate across all audited turns.Source: §4.5 · pp. 7–8

4.6 Human Evaluation of Enriched Attributes

  • Human annotators rated 94.3% of enriched-attribute claims supported or partially supported, while decision-making style had the highest unsupported or contradicted rate at 21.4%.Strict support was highest for pickiness (72.7%) and liked aspects (75.0%); pairwise weighted κ ranged from 0.40 to 0.52.Source: §4.6 · p. 8, Table 3 · p. 8

4.7 Cross-CRS Robustness

  • The aspect-entropy improvement persisted with KBRD and UniCRS under two independent judges, with paired gains from +0.265 to +0.369 and p < 0.001.Including all technically completed conversations produced positive effects from +0.275 to +0.560; the findings do not establish overall simulator fidelity.Source: §4.7 · p. 8

5 Conclusion

SimInter enriches LLM-based user simulators with behavioral attributes elicited through structured interviews and protected by turn-level auditing. On ReDial, it produces richer aspect-level preferences than a statistics-only baseline, alongside trade-offs in other diversity measures.

5 Conclusion

  • dz = 0.96 (padj < 0.001): enriched-profile simulators discuss movies along richer behavioral aspects than a statistics-only baseline.Zero-aspect users drop from 41% to 5%.Source: §5 · p. 8
  • On 300 ReDial users, simulators conditioned on enriched profiles discuss substantially richer behavioral aspects than a statistics-only baseline (dz = 0.96, padj < 0.001), and 94% of enriched-attribute claims are human-supported.Source: §5 · p. 8
  • Enrichment concentrates item mentions and reduces lexical variety, so SimInter's gains in aspect richness coexist with losses on other diversity dimensions.The paper frames CRS simulation quality as multidimensional, rather than seeking to maximize every diversity metric.Source: §5 · p. 8

Limitations

The evidence does not establish direct behavioral fidelity or the interview's independent contribution, and the evaluation remains bounded by sparse data and a movie-focused setting.

Limitations

  • Human evaluation assesses whether enriched-attribute claims are evidence-supported, not whether they faithfully match real-user behavior; independent validation against real behavior remains future work.Audit reliability is also limited by uncertain judgments about unsupported rationalizations, automated evidence retrieval, and annotator differences.Source: Limitations · p. 9
  • The ablations omit direct LLM-based attribute inference without interviewing, and sparse preference evidence limits individual-level alignment evaluation.Cross-CRS aspect findings rely on automated evaluation; generalization beyond English movie recommendation remains uninvestigated.Source: Limitations · p. 9

Ethical Considerations

The appendix bounds claims about popularity alignment because per-user evidence is sparse and liked/disliked labels are imbalanced. The paper also specifies research-only use and licensing conditions for its data, implementations, and APIs.

Ethical Considerations

  • ReDial is CC BY 4.0, MovieLens-25M is available for non-commercial research, and the listed implementations and APIs are used under their stated licenses or provider terms.The authors state that all listed resources are used solely for research.Source: Ethical Considerations · p. 9

A T2 Degeneracy Diagnostics

  • The appendix reports only pooled T2 popularity correlation in the main text and documents why per-user and persona-anchored variants are not reportable.Source: A T2 Degeneracy Diagnostics · p. 11

A.1 Per-User Popularity Correlation

  • Only 6/300 FULL users reach the ∼30-pair threshold, so paired per-user testing is infeasible and the pooled estimate treats user–movie pairs as independent observations.The authors characterize the pooled estimate as a weaker statistical claim than per-user testing.Source: §A.1 · p. 11

A.2 Persona-Anchored T2 (Liked vs. Disliked)

  • 94–97% of high-confidence liked/disliked ground-truth labels are liked in every condition, making Pearson r unreliable for discriminating conditions.The authors therefore do not interpret the apparent ordering across conditions.Source: §A.2 · p. 11

A.3 Balanced Accuracy as an Alternative

Balanced-accuracy evaluation is constrained by the scarcity of users with both liked and disliked ground-truth items. The authors recommend larger samples or prompts that elicit balanced acceptance and rejection.

A.3 Balanced Accuracy as an Alternative

  • Only 2 users per condition have both liked and disliked ground-truth classes, too few for a meaningful per-user balanced-accuracy distribution; pooled 94–96% accuracy merely matches the majority-class rate.The table caption likewise notes near-constant labels in every condition.Source: §A.3 · p. 11, Table 4 · p. 12
  • Future per-user alignment metrics require either larger samples with most users having ≥30 rated items or controlled prompts targeting a balanced acceptance/rejection ratio.The authors give 50/50 targeted CRS recommendations as an example.Source: §A.3 · pp. 11–12

B.1 Flag Taxonomy

Invented rationalizations dominate the audit flags, while structural interview failure is rare. The examples distinguish supported preferences from unsupported explanations and show the experiential vocabulary FULL adds beyond STATISTICALONLY.

B.1 Flag Taxonomy

  • 372 of 676 flags were caricature, mostly invented rationalizations; only one of 300 interviews triggered REGENERATE.The dominant issue was local rather than global structural failure.Source: §B.1 · p. 12
  • Evidence-based contradiction flags were directly checkable, while film-fact flags sometimes misclassified accurate scene descriptions and could remove grounded content.These conservative errors did not introduce unsupported claims into the profile.Source: §B.1 · p. 12

B.2 Representative Example 1: ACCEPT

  • The ACCEPT example contained no flags because every mentioned film appeared in the user’s liked or disliked lists.The interview included Juno, The Perks of Being a Wallflower, and White Noise as preference examples.Source: §B.2 · p. 12

B.3 Representative Example 2: PARTIAL (history contradiction)

  • A high-severity history contradiction excluded the claim of enjoying Blue Streak because the user had said they had not seen it.The interview also received a low-severity caricature flag.Source: §B.3 · p. 12

B.4 Representative Example 3: PARTIAL (caricature)

  • The auditor excluded a Justice League criticism because the user’s profile listed the film as liked without negative commentary.This was classified as caricature: a plausible but unsupported reason for a dislike.Source: §B.4 · pp. 12–13

B.5 Representative Example 4: PARTIAL (invented rationalization)

  • The Hangover and Superbad were disliked, but calling them “loud and obnoxious” was an invented rationale inconsistent with the user’s liking for comparable comedies.Both flagged turns were excluded from enrichment.Source: §B.5 · p. 13

D T3 Aspect Vocabulary

  • FULL introduced experiential aspects such as pacing, vibe, and energy, while STATISTICALONLY concentrated on holistic movie judgments.The top-20 aspect comparison filters generic placeholders.Source: D T3 Aspect Vocabulary · p. 13

E Cross-CRS Robustness with Trained Recommenders

With KBRD and UniCRS, enriched profiles yielded positive T3 aspect-entropy effects under both judges, including in an all-conversation sensitivity analysis. The appendix also details audit filtering and human annotation procedures.

E Cross-CRS Robustness with Trained Recommenders

  • The robustness experiment paired FULL and STATISTICALONLY profiles across KBRD and UniCRS, holding user identity, behavioral statistics, openings, budgets, and recommender configuration fixed.It generated 1,920 conversations across 120 users, two recommenders, two conditions, and four opening strategies.Source: E Cross-CRS Robustness with Trained… · p. 13
  • A separate Phase-2 audit filtered conversations for fabrication, caricature, role confusion, or persona drift before T3 extraction; it changed inclusion but did not regenerate conversations.ACCEPT and PARTIAL were retained, while REJECT conversations were excluded; paired users were the intersection evaluable in both conditions.Source: E Cross-CRS Robustness with Trained… · p. 15
  • For both recommenders and judges, the primary audit-filtered T3 effect was positive; all-conversation effects were +0.275/+0.371 for KBRD and +0.560/+0.534 for UniCRS under GPT-4o-mini/DeepSeek-chat, respectively (all p < 0.001).The paired analysis pooled aspect–sentiment pairs across each user's retained conversations before computing entropy.Source: E Cross-CRS Robustness with Trained… · p. 16, Table 7 · p. 16

F.1 Annotation Task

  • Annotators judged whether each enriched-profile claim was supported by the interview, alongside the user’s behavioral profile and transcript.Claims could be Supported, Partially supported, Unsupported, or Contradicted.Source: §F.1 · p. 16, §F.2 · pp. 16–17

F.2 Rating Scales

  • The annotation scales also rated evidence grounding, genericity, stereotype use, consistency, usefulness, overreach, and annotator confidence.Across 315 annotations, mean usefulness was 3.61/5, overreach 1.70/5, and confidence 3.00/5.Source: §F.2 · p. 17

F.3 Annotator Pool and Procedure

  • Three annotators independently rated 105 claims from 45 users, covering all six attribute types; no calibration round preceded annotation.Decision-making style had the highest Unsupported rate at 21.4%.Source: §F.3 · p. 17, Table 8 · p. 17

F.5 Inter-Annotator Agreement Details

Agreement was strongest for binary support judgments, while ordinal and grounding labels showed weaker agreement and recurring boundary disputes. The paper argues the aggregate support result remains robust across reasonable agreement assumptions.

F.5 Inter-Annotator Agreement Details

  • Quadratic-weighted κ suits the ordinal support labels because it penalizes large rating disagreements more than small ones.The labels range from Supported through Partially supported and Unsupported to Contradicted.Source: §F.5 · p. 17
  • Grounding-label κ ranged from 0.12 to 0.37, with disagreement mainly at the Generic versus Evidence-grounded boundary.Only one of 315 annotations was labeled Cliche/stereotype.Source: §F.5 · p. 17
  • Gwet’s AC1 was 0.89 for binary support and 0.99 for Contradicted versus all other labels, despite binary κ being depressed by label prevalence.All three annotators agreed on 85.7% of claims under the binary split.Source: §F.5 · p. 18
  • Supported versus Partially supported was the most common disagreement, occurring in 32% of annotator pairs per claim, and the protocol did not define the evidence threshold clearly.Annotator C was stricter, but aggregate support remained at least 91% under any single-annotator-out analysis.Source: §F.5 · p. 18

G Human Evaluation Protocol for Audit Decisions

The human audit evaluation compares Stage 4 outcomes with majority judgments on 105 stratified synthetic turns, then illustrates enriched profiles and simulated conversations. The examples show how FULL and STATISTICALONLY differ in the aspects they express.

G Human Evaluation Protocol for Audit Decisions

  • The 105 turns came from distinct users, equally divided among auditor outcomes, and three co-author annotators used majority vote as the whole-turn human reference.Annotators saw synthetic responses and retrieved historical evidence, but not auditor actions, flags, or explanations.Source: G Human Evaluation Protocol for Audit… · p. 18
  • Annotators agreed unanimously on 70/105 cases; nominal Fleiss’ κ was 0.664, while within-stratum κ values ranged from 0.268 to 0.337.The authors note that aggregate agreement is influenced by differences in label distributions across strata.Source: G Human Evaluation Protocol for Audit… · p. 18
  • 86.7% three-class accuracy and 95.2% binary filtering accuracy were achieved against majority human judgments on the 105-turn stratified sample.The three-class comparison is shown in Table 10; the binary comparison collapses the two retained categories.Source: Table 10 · p. 19, G Human Evaluation Protocol for Audit… · p. 19
  • Fourteen cases disagreed at the three-class level; nine disagreements between retained categories did not change the final filtering action.The evaluation does not establish the validity of individual issue flags or population-level recall, and its stratified sample accuracies are not unbiased estimates for all Stage 4 turns.Source: G Human Evaluation Protocol for Audit… · p. 19

H.1 Schema Definitions

  • The six enriched attributes cover aspect-level preferences, pickiness, decision-making style, viewing motivation, novelty preference, and domain expertise.The example profile for User 468 includes preferences for action choreography, rational decision-making, arousal motivation, and intermediate expertise; its pickiness score is computed from rejection rate.Source: §H.1 · pp. 19–20

I.1 FULL (opening = mood)

  • Under FULL, User 468 asks for an action thriller with relentless pacing and choreography, while STATISTICALONLY opens with a broader request for intense action and a crime or thriller plot.In the FULL exchange, the user accepts The Raid: Redemption and later asks for similar high-energy films; the STATISTICALONLY exchange includes acceptance of Heat.Source: §I.1 · p. 20, §I.2 · p. 20

I.3 Observations

  • Across four conversations, FULL yielded eight distinct aspects and filtered entropy 1.75, versus two aspects and entropy 0.69 for STATISTICALONLY.The FULL simulator voiced enriched-profile aspects such as pacing and choreography; 11 of 13 extracted STATISTICALONLY aspects were generic and filtered out.Source: §I.3 · p. 20
  • “Relentless” appeared in three of the four FULL opening requests, matching vocabulary in the enriched profile.The appendix also notes that across the four FULL openings, pace and choreography recur as movie-related requests.Source: §I.3 · p. 21

J Full System Prompts

Appendix J documents the prompts and settings for the pipeline’s interview, audit, profile-enrichment, simulation, and evaluation components, alongside API costs for the primary experiment. It specifies how interview claims are grounded, audited, and used downstream, and reports the generation and evaluation cost breakdown.

J Full System Prompts

  • The appendix reproduces prompts in condensed form, using braces for runtime-filled slots and “[...]” to mark omitted examples or instructions.Source: J Full System Prompts · p. 21

J.1 Interviewer Agent (Stage 3)

  • Interviews comprise ten messages across Account and Clarify phases; generated questions use only prior dialogue, while the interviewer avoids assumptions and probes the user’s stated preferences.The first message is one of four predefined open-ended questions selected uniformly at random; Account invites broad responses, and Clarify probes mechanisms, contrasts, or personal meaning.Source: §J.1 · p. 21

J.2 Interviewee Agent (Stage 3)

  • The interviewee receives the statistical profile and a seeker-side ReDial conversation, and must keep responses grounded in them rather than inventing unsupported personal stories.The prompt encourages concrete movie-specific reasoning and allows “I’m not sure”; an audit-triggered regeneration can append guidance to avoid unsupported claims.Source: §J.2 · pp. 21–22

J.3 Audit Agent (Stage 4)

  • The audit checks structure, movie-sentiment consistency, ReDial contradictions, fabrication, and caricature, then accepts, removes flagged turns, or regenerates by deterministic thresholds.It audits interviewee turns; fabrication requires specific ungrounded claims, while caricature flags stereotypes inconsistent with the profile. Structural errors or at least three high-severity flags trigger regeneration.Source: §J.3 · pp. 22–23

J.4 Profile Enrichment (Stage 5)

  • Profile enrichment extracts six fields from dialogue, using the existing profile only to resolve ambiguity or calibrate scores; undeterminable fields are null and pickiness is assigned separately from rejection rate.The fields are aspect preferences, pickiness, decision-making style, viewing motivation, novelty preference, and domain expertise; pickiness is deterministically mapped from rejection rate.Source: §J.4 · p. 23

J.5 User Simulator (Stage 6)

  • The simulator opens using one of four strategies, then responds consistently with its profile and pickiness level without reciting the profile or inventing watched movies.Strategies are aspect, movie, mood, and expertise; pickiness rules range from default acceptance at low scores to specific rejection reasons at high scores.Source: §J.5 · pp. 23–24

J.6 Evaluation Judges

  • During simulation, the termination judge ends conversations for specified endpoints, while movie extraction records canonical titles, introducers, and mention turns.The simulation otherwise caps conversations at eight CRS turns; movie ratings are assigned only when users express an actual evaluative reaction, not when a title is only a comparison anchor.Source: §J.6 · p. 24, §J.5 · p. 24
  • Evaluation prompts also include aspect–sentiment extraction for T3, alongside the movie-mention and rating procedures.Source: §J.6 · p. 24
  • The aspect judge extracts all user-mentioned movie aspect–sentiment pairs, with a short quote as evidence, and returns no pairs when no specific aspect opinions are expressed.Aspects are concrete movie elements, and sentiment is positive, neutral, or negative.Source: §J.6 · pp. 24–25
  • Before computing aspect entropy, post-processing removes generic placeholders such as “movie(s),” “story/stories,” “thing(s),” and “stuff.”The listed placeholders also include “film(s),” “overall movie,” “overall film,” and “experience.”Source: §J.6 · p. 25
  • The acceptance judge classifies only CRS-recommended movies as accepted, rejected, or neutral based on the user’s response.Acceptance requires positive intent or agreement to watch; rejection requires explicit decline or disinterest; ambiguous, hedged, or absent responses are neutral.Source: §J.6 · p. 25

K Model Configuration and API Cost

Related papers

This paper cites

Loading 2610.18308v1…