{"id":"7b1fa348-d53c-42bb-86b2-affb150c63ca","arxiv_id":"2607.06993","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Grounding an LLM in verbalized transaction histories via Person–Environment prompting, continued pre-training, SFT, and GRPO yields stronger retail decision simulation than frontier models, with partial cross-domain transfer.","lead":"A language model trained on real retail purchase histories can simulate individual customer decisions better than frontier general-purpose models on several shopping tasks. Retailers could use such digital twins to test promotions and product choices without relying only on surveys or generic AI priors.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"The headline claim of learned behavioral simulation is undercut by the paper’s own evidence that accuracy is dominated by in-prompt similarity cues, not internalized customer decision-making.","rationale":"The reader correctly identifies the soft spot: the paper’s strongest scientific claim is not merely that a RAG+LoRA pipeline beats GPT-5.5 on private retail tasks, but that transaction-grounded training produces a faithful simulator of individual decisions. That claim requires residual behavioral competence after explicit in-prompt evidence is controlled. The manuscript repeatedly shows the opposite pattern (Figure 3, §3.6–3.8, GRPO design) while still marketing “learned behavioral knowledge” and digital twins. In-domain tables and Lazada zero-shot transfer remain useful engineering results and keep the paper above REJECT; proprietary data, modest customer count, thin classical baselines, and failed transfers (Shopee, Tmall) already justify CONDITIONAL. No stronger independent concern (e.g., internal inconsistency of the Person–Environment math) is needed. The concrete matched-prompt ablation would settle whether CPT/GRPO add behavior beyond cue reading; until then the reader’s CONDITIONAL verdict and weakest-assumption diagnosis stand.","tokens_in":12145,"tokens_out":787,"duration_ms":6922,"concrete_test":"Rebuild B-hard v4 and classic-B with three matched prompt ablations on the same held-out baskets: (A) full Shopping-DNA + top-3 RAG with similarity/affinity lines as published; (B) same profile and candidates but remove all explicit similarity, co-occurrence, and affinity phrases; (C) candidates only, no customer history. Report LBM (CPT+SFT+GRPO) vs GPT-5.5 accuracy/AUC per task. If LBM’s advantage over GPT-5.5 collapses under (B) relative to (A), the learned-behavior claim weakens to prompt-evidence engineering.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that LBM learns customer decision-making from longitudinal transactions so a Person–Environment model outperforms frontier LLMs via behavioral knowledge (Abstract; Contributions 2,5; §6). The load-bearing condition is that gains come from internalized behavioral simulation rather than better packaging/reading of explicit evidence. The paper itself undermines that condition: §3.6 and Figure 3 show B-hard accuracy near chance with no in-prompt evidence (v1), ~60% with raw similar-item names (v2), and 95.9% with one legible similarity line (v3), and state “the bottleneck is prompt signal, not weights” and that the model “primarily operates as an evidence-conditioned decision system rather than a latent similarity engine.” GRPO is explicitly for “evidence-based calibration” over LM priors (§3.4, §3.8). Classic-B and B-hard gains (Tables 3–4), especially B2 hard-negative discrimination, are therefore consistent with superior use of Shopping-DNA/RAG cues that GPT-5.5 may not receive in the same form. CPT may still help format or domain language, but the paper does not isolate residual accuracy after stripping explicit similarity/affinity lines while holding the same retrieved candidates and profile text. Without that isolation, “learned individual behavioral simulation” and “digital twin” claims over-reach relative to evidence-conditioned prompting.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes the Large Behavior Model (LBM), a language-model framework that simulates retail customer decisions via a Person–Environment factorization: a Shopping-DNA behavioral profile (plus segment-level LoRA) for the person, and retrieval-augmented product context for the environment. Training chains continued pre-training on verbalized transactions, supervised fine-tuning for decision formats, and GRPO with verifiable rewards for evidence-based calibration. On held-out classic-B (n=5,091) and B-hard v4 (n=2,490) tasks the model beats GPT-5.5 on average (especially hard-negative discrimination B2); it also shows zero-shot and fine-tuned transfer on Lazada voucher redemption (AUC 0.772 / 0.827). Ablations attribute generalization mainly to CPT, require RAG at train and inference, and credit GRPO with shifting reliance from LM priors to in-prompt evidence.","tokens_in":12532,"tokens_out":1025,"duration_ms":13339,"significance":"If the results hold under fair, controlled prompting, the work is a useful systems contribution: a single promptable model spanning purchase, basket, promotion, and cross-domain voucher tasks, with a clear CPT→SFT→GRPO recipe and an external SIGKDD-style transfer check. The Person–Environment framing and the explicit finding that retrieval placement and prompt evidence quality dominate performance are practically valuable for retail digital-twin deployments. Strengths include multi-task evaluation with hard negatives, an external voucher benchmark, and stage-wise ablations that separate format learning from behavioral generalization. The paper does not ship code or machine-checked proofs, but the experimental design is concrete enough to be reimplemented and stress-tested.","major_comments":[{"comment":"§3.6 and Figure 3 undercut the headline claim of internalized individual decision-making. Accuracy is near chance with no in-prompt evidence (v1), ~60% with similar-item names (v2), and 95.9% with one legible similarity line (v3); the text states that “the bottleneck is prompt signal, not weights” and that the model “primarily operates as an evidence-conditioned decision system rather than a latent similarity engine.” GRPO is likewise framed as evidence-over-prior calibration (§3.4, §3.8). Tables 3–4 gains—especially B2—are therefore consistent with superior packaging/reading of Shopping-DNA/RAG cues. To support “learned behavioral simulation” and “digital twin” language (Abstract; Contributions 2,5; §6), the paper needs either (i) an isolation experiment that strips explicit similarity/affinity lines while holding the same retrieved candidates and profile text, or (ii) a clear reframing","section":null},{"comment":"Fairness of the GPT-5.5 baseline is not fully specified for Tables 3–4 and Figure 4. The manuscript must state, for each task, whether GPT-5.5 receives identical Shopping-DNA profiles, the same top-k RAG SKUs, and the same decision templates as LBM. If the frontier model is under-prompted relative to LBM, the +9.5pp / +15.7pp averages cannot be attributed to behavioral knowledge learned via CPT/SFT/GRPO. A short protocol paragraph (or appendix table) listing prompt contents per model is load-bearing for the central comparison claim.","section":null},{"comment":"Statistical reporting is insufficient for the main claims. Tables 3–4 and the Lazada AUC results report point estimates only (n=5,091; n=2,490; 12,480 cases) with no confidence intervals, bootstrap, or significance tests, and no multi-seed variance for LoRA/GRPO training. Given that B2 drives much of the average gap and that B4 is near-tied or weak (52.0 vs 51.6 on B-hard), error bars and a simple paired test (or customer-level bootstrap) are needed before claiming consistent outperformance and transfer.","section":null},{"comment":"Scale and personalization granularity limit the digital-twin claim. Primary data cover 1,500 customers (Table 1); adapters are segment-level, not user-level (§3.2), and the D2 trip-discrimination gap vs GPT-5.5 (68.9% vs 81.2%) is attributed to that choice. Claims of scaling “efficiently to millions of users” (Introduction) and of a faithful individual twin are not yet supported by the reported cohort or adapter design. Either qualify the claim to segment-conditioned simulation or add a user-level adapter scaling study on a denser subset.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing to know: this is a clean CPT→SFT→GRPO recipe that turns verbalized retail transactions plus Shopping-DNA and RAG into a single promptable model that beats GPT-5.5 on held-out purchase/hard-negative/basket/promotion tasks and transfers to Lazada voucher AUC (0.772 zero-shot, 0.827 fine-tuned). That external check and the ablations (CPT carries generalization; retrieval only helps when present at train+infer; GRPO pushes evidence over LM priors) are the real contribution.\n\nWhat is new is not Lewin, LoRA, or RAG—those are off the shelf—but the unified verbalized-transaction pipeline with segment adapters, hard-negative retail tasks, and a reproducible four-stage chain on one backbone. They do this well: tables are readable, the Person–Environment split is operationally clear, and they are honest that SFT mostly teaches format while CPT does the heavy lifting. Citation pattern is current and fair (Adobe LCBM, Shop-R1, Customer-R1, Twin-2K-500, HSTU/TIGER).\n\nThe soft spot is exactly the one the stress-test flags, and the paper itself supplies the evidence. Figure 3 / §3.6: near chance with no in-prompt evidence, ~60% with similar-item names, 95.9% with one legible similarity line; they write “the bottleneck is prompt signal, not weights” and call the model an “evidence-conditioned decision system.” So the big B2 gains and the “digital twin / learned decision-making” language over-reach relative to superior packaging and reading of explicit cues that GPT-5.5 may not get in the same form. They never strip the similarity/affinity lines while holding candidates fixed. Secondary limits: proprietary primary data, 1.5k customers, no error bars, thin classical baselines, and some cross-market transfers (Shopee, Tmall) that fail when signal lives in IDs.\n\nNone of that makes the results fake. It makes the claim “behavioral knowledge is learned” stronger as “domain-adapted evidence-conditioned prompting” than as a latent customer simulator. For people building retail agents or simulation stacks this is useful and citeable; for pure recsys theory it is incremental. I would send it to peer review—serious referees will force the claim-tightening and the isolation ablation—and I would bring it to reading group if the group cares about LLM personalization or digital twins.","headline":"Solid applied systems paper with real transfer numbers; the soft spot is that their own B-hard ladder shows prompt evidence, not internalized twins, drives most of the accuracy.","tokens_in":13208,"tokens_out":608,"would_cite":true,"duration_ms":5606,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Language models grounded in real purchase histories can simulate individual retail customers better than frontier general-purpose models.","keywords":["customer behavior modeling","large language models","digital twins","retrieval-augmented generation","continued pre-training","retail transactions","Person-Environment formulation","behavior simulation"],"falsifier":"Hold the Shopping-DNA text and retrieved product lines fixed, then compare the trained LBM against an untrained base model of the same size on the B-hard ladder: if accuracy stays near the trained model’s whenever a legible similarity cue is present, the claim that behavioral knowledge was learned into the weights fails.","tokens_in":13006,"feed_emoji":"🛒","tokens_out":939,"duration_ms":14315,"temperature":0.7,"pith_summary":"This paper argues that customer decision-making can be learned by language models when they are trained on verbalized retail transactions and conditioned on a clear split between the person and the environment. The person is a Shopping-DNA profile built from past purchases; the environment is product context supplied by retrieval. A single shared model then handles purchase prediction, hard-negative discrimination, basket completion, promotion response, and even cross-domain voucher redemption. Ablations credit continued pre-training with most of the behavioral generalization, retrieval with gains only when used in both training and inference, and reinforcement learning with teaching the model to trust explicit evidence over generic priors. A sympathetic reader would care because retailers already hold vast transaction logs yet still rely on surveys and task-specific systems; if the claim holds, one promptable digital twin could answer many decision questions and transfer across retailers.","feed_headline":"Transaction-trained models beat frontier LLMs at retail decisions","feed_subtitle":"One Person–Environment model transfers from grocery purchases to e-commerce vouchers.","key_machinery":"The Person–Environment formulation B = f(P, E). Person is a Shopping-DNA behavioral profile (plus segment-level LoRA adapters) derived from historical purchases; Environment is product context injected by retrieval-augmented generation. A four-stage pipeline—continued pre-training on verbalized behavior, supervised fine-tuning for decision formats, and GRPO reinforcement learning with verifiable rewards—turns a general language model into a promptable customer simulator.","core_discovery":"Behavioral knowledge encoded in longitudinal transaction histories can be learned by language models. Through a Person–Environment formulation—persistent customer profiles from purchase history plus retrieval-augmented product context—a single Large Behavior Model consistently outperforms frontier general-purpose language models on in-domain retail decision tasks and shows strong zero-shot and fine-tuned transfer across retailers and decision domains, with continued pre-training as the primary driver of that generalization.","pith_inferences":["If the B-hard results are right and accuracy tracks legible in-prompt similarity more than internalized taste, real deployments will depend on continuous high-quality profile writing and retrieval, not only on larger weights.","The same Person–Environment recipe may apply outside retail wherever longitudinal event logs can be verbalized—healthcare choices, media consumption, or financial products.","Replacing handcrafted Shopping-DNA with learned continuous prompt embeddings, as the authors themselves propose, is the clean test of whether the gains survive without human-written summaries.","Segment-level adapters may hit a ceiling on fine-grained discrimination tasks that need true per-user sequence memory rather than persona averages."],"forward_implications":["One shared language model can simulate purchase, basket, promotion, and survey-style decisions without separate task-specific architectures.","Behavioral representations learned from grocery transactions can transfer zero-shot to e-commerce voucher ranking and improve further with light fine-tuning.","Continued pre-training on verbalized transactions, not model size alone, is the main route to behavioral generalization.","Retailers can build scalable customer digital twins from existing transaction logs rather than surveys or many separate recommenders.","Simulation fidelity is limited more by the quality of in-prompt behavioral evidence than by the size of the underlying model."],"fun_headline_variants":["LLMs learn customer decisions from retail transaction histories","Person–Environment LBM beats frontier models on purchase tasks","Transaction-trained models transfer across retailers and domains","Continued pre-training drives retail behavioral generalization","Behavioral digital twins outperform general LLMs on retail decisions"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The measured gains come from the model having learned individual customer decision-making, not mainly from reading stronger evidence already written into the prompts.","fun_headline_variants_meta":{"raw":{"variants":["LLMs learn customer decisions from retail transaction histories","Person–Environment LBM beats frontier models on purchase tasks","Transaction-trained models transfer across retailers and domains","Continued pre-training drives retail behavioral generalization","Behavioral digital twins outperform general LLMs on retail decisions"]},"model":"grok-4.5","effort":"low","cost_usd":0.00605,"raw_usage":{"total_tokens":1597,"prompt_tokens":787,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":60500000,"prompt_tokens_details":{"text_tokens":787,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":735,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":787,"tokens_out":75,"duration_ms":5849,"temperature":1.0,"reasoning_tokens":735,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T06:46:51.687839+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Hold the Shopping-DNA text and retrieved product lines fixed, then compare the trained LBM against an untrained base model of the same size on the B-hard ladder: if accuracy stays near the trained model’s whenever a legible similarity cue is present, the claim that behavioral knowledge was learned into the weights fails.","supporting_citations":[],"review_version":2}