Source-linked AI summary
GUIDE: Generative Utility Inference and Decision Engine
Anagha Tiwari, Alexander G. Gray, Nick Feamster, Brian Jabarian, Alex Imas, Alex Kale
TL;DR
Preference elicitation struggles to efficiently capture complex preferences and ground them in domain knowledge. GUIDE combines adaptive Bayesian questioning with symbolic, data-grounded preference initialization and an observable architecture. In portfolio-optimization evaluations, it improves cold-start and early-turn recommendation quality under limited interaction budgets.
Problem
Existing elicitation approaches struggle to discover complex preferences efficiently, and standalone LLM systems can produce recommendations poorly grounded in domain knowledge or user preferences.
Source: Introduction · pp. 1–2Method
GUIDE combines Bayesian elicitation with diverse question types, symbolic domain initialization, and a transparent structure for analysis and expert oversight.
Source: Introduction · p. 2Results
GUIDE consistently improves cold-start and early-turn recommendation quality under limited interaction budgets in financial portfolio-optimization evaluations.
Source: Introduction · p. 2Takeaways & Limitations
GUIDE's preference state can serve as a portable user profile supporting personalization in downstream applications.
Source: Conclusion and Future Work · p. 8The authors identify evaluation with human subjects in varied applications as future work.
Source: Conclusion and Future Work · p. 8
Abstract
from arXiv · showhide
Measuring the preferences of human users remains a fundamental challenge of AI alignment. Existing elicitation approaches struggle to efficiently discover multidimensional preferences or accurately ground these inferences in domain knowledge. To address this, we introduce GUIDE, an LLM-driven elicitation architecture that infers user preferences through conversations by combining Bayesian adaptive sampling for question selection and symbolic representation learning to initialize domain-specific preference models. GUIDE generalizes adaptive sampling to diverse elicitation questions through an extensible type system of transforms on a parameterized preference state. GUIDE produces domain-specific preference representations through an initialization process using symbolic rule-based learning to capture world knowledge and set priors over preference dimensions grounded in data about decision alternatives. The architecture provides observability and steerability to facilitate deployment and analyze elicitation processes. In silico experiments on investment portfolio optimization demonstrate that GUIDE improves cold-start and minimizes recommendation regret consistently within early elicitation interactions across user personas compared to prior work, LLM-only baselines, and ablated GUIDE versions.
Introduction
Preference elicitation matters for informed decisions and AI alignment, yet existing systems struggle to capture complex preferences through narrow queries or ungrounded LLM inferences. GUIDE combines adaptive Bayesian elicitation, empirical domain initialization, and observable, steerable components; evaluations report improved cold-start and early-turn recommendation quality.
Introduction
- Existing Bayesian approaches often rely on pairwise questions, requiring many narrow elicitation cycles and struggling with preferences better expressed in open-ended formats.These approaches maintain uncertainty over latent utility, select queries, and update a posterior with new evidence.Source: Introduction · p. 1
- Standalone LLM preference elicitation lacks explicit preference uncertainty and consistent multi-step reasoning, while implicit domain knowledge can produce poorly grounded recommendations.These limitations motivate using LLMs as interfaces within structured elicitation, with human oversight and steering.Source: Introduction · pp. 1–2
- GUIDE combines Bayesian elicitation with diverse question types, symbolic domain initialization, and a transparent structure for analysis and expert oversight.Its initialization constructs interpretable rules and population priors grounded in domain-specific data.Source: Introduction · p. 2
- GUIDE consistently improves cold-start and early-turn recommendation quality under limited interaction budgets, while heterogeneous transforms introduce a late-stage stability trade-off.Evaluations cover diverse investor profiles and compare GUIDE with ablated variants, LLM-only approaches, and prior work.Source: Introduction · p. 2
- GUIDE makes LLM-based elicitation design more learnable by exposing preference representations and question-selection policies as observable, editable components.This structure supports systematic study of elicitation strategies and interactions between statistical and generative AI methods.Source: Introduction · p. 2
Preliminaries and Related Works
- Unlike OPEN's pairwise comparisons, GUIDE supports extensible question types and grounds preference dimensions in real domain data about decision alternatives.The question types include rule elicitation, dimension dominance and proposal, and free-text inputs.Source: Preliminaries and Related Works · p. 2
System Architecture
GUIDE maintains a structured preference posterior and adaptively selects questions through a two-phase elicitation loop. It combines domain initialization, diverse question transforms, developer review, and a recommendation process that can search between catalog alternatives.
System Architecture
- GUIDE updates a structured posterior after each turn, selects questions through a two-phase controller, and recommends when the particle posterior converges.The representation includes a posterior over utility weights, active dimensions, preference rules, and an interaction log.Source: System Architecture · p. 3, Structured Posterior State · p. 3
Structured Posterior State
- The posterior interaction log is replayed during particle rejuvenation, while soft rules shape inference and hard rules constrain recommendations.Source: Structured Posterior State · p. 3
Two-phase Question Selection
- Dimension discovery runs first; calibration then selects among pairwise, rule, and dimension-dominance questions, with a return to discovery when a new dimension is found.Population priors enter at the transition to calibration.Source: Two-phase Question Selection · pp. 3–4
Elicitation Question Transforms
- Question transforms compete by expected information gain, with GUIDE supporting pairwise, rule, dimension-dominance, free-text, and dimension-proposal inputs.Free-text statements become quasi-likelihood updates to the same particle posterior, while accepted proposed dimensions expand the axis set.Source: Elicitation Question Transforms · p. 4
Decision Alternatives and Perturbations
- Portfolio users whose optima fall between catalog entries motivate structured perturbations that add interior candidates at calibration entry.The evaluation disables perturbations so methods share a discrete action space.Source: Decision Alternatives and Perturbations · p. 4
Convergence, verification, and recommendation
- GUIDE stops after ranking entropy stays low and stable for several cycles, then recommends the highest-expected-utility feasible alternative under the converged posterior.Soft rules shape inference, hard rules filter infeasible alternatives, and the system also returns a runner-up and confidence measure.Source: Convergence, verification, and… · p. 4
Offline Domain Initialization
- An offline pipeline uses population data to produce a world model, feature-loaded preference dimensions, and per-dimension population priors for warm-starting users.In the portfolio domain, the pipeline uses the National Financial Capability Study Investor Survey.Source: Offline Domain Initialization · pp. 4–5
Developer Steering Checkpoint
- Mined population rules pass through developer review before use; their dimensions, directions, thresholds, confidences, and prior strengths are inspectable and editable.The checkpoint permits accepting, dropping, or editing each rule before it is cached for reuse.Source: Developer Steering Checkpoint · p. 5
Evaluation Methodology
The evaluation uses simulated investor personas and a shared portfolio catalog to measure recommendation regret over limited elicitation turns. It compares GUIDE with LLM baselines, ablations, and prior methods while measuring regret and convergence efficiency.
Evaluation Methodology
- The evaluation measures how quickly methods identify each simulated investor’s utility-maximizing portfolio, using recommendation regret over elicitation cycles.The task uses a fixed catalog and varied personas to assess methods against hidden ground-truth preferences.Source: Evaluation Methodology · p. 5, Experimental Setup · p. 6
LLM Baselines
- Four LLM baselines combine GPT-5-4 mini or Claude Opus 4.7 with either open-ended or pairwise questions.Source: LLM Baselines · p. 5
GUIDE Ablations
- Telescoping GUIDE ablations add domain-specific priors, dimension discovery, and preference transforms beyond pairwise comparisons.DI denotes domain initialization, comprising population priors and domain rule mining.Source: GUIDE Ablations · p. 5
PE Methods from Prior Work
- OPEN and PEBOL provide prior-work comparisons that test preference-dimension grounding and parameterization against learning decision boundaries directly.The implementations use GPT-5.4-mini for their LLM-dependent steps.Source: PE Methods from Prior Work · p. 5
Experimental Setup
- All methods select from the same 75-portfolio catalog across 50 simulations and 15 elicitation turns per persona.GUIDE’s continuous perturbations are disabled for the shared discrete action space.Source: Experimental Setup · p. 6
Evaluation Personas
- Five personas span aggressive, balanced, conservative, ESG-balanced, and global-balanced investment preferences, including less salient multidimensional trade-offs.A shared simulated client LLM responds naturally without revealing latent preference weights or dimensions.Source: Evaluation Personas · p. 6
Evaluation Metrics
- Regret measures the gap from each persona’s ground-truth utility, while NRR measures initial regret eliminated per turn through best performance, normalized for cold-start differences.The table reports NRR averaged across simulations per persona; higher values indicate greater efficiency.Source: Evaluation Metrics · p. 6, Table 2 · p. 6
Evaluation Results
Across evaluation results, GUIDE improves early elicitation efficiency through expressive querying and empirical grounding, while its observability supports analysis of preference dimensions and model coverage.
Evaluation Results
- GUIDE converges faster and more consistently than prior work and ablations through expressive querying, lowers cold-start regret through empirical grounding, and exposes interpretable coverage and recovery signals.These findings address the paper’s three requirements: expressive querying, empirical grounding, and observability.Source: Evaluation Results · p. 6
R1: Expressive Querying for Efficient & Consistent Convergence
- GUIDE’s NRR exceeds both prior-work systems on all five personas and the strongest ablation on four of five; early regret reaches a minimum as quickly as or faster than LLM approaches.Figure 2 reports regret across elicitation turns for investor personas, averaged over 50 simulations.Source: R1: Expressive Querying for Efficient… · p. 6, Figure 2 · p. 7
- Through turn 5, GUIDE full matches or outperforms pairwise-only, but later fluctuations let pairwise-only achieve lower final regret for Aggressive Young and Conservative Retiree.The authors attribute this bounce-back to competition among diverse transforms; they recommend full GUIDE for costly, limited interactions and simpler transforms for longer horizons.Source: R1: Expressive Querying for Efficient… · pp. 6–7, Figure 3 · p. 7
R2: Empirical Grounding for Lower Regret
- At turn 1, mean regret is 0.278 for GUIDE, 0.678 for OPEN, and 0.651 for PEBOL; by turn 5, GUIDE’s 0.038 regret is over 90% lower than both baselines.Values aggregate regret across 50 simulations and all personas.Source: R2: Empirical Grounding for Lower Regret · p. 7
- Symbolic domain initialization supplies population-grounded priors before elicitation, while dimension discovery identifies relevant preference axes during interaction, reducing cold-start effects.The pairwise-only ablation retaining both components has initial regret 0.279, while variants without initialization or discovery exceed 0.44.Source: R2: Empirical Grounding for Lower Regret · p. 7
R3: Observability into Model Coverage and PE
- For Balanced Midcareer, dimension contributions expose what drives recommendation margins, while diversification-balance coverage shows full and pairwise ablations overlapping ground truth, unlike ablations omitting initialization and discovery.Coverage analysis can inform revisions to initialization rules, vocabulary, or query transformations.Source: Figure 4 · p. 7, R3: Observability into Model Coverage… · pp. 7–8
Conclusion and Future Work
GUIDE combines expressive querying, empirical grounding, and observability in a framework for preference elicitation. The authors identify human-subject evaluation across applications as future work.
Conclusion and Future Work
- GUIDE improves early elicitation efficiency through expressive querying and empirical grounding, while observability enables analysis of the preference-elicitation process.The paper presents GUIDE as an LLM-assisted Bayesian framework for initializing domain-specific preference models.Source: Conclusion and Future Work · p. 8
- Future work should evaluate GUIDE with human subjects in various applications.The architecture is designed to facilitate this research.Source: Conclusion and Future Work · p. 8
- GUIDE’s extensible architecture supports new question types, selection policies, uncertainty representations, and domain logic; its preference state can serve as a portable profile for downstream personalization.The authors envision combining adaptive learning, symbolic reasoning, and human oversight across decision-making domains.Source: Conclusion and Future Work · p. 8
Particle Posterior Mechanics
GUIDE represents preferences with weighted latent dimensions, updates a particle posterior under rules and priors, and adds population priors when discovery transitions to calibration.
Dimensions and loading mappings.
- A preference dimension is a latent evaluative axis whose signed feature loadings encode preference direction and feature importance.Interior preferences can be represented with features shaped to peak at a desired point, such as a 50/50 equity/bond split.Source: Dimensions and loading mappings · p. 9
Prior form.
- The weight prior factorizes into exponential distributions, a maximum-entropy choice on nonnegative weights given a mean, with finite density at zero.This permits negligible weight on an irrelevant axis without penalty.Source: Prior form · p. 9
Degeneracy control and rejuvenation
- When normalized effective sample size falls below 0.5, GUIDE resamples and rejuvenates particles using jittered log-weights and Metropolis–Hastings acceptance.This restores particle diversity without biasing the targeted posterior.Source: Degeneracy control and rejuvenation · p. 9
Rules affecting posterior
- Soft rules penalize violating particles in proportion to confidence and violation magnitude, while hard rules filter infeasible alternatives at recommendation time.Later evidence can overturn a mistaken soft rule through rejuvenation.Source: Rules affecting posterior · p. 9
Prior injection at calibration entry.
- Population priors enter at the discovery-to-calibration transition, steering posterior weights and personalizing the feasible action space without biasing initial exploration.The personalized action space can reach a user’s interior optimum rather than being limited to published alternatives.Source: Prior injection at calibration entry · p. 9
Elicitation Transforms
GUIDE uses lookahead, rule elicitation, and dimension proposal transforms alongside posterior reconciliation to support diverse preference queries and updates.
Pairwise Lookahead
- Optional depth-bounded beam search evaluates question–outcome–follow-up chains, discounting later gains and pruning projection-equivalent pairs.The stated settings are depth 2, beam width 3, and discount δ=0.85.Source: Pairwise Lookahead · p. 9
Rule Elicitation.
- Rule elicitation probes contradictions, thresholds, and dependencies, filtering candidate questions with per-dimension beliefs about hard, soft, or absent constraints.Evidence transfers partially across dimensions according to overlap in their loaded feature sets.Source: Rule Elicitation · pp. 9–10
Dimension proposal.
- Dimension proposal adds axes only after grounding, feature-novelty, and coherence gates; passing axes receive prior-scale imputation and remain auditable through logged gate failures.Discovery can reconsider a rejected concept on later free-text turns, and user responses separately update its Beta confidence.Source: Dimension proposal · p. 10
Utility Updates
- After each cycle, GUIDE reconciles the posterior, reapplies dimension confidences, and reevaluates rules, promoting supported soft rules and demoting hard rules contradicted by user choices.This prevents a single mis-elicited constraint from dominating the final recommendation.Source: Utility Updates · p. 10
EIG Implementation Details
GUIDE scores candidate questions by expected reduction in ranking entropy, making diverse question types comparable. It adjusts these scores to favor productive questions and limit stagnation.
Expected information gain overview
- Each transform evaluates its candidate questions by counterfactual posterior reweighting, and GUIDE asks the single question with the highest expected information gain across transforms.The score uses expected reduction in ranking entropy, with outcomes weighted by their particle-mixture predictive probabilities.Source: Expected information gain overview · p. 10
EIG Augmentation Formulas
- Three multipliers adjust raw information gain: an effective-sample-size discount penalizes particle collapse, while realized success scales question types by their track record of reducing entropy.The success rate is tracked separately for each type using an exponential moving average.Source: EIG Augmentation Formulas · pp. 10–11
- A curvature bonus promotes questions on rising information surfaces and reduces scores near local peaks, while force-switching and resolved-axis deprioritization address local optima.Force-switching suppresses a dominant question type when recent entropy reduction is low; resolved-axis deprioritization down-weights dimensions with low posterior variance.Source: EIG Augmentation Formulas · p. 11
Domain World Model Initialization
GUIDE initializes domain-specific preference models from population data, then makes mined rules and their effects available for developer review. Its initialization combines numeric priors with symbolic rules for preferences that are categorical or conditional.
Developer Domain Inputs
- Developers specify behavioral and demographic variables, item-ownership indicators, and preference dimensions with signed mappings from behavioral proxies.These inputs define the vocabulary used to ground dimensions and condition rules on user segments.Source: Developer Domain Inputs · p. 11
Segmentation and discriminative rule discovery
- ILASP learns segment-conditioned rules from respondent records, using demographic and ownership values as antecedents and binned behavioral proxies as consequents.Proxy values are divided into quantile-based low, medium, and high groups; a rule is emitted when a segment’s mean differs from the population mean by at least the threshold, subject to minimum support.Source: Segmentation and discriminative rule… · p. 11
Symbolic preference knowledge rules
- A symbolic layer represents categorical or conditional preferences as hard constraints that disqualify alternatives or soft rules that penalize violations; population rules and dialogue-elicited user rules remain distinct.The rule forms include bounds on dimension weights, feature thresholds, and conditionals.Source: Symbolic preference knowledge rules · p. 12
Offline Domain Initialization Outputs
- Offline initialization produces a world model, dimension parameterizations, and population priors; segment rules shift starting priors toward the population segments matched at intake.Population statistics set the scale of dimension-weight priors and determine which dimensions are seeded, while segment rules help avoid cold-start.Source: Offline Domain Initialization Outputs · p. 12
Developer Steering Checkpoint: Full Mechanics
- A developer checkpoint lets experts accept, edit, or drop mined rules before prior injection, with validation checks and visibility into joint rule implications.The checkpoint can surface conflicting or redundant rules before they affect a live session.Source: Developer Steering Checkpoint: Full… · p. 12
Recommendation Details
GUIDE re-scores a fixed candidate set under the converged posterior and filters out alternatives that violate hard rules. It also challenges the leading recommendation with adversarial pairwise probes before finalizing.
Adversarial verification.
- Before finalizing, adversarial pairwise probes target flip fragility; GUIDE confirms the recommendation only if it survives unchanged, otherwise returning to calibration with the new evidence.Source: Adversarial verification · p. 12
Recommendation and Search
- The candidate set includes the catalog and perturbations built at calibration entry, allowing an interior-point recommendation; it is re-scored under the converged posterior and re-centered during calibration only when a perturbation wins pairwise choices.The recommendation maximizes expected posterior utility among alternatives satisfying every hard rule.Source: Recommendation and Search · p. 12
- Soft rules shape weights through penalties, while hard rules filter violating alternatives; GUIDE also returns a runner-up, feature deltas, and confidence derived from the utility gap.Source: Recommendation and Search · p. 12
- Table 3 reports regret values averaged across simulations and personas for each method.Source: Table 3 · p. 13
Evaluation Methodology
The evaluation compares implementations of OPEN and PEBOL with GUIDE using simulated investor personas whose hidden preferences define ground-truth utility. Table 3 summarizes regret across personas and simulations.
Prior Work Implementation Details
- OPEN uses a GPT-5.4-mini-generated binary-feature ontology and pairwise queries, while PEBOL asks binary questions about investment aspects and updates a Beta posterior over portfolio utilities.Source: Prior Work Implementation Details · p. 13
Simulated Persona Agents
- Each simulated investor has a description, opening message, and hidden linear utility function whose weights determine ground-truth preferences and simulated choices, but are not exposed to elicitation methods.Source: Simulated Persona Agents · p. 13
Summary Regret Table
- Table 3 summarizes regret values across persona agents for each method.The table caption specifies that values are averaged across all simulations and personas.Source: Summary Regret Table · p. 13