Ask, Clarify, Optimize: Human-LLM Agent Collaboration for Smarter Inventory Control¶

🕒 Published (v1): 2025-12-31 21:45 UTC · Source: Arxiv · link

Ask a follow-up

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

TL;DR¶

LLM end-to-end solvers incur a "hallucination tax" on stochastic inventory control because language models cannot perform grounded probabilistic reasoning. This paper proposes a hybrid agentic framework that strictly decouples natural-language elicitation (LLM) from mathematical optimization (classical (s,S) and DQN solvers), reducing total inventory costs by 32.1% over a GPT-4o interactive baseline. To evaluate the interactive system at scale, the authors introduce a fine-tuned "Human Imitator" as a reproducible digital twin of a boundedly rational manager.

Problem¶

Small and medium-sized businesses cannot access mature operations research tools for inventory control because translating informal business context into formally specified optimization inputs requires expert knowledge. Direct LLM solvers produce fluent but structurally unsound policies—misinterpreting lead times, conflating lost sales with backorders, ignoring Bellman optimality—compounding into recurring inefficiencies. Existing LLM frameworks also assume fully specified, static problem descriptions and cannot handle the ambiguity and bounded rationality of real managerial dialogue.

Method¶

The framework decomposes inventory decision-support into three specialized agents in a sequential pipeline:

Information Extraction Agent (GPT-5-mini): Maintains an explicit Parameter Specification Table (12 fields: time horizon, demand distribution, holding/penalty/setup cost, lead time, etc.). In each conversational round it updates defined fields, detects logical contradictions, and generates targeted clarifying questions. Dialogue terminates only when all parameters are marked defined and conflict-free. Outputs extracted_params.csv + conversation_log.txt.
Optimization Agent: Acts as a solver dispatcher, not a reasoner. Elicits a risk-tolerance parameter λ ∈ [−10, 10] and minimizes L(π) = E[Cost(π)] + exp(−λ)·Std[Cost(π)]. Grid-searches (s,S) parameter space and trains a Deep Q-Network (DQN) over the MDP formulation; selects the policy minimizing L(π). Serializes weights and evaluation results to deployment_artifacts/.
Policy Interpretation Agent: Ingests artifacts and translates policy into actionable guidance. If (s,S) wins, explains reorder point and order-up-to level in plain language. If DQN wins, frames it as an on-demand oracle and reports the cost penalty of the simpler (s,S) alternative.

Human Imitator for evaluation: Qwen2.5-7B fine-tuned via LoRA (SFT) on 1,184 human–machine conversational turns from 66 participants. Trained to reproduce bounded rationality: vague estimates, omissions, informal language. Perplexity drops from 37.99 (base) to 3.49 (Epoch 2 checkpoint); SBERT similarity rises from 0.282 to 0.493.

Key Contributions¶

Hallucination tax quantification: 32.1% cost reduction of hybrid framework over interactive GPT-4o end-to-end baseline.
Computational ceiling proof: Providing GPT-4o with perfect ground-truth parameters yields no statistically significant improvement, confirming the bottleneck is computational, not informational—prompt engineering cannot bridge the gap.
Complexity premium: Performance advantage scales with problem difficulty; the gap widens under longer lead times and high-penalty, high-flexibility regimes.
Human Imitator methodology: Scalable, reproducible evaluation testbed for interactive decision-support systems via supervised fine-tuning on real human–LLM dialogue; generalized evaluation template beyond inventory control.

Results¶

Hybrid agentic framework reduces total inventory cost by 32.1% relative to interactive GPT-4o end-to-end solver.
"Perfect information" counterfactual (ground-truth parameters fed directly to GPT-4o): no statistically significant improvement over standard GPT-4o baseline.
Human Imitator SFT: PPL 37.99 → 3.49 (Epoch 2); BERTScore 0.800 → 0.845; ROUGE-L 0.104 → 0.244; SBERT 0.282 → 0.493 on 1,184-turn corpus.
Performance gap increases with lead time length and in high-penalty/high-flexibility cost regimes.
Baselines: interactive GPT-4o as end-to-end solver; GPT-4o with oracle parameter injection.

Limitations¶

Human Imitator trained on only 66 participants (1,184 turns); behavioral diversity may be narrow and not representative of all SMB manager personas.
Epoch 3 fine-tuning shows BERTScore degradation (0.845 → 0.818), indicating risk of overfitting to lexical patterns; evaluation dataset and training set likely overlap.
Optimization Agent currently limited to (s,S) and DQN; more complex multi-echelon, multi-product, or supply-disruption scenarios are not evaluated.
No ablation isolating the contribution of explicit memory (Parameter Specification Table) versus LLM backbone strength.
Evaluation conducted with GPT-5-mini and GPT-4o; generalizability to open-source or smaller LLMs is not established.
Cost reduction metric is simulation-based; no real-world deployment validation.

Relevance to Harnesses / Meta-Harnesses¶

This paper is a concrete instantiation of a task-decomposition meta-harness: the system orchestrates three specialized sub-agents with strict interface contracts (CSV parameter file, JSON config, artifact directory) rather than routing all reasoning through a single model. The explicit memory architecture—a persistent Parameter Specification Table updated across turns—directly addresses context drift and hallucination that plague single-context harnesses handling long, multi-turn interactions. The modular solver-dispatch design (swappable (s,S) / DQN backends) exemplifies the harness pattern of separating orchestration logic from execution engines, enabling backend updates without altering the linguistic interface. The Human Imitator additionally provides a reusable evaluation harness for interactive multi-agent systems, addressing the benchmark gap that plagues harness-level assessment.