Skip to content

ValuePilot: A Two-Phase Framework for Value-Driven Decision-Making

🕒 Published (v1): 2025-12-09 12:15 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

ValuePilot is a two-phase framework that enables AI agents to make personalized decisions aligned with individual users' value preferences rather than task-oriented objectives. It combines an LLM-based dataset generation toolkit (DGT) with a decision-making module (DMM) that uses T5 encodings and the PROMETHEE multi-criteria ranking method to select actions consistent with user-specified value weights across six dimensions (Curiosity, Energy, Security, Happiness, Intimacy, Fairness). The DMM outperforms GPT-5, Claude Sonnet 4, and other frontier LLMs at matching human action rankings in a held-out human study.

Problem

Existing LLM decision-making approaches (ReAct, AutoPlan, RLHF, DPO) optimize task completion or aggregate collective preferences, ignoring inter-individual variability in value priorities. No suitable dataset annotates decision scenarios with fine-grained, per-dimension value scores for training personalized value-driven agents. The result is agents that cannot generalize to novel scenarios while respecting individual value trade-offs.

Method

DGT (Dataset Generation Toolkit): A multi-stage pipeline using GPT-4 to generate domestic multi-agent scenarios with implicit value dimensions (keywords suppressed to prevent trivial recognition), 10 candidate actions per scenario, and bipolar value scores in [−1, +1] per dimension. Automatic re-evaluation filters samples where inferred values diverge from targets; a four-person human review team further curates quality.

DMM (Decision-Making Module): 1. Value Assessment Network: T5 encoder + 4-head self-attention + 2-layer MLP with tanh output produces an objective [−1, +1] score per (scenario, action, value dimension) triple. 2. Contextualized Scoring: User value preference vector p is sigmoid-scaled (eq. 1) to correct for moderate-response bias. Preference discrepancy scores d blend objective scores ρ with user preferences at weight w=0.3 (eq. 3). Action scores are further scaled by scenario relevance (eq. 4). 3. Action Selection via PROMETHEE: Pairwise preference degrees are computed per dimension (eq. 5), aggregated by user preference weights (eq. 6), and resolved into net outranking flows ϕᵢ (eq. 7) to produce an interpretable action ranking.

Key Contributions

  • Multi-stage LLM-based DGT pipeline generating 11,938 scenarios / 100,255 actions with interpretable per-dimension value annotations across 6 dimensions.
  • DMM that explicitly separates objective value assessment (Value Assessment Network) from subjective personalization (Contextualized Scoring + PROMETHEE), enabling interpretable action rankings.
  • PROMETHEE integration for principled multi-dimensional value trade-off resolution within a learned framework.
  • Novel Order-Sensitive Similarity (OS-Sim) metric extending the Jaccard index via progressive prefix comparison for evaluating ranked action lists against human ground truth.
  • Human study (40 subjects, 11 scenarios) demonstrating DMM surpasses frontier LLMs on both holistic alignment and first-choice accuracy.

Results

  • Value Recognition (Value Assessment Network vs. open-source LLMs):
  • AvgAcc at t=0.2: 66.70% vs. best baseline Gemini-1.5-Flash at 51.61% (+15.09 pp)
  • AvgAcc at t=0.05: 40.00% vs. 25.64% (+14.36 pp)
  • MAE: 0.19 vs. 0.24 (36.7% relative reduction)
  • Decision-Making (DMM vs. LLM baselines, 40 subjects):
  • OS-Sim: 73.16% ± 0.43% vs. GPT-5 at 69.23% ± 0.71% (+3.93 pp)
  • First-Action Accuracy: 46.14% ± 4.09% vs. GPT-5 at 38.01% ± 3.81% (+8.13 pp)
  • Ablation (OS-Sim / First-Acc):
  • Full DMM: 73.16% / 46.14%
  • w/o Scenario context: 69.99% / 43.64%
  • w/o Subjective preference adjustment: 68.93% / 43.45%
  • w/o Preference entirely: 61.07% / 31.82%
  • Action-only: 60.23% / 32.27%

Limitations

  • Human study is small (40 subjects, 11 scenarios), limiting statistical power and demographic generalizability.
  • Only 6 value dimensions are tested; the framework claims extensibility but does not demonstrate it empirically.
  • Scenarios are limited to domestic multi-agent settings; generalization to other domains (e.g., professional, adversarial) is unvalidated.
  • User value preferences are self-reported via questionnaire, which may not reflect revealed preferences in actual behavior.
  • Dataset is synthetic (GPT-4-generated) despite human filtering; distributional gaps vs. real-world value judgments remain unquantified.
  • The PROMETHEE ranking step is non-differentiable and cannot be end-to-end trained with the Value Assessment Network.

Relevance to Agentic AI / LLM Agents

ValuePilot directly addresses the personalization gap in agentic systems: how an autonomous agent should act not just to complete a task but to reflect a specific user's stable value priorities. The PROMETHEE-based multi-criteria action selection is a concrete alternative to prompting LLMs with preference descriptions, showing that a lightweight trained module with explicit value decomposition can outperform GPT-5 at matching human choices. For researchers building agents that must operate over long horizons in human environments (household robots, personal assistants, social agents), this work establishes value-driven decision-making as a tractable engineering pathway separate from reward shaping or aggregate alignment. The DGT toolkit also offers a reusable approach for bootstrapping value-annotated training data where none exists.