Could giving AI shopping agents pictures or more options to pick from make their choices less robotic?
Can multimodal cues or larger choice sets reduce agents' rule-like decision patterns?
This explores whether AI shopping agents' habit of snapping to rigid, all-or-nothing choices (for example, always picking the cheaper or higher-rated product) could be softened by giving them richer inputs, such as images and other sensory cues, or more options to choose from.
This explores whether the rigid, rule-like way AI agents make choices could be loosened by changing what they see (multimodal cues) or how many options they get (larger choice sets). The short answer: the corpus doesn't directly test either fix. What it does show suggests the rigidity comes from the model itself, not from thin inputs. That matters because it points to different remedies.
Start with the problem. When 17 LLMs and a group of humans faced the same price, rating, and nudge manipulations on binary product choices, humans shifted their choices modestly and probabilistically, while agents swung hard and collapsed into near-deterministic rules Do shopping agents respond more strongly to price and rating cues?. Note the setup: two options at a time. So the 'larger choice sets' question is genuinely open. The benchmark that best documents the problem never tested the variable you're asking about.
On modality, the strongest evidence points the other way. Researchers stripped agents down to headless, text-only interfaces and told them explicitly to ignore where a product appeared in the list. The agents still showed strong position bias, and each model had its own particular version of it Do AI shopping agents show persistent position bias in text interfaces?. If the bias survives removing the visual layer, adding a visual layer is unlikely to dissolve it. The quirks seem built into the model and resurface after every model update. Multimodal input does help agents in a different job: inferring what a user prefers by watching them over time, stored in an entity-centric memory Can agents learn preferences by watching rather than asking?. Richer cues there improve what the agent knows about you. Nothing shows they make its choices more humanlike.
The more surprising lead is where the rigidity might come from. Post-training appears to 'sharpen' models: it pushes outcomes toward all-or-nothing, wins easy cases, and wipes out rare-but-reachable alternatives. Base models with light prompting keep a wider spread of behavior Do base models find more solutions than post-trained ones?. That work studied agentic task-solving, not shopping, but the pattern looks a lot like agents collapsing onto a single rule. If the link holds, the lever is training, not input. A related result: LLMs fine-tuned on psychology-experiment data predict human decisions better than classic cognitive models and capture differences between individuals Can language models learn to model human decision making?. In other words, you can teach models to be probabilistic the way people are.
Why it matters beyond benchmarks: once agents carry out purchases rather than just recommend them, their decision habits feed back into what consumers come to want Do AI shopping agents reshape what consumers actually want?. A rigid rule inside the agent can turn into a shaped preference in the person. So the real question may be less 'what should we show the agent?' and more 'what has training taught it to be certain about?'
Sources 6 notes
A benchmark testing 17 LLM models against humans on binary product choices found agents exhibited much larger choice shifts in response to price, rating, and nudge interventions—collapsing into rule-like patterns rather than the modest probabilistic shifts humans showed.
ACES found that AI agents exhibit strong, model-specific position biases even in headless text-only interfaces with explicit instructions to ignore position. The bias is a deep-seated model characteristic, not a superficial visual artifact, and recurs after each model update.
M3-Agent demonstrates that separating episodic events from semantic knowledge in an entity-centric graph, combined with parallel memorization and control processes, allows agents to infer and act on user preferences without asking. This architecture mirrors human cognitive systems that bind disparate information about individuals across sensory modalities.
Across 14 model pairs and three agentic benchmarks, base models equipped with only relaxed system prompts eventually surpass post-trained counterparts in pass@K coverage as rollout budget grows. Post-training bimodalizes task outcomes, sharpening performance on easy cases while eliminating rare-but-reachable solutions.
LLMs finetuned on psychology experiment data predict human behavior more accurately than theory-driven models in decision tasks, capture individual differences in their embeddings, and transfer learning across tasks without task-specific design.
Show all 6 sources
Bilgihan and colleagues argue that transactional AI agents—which execute purchases rather than just recommend—end the consumer's sole cognitive role in preference formation. Preferences emerge through bidirectional feedback loops between agent learning and consumer trust, not independently beforehand.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments
- What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce
- Shopping by algorithm: How agentic AI deploys human heuristics as a surrogate consumer
- When Agents Shop for You: Role Coherence in AI-Mediated Markets
- From Product Search to Preference Articulation: The Economics of Agentic Commerce
- The AI Agent in the Room: Rethinking Consumer Decision-Making in the Age of Autonomous AI
- The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets
- Determinants of LLM-assisted Decision-Making