Paper deep dive
MACS: A Hybrid Multi-Agent Framework for Reliable Conversational E-Commerce Recommendation
Juli Huang, Hannah Clay, Sajjad Beygi, Thomas Sarda, Negin Golrezaei, Amin Saberi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 89%
Last extracted: 8/17/2026, 4:56:50 AM
Summary
The paper introduces MACS (Multi-Agent Commerce System), a hybrid framework for reliable conversational e-commerce recommendation within fixed merchant catalogs. MACS decouples LLM-based natural language interaction (shopping agent) from deterministic constraint enforcement and retrieval (merchant agent). It utilizes SQL for hard constraint filtering and a knowledge graph for structural queries, ensuring zero constraint drift and perfect brand compliance across multi-turn sessions.
Entities (10)
Relation Signals (9)
MACS → uses → Shopping Agent
confidence 95% · MACS is composed of two cooperating agents: a shopping agent... and a merchant agent
MACS → uses → Merchant Agent
confidence 95% · MACS is composed of two cooperating agents: ... a merchant agent that enforces hard constraints deterministically
Knowledge Graph → queriedby → Cypher
confidence 90% · MACS uses a knowledge graph operationally for structural queries ... via explicit Cypher queries over a knowledge graph
Merchant Agent → uses → SQL
confidence 90% · Constraint filtering, brand exclusion, and progressive relaxation all execute as SQL queries against this store.
Merchant Agent → uses → Knowledge Graph
confidence 90% · The merchant agent ... resolves complex structural queries via knowledge graph traversal.
Shopping Agent → communicatesvia → UCP
confidence 85% · The shopping agent communicates with the merchant agent via structured commerce protocols, specifically Universal Commerce Protocol (UCP)
Shopping Agent → communicatesvia → ACP
confidence 85% · Universal Commerce Protocol (UCP) ... or Agentic Commerce Protocol (ACP)
MACS → outperforms → GPT+Catalog
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Conversational recommendation for e-commerce is increasingly mediated by large language models (LLMs), yet many real-world deployments operate under a stricter requirement: recommendations must be drawn only from a merchant's fixed catalog, without web search or unsupported product claims. In this setting, the main challenge is reliability under hard constraints: the system must satisfy user requirements, remain grounded in available inventory, and preserve preferences across multiple conversational turns. We present MACS (Multi-Agent Commerce System), a hybrid multi-agent framework for reliable conversational recommendation in fixed-catalog settings. MACS uses LLMs for language-facing tasks such as interpreting user requests, eliciting preferences, and generating responses, while correctness-critical operations, including product retrieval, hard-constraint filtering, brand exclusion, and progressive relaxation, are executed deterministically by the merchant agent. A session-persistent preference layer tracks constraints across turns, enabling consistent handling of budget overwrites and exclusion reversals. On a 140-query single-turn benchmark, MACS achieves the highest pass rate (87.1%) and perfect brand compliance (1.000). On a 10-scenario multi-turn benchmark, MACS achieves the strongest macro Pass@5 (72% vs. 56% GPT+Catalog / 52% Gemini+Catalog) with zero constraint drift. The advantage is sharpest on exclusion reversal (100% vs. 20% / 0%) and constraint accumulation (100% vs. 60% / 40%). Mean judged response quality is similar across systems (0.751 vs. 0.736). These results suggest that hybrid architectures combining deterministic constraint enforcement with session-persistent preference tracking provide stronger reliability-oriented performance than catalog-bound prompt-only baselines in the fixed-catalog merchant setting.
Tags
Links
- Source: https://arxiv.org/abs/2608.14068v1
- Canonical: https://arxiv.org/abs/2608.14068v1
Trouble viewing inline? Open PDF directly →
Full Text
55,299 characters extracted from source content.
Expand or collapse full text
MACS: A Hybrid Multi-Agent Framework for Reliable Conversational E-Commerce Recommendation CCS: Information systems Recommender systemsCCS: Computing methodologies Question answeringCCS: Computing methodologies Multi-agent systemsCCS: Applied computing Electronic commerce Juli Huang Affiliation: Stanford University , Department of Computer Science , Stanford, CA, USA email: julih@stanford.edu , Hannah Clay Affiliation: Stanford University , Department of Computer Science , Stanford, CA, USA email: hclay116@stanford.edu , Thomas Sarda Note: Now at Google DeepMind. Affiliation: Stanford University , Department of Computer Science , Stanford, CA, USA email: tsarda@stanford.edu , Sajjad Beygi Affiliation: University of Southern California , Los Angeles, CA, USA Affiliation: Amazon , USA email: beygi.e@gmail.com , Negin Golrezaei Affiliation: Massachusetts Institute of Technology , MIT Sloan School of Management , Cambridge, MA, USA email: golrezae@mit.edu and Amin Saberi Affiliation: Stanford University , Department of Management Science & Engineering , Stanford, CA, USA email: saberi@stanford.edu Abstract. Conversational recommendation for e-commerce is increasingly mediated by large language models (LLMs), yet many real-world deployments operate under a stricter requirement: recommendations must be drawn only from a merchant’s fixed catalog, without web search or unsupported product claims. In this setting, the main challenge is reliability under hard constraints: the system must satisfy user requirements, remain grounded in available inventory, and preserve preferences across multiple conversational turns. We present MACS (Multi-Agent Commerce System), a hybrid multi-agent framework for reliable conversational recommendation in fixed-catalog settings. MACS uses LLMs for language-facing tasks such as interpreting user requests, eliciting preferences, and generating responses, while correctness-critical operations, including product retrieval, hard-constraint filtering, brand exclusion, and progressive relaxation, are executed deterministically by the merchant agent. A session-persistent preference layer tracks constraints across turns, enabling consistent handling of budget overwrites and exclusion reversals. On a 140-query single-turn benchmark, MACS achieves the highest pass rate (87.1%) and perfect brand compliance (1.000). On a 10-scenario multi-turn benchmark, MACS achieves the strongest macro Pass@5 (72% vs. 56% GPT+Catalog / 52% Gemini+Catalog) with zero constraint drift. The advantage is sharpest on exclusion reversal (100% vs. 20% / 0%) and constraint accumulation (100% vs. 60% / 40%). Mean judged response quality is similar across systems (0.751 vs. 0.736). These results suggest that hybrid architectures combining deterministic constraint enforcement with session-persistent preference tracking provide stronger reliability-oriented performance than catalog-bound prompt-only baselines in the fixed-catalog merchant setting. Keywords: conversational recommendation, agentic commerce, multi-agent systems, preference elicitation, knowledge graph reasoning, semantic search, limited-catalog shopping, Model Context Protocol 1. Introduction Conversational recommender systems increasingly rely on large language models (LLMs) to interact with users in natural language, elicit preferences, and generate recommendations. A particularly important and underexplored deployment setting is limited-catalog shopping: a merchant agent using its own fixed product database without web search to enforce the recommendation of verified inventory while respecting user constraints. This setting gives merchants control over product data and avoids the latency and hallucination risks of web-augmented search. However, it places strict reliability demands on the recommender. LLMs, despite their fluency, remain unreliable in this setting: they may hallucinate product specifications not present in the catalog, forget earlier brand exclusions or other user preferences across conversational turns, or apply budget filters inconsistently. Therefore, a recommendation system that returns out-of-catalog products or violates a user’s stated constraints fails a core reliability requirement. MACS (Multi-Agent Commerce System) is a hybrid agent-to-agent e-commerce framework for reliable conversational recommendation in fixed-catalog settings. MACS is composed of two cooperating agents: a shopping agent that handles natural-language input, intent routing, and response generation; and a merchant agent that enforces hard constraints deterministically and retrieves products exclusively from the fixed catalog. The shopping agent communicates with the merchant agent via structured commerce protocols, specifically Universal Commerce Protocol (UCP) (Google 2026) or Agentic Commerce Protocol (ACP) (OpenAI 2025). Because the shopping agent cannot retrieve products directly, catalog grounding is enforced at the architectural level rather than through prompt instructions. Three core contributions are made as follows: • Architecture. A hybrid shopping-agent to merchant-agent architecture separates LLM-based natural-language interaction from deterministic constraint enforcement, enabling reliable product recommendation over a fixed, merchant-owned catalog without web search. • Session state. Session-persistent preference state management preserves user constraints across turns, supports updates such as budget overwrites and exclusion reversals, and enables consistent multi-turn recommendation behavior. • Evaluation framework. A hybrid evaluation methodology separates deterministic constraint correctness from response quality, enabling faithful assessment of conversational recommenders in both single-turn and multi-turn limited-catalog settings. 2. Related Work MACS sits at the intersection of three active research areas: conversational recommender systems, knowledge-graph-augmented retrieval, and LLM reliability under hard constraints. Prior work in each area motivates a component of the MACS design. However, none addresses the combination of deterministic constraint enforcement, session-persistent preference state, and fixed-catalog grounding that defines limited-catalog commerce-agent settings. Conversational Recommender Systems. CRSs elicit user preferences through multi-turn interaction (Jannach et al. 2021), ranging from bandit-based question selection (Christakopoulou et al. 2016) and RL-based dialogue policy learning (Sun and Zhang 2018) to LLM-integrated pipelines for preference understanding and response generation (Friedman et al. 2023; Feng et al. 2023; Liu et al. 2023c). Retrieval augmentation (Yang and Chen 2024), cross-session memory (Xi et al. 2024), knowledge-enhanced sequential modeling (Zou et al. 2024), and behavioral evaluation (Yang et al. 2024) have further extended the paradigm. Kostric et al. (Kostric et al. 2024) show that generating usage-oriented clarifying questions yields richer preference signals in multi-turn CRS, underscoring the need for explicit cross-turn preference-tracking mechanisms. MACS addresses a gap these works leave open: deterministic enforcement of hard user constraints (excluded brands, budget ceilings, specification minimums) across turns, with safeguards to combat LLM unreliability. Knowledge Graphs and LLM Agents for Shopping. Knowledge-aware recommenders (CKE (Zhang et al. 2016), KGCN (Wang et al. 2019b), KGAT (Wang et al. 2019a)) incorporate a graph structure into learned representations; MACS uses a knowledge graph operationally for structural queries (substitutes-(products that serve the same purpose at a similar price point), compatibility) via explicit Cypher queries over a knowledge graph, so the retrieval path can be directly inspected. By contrast, learned embeddings produce similarity scores that are opaque to inspection. Shopping-agent benchmarks (WebArena (Zhou et al. 2023), AgentBench (Liu et al. 2023b), ShoppingBench (Wang et al. 2025), τ-Bench (Yao et al. 2024)) study the shopping-agent side; MACS studies both the shopping-agent and complementary merchant-agent design that provides structured, constraint-enforcing catalog access. LLM Reliability and Preference Elicitation. LLM faithfulness under hard constraints (e.g., brand exclusions, budget ceilings, specification minimums) is a known open problem. Chain-of-thought prompting reduces but does not eliminate constraint violations in multi-turn settings (Jannach et al. 2021; Feng et al. 2023). Preference elicitation approaches such as EAR (Lei et al. 2020) and UniCRS (Wang et al. 2022) ask clarifying questions to narrow item sets before retrieval, but treat constraint enforcement as an LLM inference task rather than a deterministic predicate (Christakopoulou et al. 2016; Sun and Zhang 2018). MACS decouples these two responsibilities: preference elicitation remains an LLM task, while constraint enforcement is delegated to SQL predicates persisted across turns in a typed session-state layer. As a result, constraint satisfaction does not depend on the quality of the LLM’s generation. Prompt-based approaches cannot guarantee this, because they still rely on the model itself to enforce constraints. 3. Problem Statement and System Overview We formalize limited-catalog conversational recommendation as a constrained multi-turn retrieval problem over a fixed catalog: Problem definition. A merchant catalog C contains products with structured attributes (for example, in the electronics domain: price, brand, RAM, storage, and display size). Over a multi-turn conversation, the system must extract user constraints, retrieve Rt⊆R_t satisfying all accumulated constraints at turn t, and generate factually grounded responses. It also handles add-to-cart, checkout, and service queries (return policy, shipping, warranty) through the same session. Three reliability requirements follow from the limited-catalog setting: catalog grounding (every returned product exists at a verified price), constraint correctness (hard requirements enforced deterministically, not inferred), and cross-turn persistence (constraints active until explicitly revised, never silently dropped). System overview. MACS separates a shopping agent (LLM-based preference elicitation, intent routing, response generation) from a merchant agent (SQL constraint filtering, knowledge graph traversal, session-persistent state). The shopping agent never retrieves products directly; all calls go through the merchant agent’s structured API, so the LLM cannot hallucinate catalog contents. Figure 1 illustrates the two agents and their interaction diagram. Sections 4 and 5 describe each component. 4. The Shopping Agent On each turn, the shopping agent interprets the user’s message, determines the action to take (for example, new search, refinement, comparison, add-to-cart, checkout, or FAQ), extracts structured preference slots, and requests catalog results from the merchant agent before generating a response. All correctness-critical operations, including product retrieval, hard-constraint enforcement, exclusion of disallowed brands or attributes, and progressive relaxation of lower-priority constraints when necessary, are handled by the merchant agent (Section 5). 4.1. Pipeline: Rewriting, Routing, and Extraction Each turn passes through query rewriting, intent routing, structured extraction, and catalog-grounded execution. Query rewriting injects structured hints before any LLM call (e.g., “for my son” → [use_case: school]; “gaming Chromebook” → a ChromeOS-compatibility note). Intent routing intercepts common intents (compare, refine, add-to-cart) with deterministic keyword rules, while ambiguous requests are handled by an LLM router. Structured extraction converts the user’s message into typed slot-value pairs, such as budget, excluded brands, and required specifications, and resolves follow-up references across turns, including requests like “show me something cheaper” or “add the second one to the cart.” 4.2. Session State and Multi-Turn Management The session state is a typed slot dictionary with three key properties: • Cross-turn accumulation and ordinal resolution. Slot values persist across turns unless explicitly revised. References such as “add the second one” are resolved against the most recent recommendation context. Brand exclusions are enforced at the SQL level and in a post-retrieval title filter, so no excluded brand appears regardless of database condition labels. • Preference pivot detection. Updated constraints override earlier ones when appropriate: a new budget ceiling replaces the previous value, and a statement such as “actually HP is fine” reverses an earlier brand exclusion. • Underspecification handling. When the query contains no hard constraint (brand exclusion, price ceiling, or specification minimum) and the message is fewer than four words, the agent asks a targeted follow-up question rather than guessing. When at least one hard constraint is present with a sufficiently specific message, it proceeds directly to search. The recommendation output presents products that satisfy the current constraints, explains why they were selected, and provides follow-up options for refinement, comparison, and further catalog browsing. 5. The Merchant Agent The merchant agent enforces hard constraints deterministically (SQL filtering, brand exclusion, progressive relaxation) and resolves complex structural queries via knowledge graph traversal. Operating exclusively over the merchant’s catalog, it guarantees that every product exists at a verified price. Flow 1: Search — “looking for a gaming laptop”User“gaming laptop”AI Shopping AgentParse requestGenerate UCP requestMerchant AgentParse UCP requestUse KG → itemsPrices, shippingCreate UCPresponseAI Shopping AgentReceive responsePresent to userUserSees resultsFlow 2: Add to Cart — “add this to cart”UserSelect productAdd to cartAI Shopping AgentParse queryGenerate UCP requestMerchant AgentParse UCP requestAdd to cartValidate inventoryCreate UCPresponseAI Shopping AgentReceive responsePresent to userUserCart updated Figure 1. MACS interaction flows. The AI shopping agent handles natural-language interaction and protocol generation; the merchant agent enforces hard constraints, validates inventory, and returns structured responses. Flow 1 (Search): the user’s query is parsed, constraint-filtered via the knowledge graph and SQL store, and results are returned. Flow 2 (Add to Cart): the merchant agent validates inventory before confirming. All constraint enforcement occurs exclusively in the merchant agent.Two interaction flow diagrams. Flow 1 (Search): User to AI Shopping Agent (parse request, generate UCP) to Merchant Agent (parse UCP, use KG, prices) to Create UCP response to AI Shopping Agent (present to user) to User (sees results). Flow 2 (Add to Cart): User to AI Shopping Agent to Merchant Agent (add to cart, validate inventory) to Create UCP response to AI Shopping Agent to User (cart updated). 5.1. Protocol and Data Layer MACS exposes three protocol families through a shared backend interface, so adding a new protocol requires only a request adapter and response formatter. Model Context Protocol (MCP) (Anthropic 2024) provides four typed tool calls (search_products, get_product, add_to_cart, checkout) with observable response envelopes. Universal Commerce Protocol (UCP) (Google 2026) follows Google’s schema for Gemini-based agents. Agent Commerce Protocol (ACP) (OpenAI 2025) adds a checkout session lifecycle and webhooks. A configuration variable selects UCP or ACP at deployment time. 5.2. Data Layer Three data stores serve distinct roles. The relational product store holds the authoritative product catalog (21,000+ items; 1,490+ laptops with 12 structured specification fields per product: CPU, RAM, GPU, storage type, display size, battery life, weight, OS, price, brand, review count, and rating). Constraint filtering, brand exclusion, and progressive relaxation all execute as SQL queries against this store. The knowledge graph encodes product relationships across 2,400 nodes and 8,500 edges in two actively-queried relationship types: Similar_To (same use-case substitutes at a similar price) and Compatible_With (accessory compatibility). Complex intent queries (“show me similar items,” “find a compatible accessory”) are resolved via Cypher traversal in ∼ 17 ms with no LLM call. The graph degrades gracefully to SQL-based search when unavailable, ensuring no hard dependency on the graph layer. The cache and session layer caches search results, product summaries, price data, and LLM-generated narratives using per-entry Time-To-Live (TTL) expirations (30 secs-30 min depending on data freshness requirements). This caching layer achieves ∼ 12× latency reduction on cache hits (36 ms vs. 446 ms cold read, measured over n=10n=10 runs). 5.3. Hard Constraint Filtering Extracted slot values map directly to SQL WHERE clauses. Brand exclusions are enforced both at the SQL level and in a post-retrieval title filter. This dual enforcement is necessary because some entries in our catalog store condition metadata (e.g., “New”, “Recertified”) in the brand field rather than the manufacturer name. 5.4. Progressive Relaxation When hard constraints produce fewer than three results, the system applies a sequence of relaxation steps ordered by constraint importance. At each step, the least-critical optional specification constraint is dropped first. Price ceilings and brand exclusions, however, are never relaxed. The user always sees real, purchasable products; the system never returns an empty result set or fabricated alternatives. When relaxation occurs, the response explicitly discloses which constraints were loosened, so the user is not misled about what was actually enforced. Figure 2 shows the comparison and checkout view, in which side-by-side pros/cons and delivery options are resolved from the relational store and presented without LLM-generated fabrication. Figure 2. MACS comparison and checkout view. Pros/cons are derived from structured catalog attributes, and delivery options and pricing are retrieved directly from the merchant agent’s relational store, not generated by the LLM.Side-by-side product comparison showing Lenovo IdeaPad vs.\ ThinkBook with pros/cons bullet points, delivery options (Standard, Express, Overnight), and itemized pricing breakdown. 5.5. Best-Value Scoring The best-value pick uses a weighted scoring function tuned to the detected use case: (1) score(p)=wprice⋅rp+wrating⋅r^+min(n/200,cvol)+wspec⋅ϕ(p)score(p)=w_price· r_p+w_rating· r+ \! (n/200,\;c_vol )+w_spec·φ(p) where p denotes a product; rp∈[0,1]r_p∈[0,1] is the normalized price score computed as 1−(pricep−pricemin)/(pricemax−pricemin)1-( price_p- price_ )/( price_ - price_ ), so lower-priced products receive higher scores within the retrieved set; r r is the star rating; n is review count; cvolc_vol is a use-case-specific cap on the review-volume confidence boost (so that a single viral product cannot dominate by review count alone); and ϕ(p)∈[0,1]φ(p)∈[0,1] is a use-case specification score: a deterministic weighted sum over structured specification fields (e.g., GPU tier for gaming, RAM tier for ML/video editing, battery life for student use). Unlike a learned similarity score, ϕ(p)φ(p) is computed directly from catalog attributes. Table 1 reports the full weight vectors. Table 1. Best-value scoring weights and ϕ(p)φ(p) spec-score components per use case (cvol=0.05c_vol=0.05 for all). Use case wpricew_price wratingw_rating wspecw_spec ϕ(p)φ(p) components Gaming 0.15 0.20 0.60 GPU tier (0–0.40) + refresh rate (0–0.15) + RAM ≥ 16 GB (+0.03) ML 0.15 0.20 0.60 RAM tier (0–0.40) + GPU (0–0.20) Video edit. 0.15 0.20 0.60 GPU tier (0–0.30) + RAM tier (0–0.20) Student 0.45 0.30 0.20 Battery tier (0–0.30) + RAM tier (0–0.15) Programming 0.20 0.30 0.45 RAM tier only (0–0.20) Default 0.35 0.35 0.20 RAM tier only (0–0.20) 6. Evaluation This section describes the evaluation framework and reports all results. No human participants were involved in benchmark construction or scoring; all evaluations use automated deterministic checks and LLM-as-judge scoring. First, Section 6.1 defines single-turn evaluation and its associated metrics. Section 6.2 defines multi-turn evaluation and scenario design. Next, Section 6.3 describes the catalog-bound baselines and justifies their selection. Section 6.4 reports all numerical findings. 6.1. Single-Turn Evaluation A single-turn interaction is defined as follows: the user submits a natural-language request, the system may ask at most one clarifying question, and the system then returns a catalog-grounded recommendation. This protocol evaluates recommendation quality in isolation from multi-turn preference accumulation, which is assessed in Section 6.2. The evaluation separates constraint correctness (enforcing price, brand, and specification requirements) from response quality (helpful, well-explained recommendations). Hard constraints are evaluated deterministically via SQL verification. For response quality, G-Eval (Liu et al. 2023a) is applied: an LLM-as-judge evaluation framework that scores free-text responses against a rubric without requiring gold-standard reference answers. G-Eval is restricted to quality assessment to avoid circularity with the constraint-checking component. Deterministic ground truth (hard constraints). For each benchmark query, expected filter values and exclusions are derived by executing the constraints directly against the relational product store as SQL predicates (e.g., price <= 1000 AND brand NOT ILIKE ’%HP%’). This SQL ground truth is independent of any LLM: violations are verifiable, objective failures. LLM-as-judge quality scoring. For response narrative quality, G-Eval is applied with a custom rubric evaluating engagement (0–4), tone calibration (0–3), and factual accuracy/coherence (0–3), normalized to [0,1][0,1]. The judge is GPT-4o-mini at temperature 0 for reproducibility. It evaluates whether explanations are factually grounded in the returned products, not whether they resemble a reference answer. Calibration examples anchor the three discrete score levels: 0.0 (poor or irrelevant response), 0.5 (partially satisfactory response), and 1.0 (excellent response that fully addresses the request); higher values are strictly better. Composite scoring formula. The final per-query score combines deterministic and quality components: (2) s=0.40stype+0.20sbrand+0.10sfilter+0.05sstock+0.10sexplain+0.15squalitys=0.40\,s_type+0.20\,s_brand+0.10\,s_filter\\ +0.05\,s_stock+0.10\,s_explain+0.15\,s_quality where stype∈0,1s_type∈\0,1\ is a binary response-type match (1 if the system produces a recommendation when one is expected, or a clarifying question when the query is underspecified; 0 otherwise); sbrand∈[0,1]s_brand∈[0,1] is the fraction of returned products respecting all stated brand exclusions; sfilter∈[0,1]s_filter∈[0,1] is the fraction of returned products satisfying price and specification constraints; sstock∈[0,1]s_stock∈[0,1] is the fraction of returned products that are currently available in inventory (i.e., not marked out-of-stock or discontinued in the catalog at query time); sexplain∈[0,1]s_explain∈[0,1] measures response explainability deterministically via three equal-weight sub-checks: whether the message contains structured explanation elements (bullet lists or explanatory connectives such as because or since), at least one specific product attribute (e.g., RAM, battery life, price, or storage), and an explicit disclosure of applied constraints (e.g., “under $X,” “at least,” or “matching your”); each sub-check contributes equally, so sexplain=checks_passed/3s_explain=checks\_passed/3; and squality∈[0,1]s_quality∈[0,1] is the G-Eval narrative quality score. Weights for components inapplicable to a given query (e.g., sbrands_brand when no brand exclusion is present, or sexplains_explain for clarification-only responses) are redistributed to squalitys_quality, raising its effective weight to 45–60%. This redistribution preserves the total weight budget and avoids penalizing queries that do not exercise a particular constraint type (a query with no brand exclusion receives no fixed zero for brand compliance). For text-only baselines (GPT, Gemini), only stypes_type and squalitys_quality contribute to the composite score. Brand compliance is evaluated separately via a negation-window regex over free-text responses and is reported in Table 3; it is not incorporated into the composite. To note, filter compliance is marked N/A for text-only systems, as price verification from free text is unreliable. Query construction. Queries were authored by the research team to cover five user intent classes: direct purchase (e.g., “gaming laptop under $1,000 with at least 16 GB RAM, no HP”), multi-constraint filtering, vague preference (e.g., “I need something for school”), service inquiry (e.g., “What is your return policy on refurbished laptops?”), and product comparison. Each query is assigned a buyer persona (student, professional, gamer, or traveler) and classified as either specified (containing ≥ 1 extractable hard constraint: brand exclusion, price ceiling, or minimum specification) or underspecified (expressing intent without measurable constraints). Query construction details are provided in Appendix B. Benchmark scale. Two nested sets are used. The shared 140-query benchmark (85 specified / 55 underspecified) supports a three-way comparison across MACS, GPT+Catalog, and Gemini+Catalog. The expanded 225-query benchmark (138 specified / 87 underspecified) covers 40 query groups, including catalog exploration, follow-up service questions, post-recommendation refinements, and orchestrator routing; evaluated on MACS only (baselines run on 140-query subset). Pass threshold: score ≥0.5≥ 0.5. Stochastic robustness. The G-Eval judge introduces ≈ 0.02–0.03 score variance between identical runs at temperature 0. Pass@k (Chen et al. 2021), defined as the fraction of k independent evaluation runs that pass, provides a robustness measure for this variance. Pass@5 is measured on the full 225-query benchmark using a pre-fix code snapshot: MACS achieves a macro Pass@5 of 0.7420.742, with mean score 0.609±0.0100.609± 0.010 (SD) across five independent runs (per-run averages: 0.622,0.614,0.611,0.603,0.5960.622,0.614,0.611,0.603,0.596; pass rate 74.2%±4.274.2\%± 4.2p). The ±0.010± 0.010 Standard Deviation (SD) confirms that single-run results lie within one SD of the multi-run mean; the same variance bound applies to the improved result of 0.6810.681 (Table 4) from the final code version. 6.2. Multi-Turn Evaluation Single-turn evaluations cannot assess the properties that distinguish agentic systems from stateless LLM calls: constraint persistence, preference updates, and brand-exclusion reversals across turns. A 10-scenario benchmark is designed covering three fundamental slot operations: accumulation (new constraint added), update (existing constraint replaced), and reversal (prior constraint removed). Scenarios span six constraint-persistence patterns: constraint accumulation, use-case pivot (gaming → video editing), brand-exclusion persistence, vague-to-specific elicitation, budget overwrite, and exclusion reversal. Additionally, one 5-turn long session and one dense simultaneous-enforcement first turn are included; each scenario contains 3–5 scripted turns. Evaluating at K=5 runs per scenario yields 50 data points per system for variance-aware macro Pass@5 comparison. Table 6 lists all ten scenarios by category. Scoring. The multi-turn score combines a judge score with deterministic constraint checks: (3) sMT=0.55sconstraint+0.45sjudges_MT=0.55\,s_constraint+0.45\,s_judge sconstraints_constraint is computed from deterministic checks against the system’s final-turn product list: brand exclusions present/absent, budget ceiling respected. Importantly, constraint checks are computed before the judge is called, and their pass/fail results are included in the judge prompt. This prevents the judge from incorrectly marking a response as violating constraints that were enforced during retrieval but not explicitly mentioned in the text. sjudges_judge is GPT-4o-mini (temperature 0) evaluating the full conversation transcript on constraint satisfaction (0–4), preference tracking accuracy (0–3), and response appropriateness (0–3). Pass threshold: ≥0.65≥ 0.65 (raised from 0.50 to eliminate floor effects: a system with sconstraint=1.0s_constraint=1.0 and sjudge=0s_judge=0 would auto-pass at threshold 0.50 since 0.55×1+0.45×0=0.55≥0.500.55× 1+0.45× 0=0.55≥ 0.50). Mean score is the average of sMTs_MT across all K independent runs for each scenario, then averaged across scenarios. Macro Pass@5 is the mean over scenarios of the fraction of 5 runs exceeding the threshold (i.e., the fraction of runs that pass, averaged over all scenarios). 6.3. Baselines Two catalog-bound baselines are evaluated. GPT+Catalog uses GPT-4o-mini with MACS’s live catalog injected into the system prompt before each query; it receives the same top-K products as MACS, where K denotes the highest-ranked products returned by MACS’s catalog search for that query, but applies no SQL-layer constraint enforcement. Gemini+Catalog uses Gemini 2.5 Flash Lite under the same injection protocol. Both systems along with MACS operate on the same query set, receive the same catalog snapshot per query, use the same token budget, are prohibited from web search, and are scored by the same pipeline. This design yields a fair, controlled comparison that reduces database-access asymmetry and focuses the analysis on architectural differences in constraint enforcement. The catalog injection approach is adopted rather than vanilla LLM baselines because it controls for catalog grounding, allowing observed quality differences to reflect architectural choices rather than information asymmetry. Concretely, the baselines test constraint-aware response generation given a shared candidate pool. They are not full end-to-end retrieval systems, and the comparison is explicitly scoped to that narrower question. Both baselines share an identical five-rule system prompt instructing the LLM to recommend only from the injected catalog, respect stated constraints (budget, brand exclusions, required specs), avoid web search, and disclose when no matching product exists. Products are injected as a numbered list, with K=8K=8 (top-8) for single-turn queries and K=10K=10 for multi-turn scenarios. Generation uses temperature 0.3 and a 600-token output budget. The quality judge is GPT-4o-mini at temperature 0, identical to MACS’s scoring pipeline, so all quality scores are directly comparable. 6.4. Results Latency (single-turn path). Table 2 reports measured median latencies by pipeline phase for the single-turn recommendation path; multi-turn interactions compound per-turn latency additively across turns. LLM inference accounts for 85–92% of total latency on cache misses. The knowledge graph re-ranking overlay (207 ms median) is 11–28× faster than equivalent LLM-based narrative generation (2,256 ms median) and produces deterministic, auditable results. Consequently, structural queries (substitutes, best value) are routed through the graph rather than a prompt. The cache layer reduces effective latency ∼ 12× on repeated queries (446 ms cold vs. 36 ms cached). Table 2. Measured median latencies by pipeline phase (n=10n=10 runs per phase, direct-call benchmark; median reported to reduce sensitivity to occasional LLM cold-start outliers). LLM calls dominate total latency, and knowledge graph traversal and cache hits are negligible. Phase Median (ms) Notes CPU phases (rewrite, filter) <<1 Pure Python, no I/O Domain detection — fast path <<1 Keyword dictionary Domain detection — LLM ∼ 1,630 gpt-4o-mini; ambiguous only Criteria extraction (LLM) 1,927 Primary bottleneck Question generation (LLM) 1,645 Preference elicitation Post-rec intent detection 556 gpt-4o-mini classification Rec. narrative (LLM) 2,256 Recommendation explanation Filter refinement (LLM) 2,187 gpt-4o-mini refinement SQL search (cache miss) 446 Catalog store round-trip SQL search (cache hit) 36 Cache GET + deserialize KG re-ranking overlay 207 FAISS + graph + scoring Total (cache miss, rec.) ∼ 4,600 1,927++446++2,256; KG overlay adds ++207 ms when triggered Total (cache hit, rec.) ∼ 2,300 36++2,256; criteria extraction and SQL search bypassed Single-turn results. Table 3 reports the three-way comparison on 140 queries. MACS led all systems on pass rate at 87.1%, with perfect brand compliance (1.000) and near-perfect filter compliance (0.970). GPT+Catalog achieved 72.1% and Gemini+Catalog 68.6%. The gap reflects MACS’s SQL-enforced constraint pipeline, which text-only systems cannot replicate structurally. Remaining MACS failures were catalog-level: absent inventory, marketplace risk (users asking whether off-platform deals are legitimate, which requires seller-trust signals outside the product database), and multi-category budgets. Table 3. Three-way single-turn comparison (n=140n=140 unique queries, pass threshold ≥0.5≥ 0.5. MACS Brand/Filter: deterministic SQL-layer compliance. Baseline Brand†: regex negation-window check on free-text responses (6 exclusion queries; 1 Gemini violation on Q26 “no HP, no Acer”). Filter N/A: cannot verify price compliance from free text. Quality: LLM-judge narrative score (GPT-4o-mini, temperature 0). System Avg ↑ Pass% ↑ Brand ↑ Filter ↑ Quality ↑ MACS 0.681 87.1% 1.000 0.970 0.394 Gemini+Catalog 0.662 68.6% 0.833† N/A 0.427 GPT+Catalog 0.639 72.1% 1.000† N/A 0.392 Scope note: all evaluations use a consumer-electronics catalog (laptops and accessories). Generalization to other product domains is not claimed and represents a direction for future work. Expanded 225-query benchmark. Table 4 reports per-group pass rates on the full 225-query benchmark (MACS only; all queries evaluated independently with no prior session context). Across all 225 queries, MACS achieved avg. 0.681 (pass rate 85.3%). Gains from targeted fixes included: follow-up service QA (0%→ 100%), catalog exploration (50%→ 75%), orchestrator routing (25%→ 33%), post-rec refinement (43%→ 86%), and gaming-specific (25%→ 100%). The multi-constraint group (17 queries) achieved 94.1% pass rate, confirming that the SQL pipeline enforces four simultaneous constraints deterministically. Table 4. Per-group single-turn results on the expanded 225-query benchmark (MACS only). Per-group Pass% uses threshold ≥0.65≥ 0.65 (stricter, to identify reliably-passing groups); the "All" row uses ≥0.5≥ 0.5 for comparability with Table 3. All queries evaluated independently (no prior session context). Group N Avg ↑ Pass% ↑ Strong groups Multi-intent rigidity 5 0.775 100.0% One-liner purchase intent 6 0.833 100.0% Refinement 3 0.756 100.0% Follow-up service QA 4 0.810 100.0% Multi-constraint 17 0.791 94.1% Post-rec refinement 7 0.738 85.7% Weak groups Context-free comparison 6 0.393 0.0% Orchestrator routing 12 0.502 33.3% Preference discovery 8 0.473 25.0% All (225 queries) 225 0.681 85.3% Multi-turn results. Table 5 reports multi-turn results across all three systems (single-turn results are in Table 3). MACS led all systems on single-turn pass rate (87.1%) and achieved 100% Pass@5 on the two hardest state-management scenarios: Exclusion Reversal (S8), in which GPT+Catalog scored only 20% and Gemini+Catalog 0%, and Constraint Accumulation (S1), in which GPT+Catalog scored 60% and Gemini+Catalog 40%. On macro Pass@5, MACS led (72% vs. 56% GPT+Catalog vs. 52% Gemini+Catalog). Mean judged response-quality scores were similar across systems (MACS 0.751, GPT+Catalog 0.736, within one SD). The 16-point Pass@5 gap reflects scenario-level consistency advantages that mean scores do not capture. Table 5. Multi-turn results (MT, 10 scenarios, Pass@5, threshold ≥0.65≥ 0.65). MT scoring: 0.55×0.55×constraint (deterministic) ++ 0.45×0.45×judge (GPT-4o-mini, temp. 0). All systems catalog-bound; same token budget; no web search. Pass@5 = macro fraction of 5 runs passing threshold. Drift: fraction of run-scenario pairs with a constraint violation. MACS uniquely achieves 100% Pass@5 on the two hardest state-management scenarios (S1: Constraint Accumulation, S8: Exclusion Reversal); aggregate mean score differences lie within one SD. System MT Avg± MT Pass@5 Drift MACS 0.751± 0.095 72% 0.000 GPT+Catalog 0.736± 0.090 56% 0.000 Gemini+Catalog 0.717± 0.118 52% 0.014 Table 6 provides the multi-turn category breakdown. MACS uniquely dominated exclusion reversal (100% Pass@5 vs. 20% GPT / 0% Gemini) and constraint accumulation (100% vs. 60%/40%). Both scenarios require mid-conversation state updates via session-persisted slot dictionaries. MACS also led on dense constraint (100% vs. 80%/80%) and comparison (40% vs. 0%/0%). Budget overwrite and long session (S9) were strong across all systems. Notably, MACS brand exclusion Pass@5 (20%) falls below GPT+Catalog and Gemini+Catalog (40% each) on the brand-exclusion scenario; this is consistent with stochastic variance at K=5 on a single scenario (5 data points) and does not contradict the single-turn brand_score of 1.000, which is measured deterministically over 140 queries. Intent pivot remains hard across all systems (0%), confirming it is a structural failure mode rather than a system-specific weakness. Table 6. Multi-turn category breakdown (Pass@5 runs; avg score / Pass@5; ✓ = Pass@5 ≥60≥ 60%). Categories map 1:1 to the 10 scenarios. Category MACS GPT+Catalog Gemini+Catalog Brand Exclusion 0.640 / 20% 0.667 / 40% 0.685 / 40% Budget Overwrite ✓0.793 / 100% ✓0.739 / 80% ✓0.712 / 80% Budget Refinement ✓0.748 / 60% ✓0.802 / 80% ✓0.811 / 80% Clarification ✓0.793 / 100% ✓0.802 / 100% ✓0.730 / 100% Comparison 0.658 / 40% 0.631 / 0% 0.586 / 0% Constraint Accum. ✓0.784 / 100% ✓0.739 / 60% 0.640 / 40% Dense Constraint ✓0.721 / 100% ✓0.811 / 80% ✓0.829 / 80% Exclusion Reversal ✓0.955 / 100% 0.640 / 20% 0.595 / 0% Intent Pivot 0.640 / 0% 0.640 / 0% 0.622 / 0% Long Session ✓0.775 / 100% ✓0.892 / 100% ✓0.955 / 100% Overall 0.751 / 72% 0.736 / 56% 0.717 / 52% Reliability, grounding, and disclosure. Three pillars are evaluated deterministically (Table 7). Grounding: all systems scored 1.000, as all retrieve from MACS’s catalog layer. Drift: MACS and GPT+Catalog scored 0.000; Gemini+Catalog recorded 0.014 (one violation across the 50 run-scenario pairs). Disclosure: eight catalog-impossible queries (e.g., “RTX 4090 laptop, $150 budget”) were scored 1.0/0.5/0.0 for explicit disclosure / silent relaxation (constraints are loosened without disclosure) / violation at pass threshold 0.7. MACS scored 0.925, achieving explicit disclosure on 6 of 8 queries via its progressive relaxation pipeline. By contrast, GPT+Catalog and Gemini+Catalog both scored below 0.6, as both receive MACS’s already-relaxed product list with no mechanism to detect or disclose the constraint gap. The resulting disclosure gap (+0.338+0.338 vs. GPT+Catalog, +0.375+0.375 vs. Gemini+Catalog) is the clearest architectural differentiator between the systems. Table 7. Reliability, grounding, and disclosure evaluation (n=8n=8 catalog-impossible queries for Disclosure; n=10n=10 scenarios × K=5 runs for Drift). Grounding: fraction of returned products verifiable in the catalog at stated price. Disclosure: fraction of impossible queries acknowledged explicitly. Drift: fraction of run-scenario pairs with a constraint violation. ↑ higher is better; ↓ lower is better. System Grounding ↑ Disclosure ↑ Drift ↓ MACS 1.000 0.925 0.000 GPT+Catalog 1.000 0.588 0.000 Gemini+Catalog 1.000 0.550 0.014 Session-state ablation. To measure the contribution of the accumulated slot dictionary, a MACS-NoSession ablation is conducted in which the slot dictionary is cleared after each turn while LLM conversation history is preserved, re-evaluating all ten multi-turn scenarios at K=5. MACS-Full achieved macro Pass@5 = 72% (mean score 0.751); MACS-NoSession dropped to 52% / 0.673 (−-20 p, −-0.078). The drop concentrated in two scenarios: S1 (constraint accumulation across four turns, 100%→ 0%) and S7 (budget overwrite, 100%→ 0%). Both require constraints set in earlier turns to gate SQL predicates in later turns. Without the slot dictionary, those predicates reset to unconstrained, and the correct products were no longer filtered in. Scenarios driven by within-turn signals were unaffected: S8 (exclusion reversal, 100%), S6 (price refinement, 100%), S9 (long session, 100%), and S10 (dense first-turn constraint, 100%) remained unchanged. This ablation provides direct evidence that the session-persistent slot state is the primary driver of cross-turn constraint enforcement in MACS. SQL-enforcement ablation. To measure the contribution of deterministic SQL predicates, a MACS-NoSQL ablation is conducted (environment flag ABLATION_NO_SQL=1) that removes all hard-constraint WHERE clauses (brand, price range, OS exclusions) while keeping catalog injection and specification filtering. Over the 225-query single-turn benchmark, MACS-Full achieved avg=0.681 / pass=85.3%; MACS-NoSQL dropped to avg=0.628 / pass=81.3% (−-0.053 avg, −-4.0 p). The sharpest effect appeared on the 19 brand-constrained queries: brand_score fell from 1.000 to 0.684 (−-0.316), confirming that SQL predicates, not prompt instructions, enforce brand compliance. filter_score was largely unchanged (0.970 vs. 0.970) because some price-related filtering still occurs after retrieval, and LLM instruction-following can partially compensate for the removal of SQL-side price constraints on underspecified queries. Metric sensitivity. The multi-turn composite score uses a reliability-oriented weighting of 0.55×sconstraint+0.45×sjudge0.55× s_constraint+0.45× s_judge. Re-evaluating MACS macro Pass@5 across five evenly-spaced weight configurations yields: judge-heavy (0.60j/0.40c0.60j/0.40c): 42%; (0.55j/0.45c0.55j/0.45c): 58%; equal (0.50j/0.50c0.50j/0.50c): 72%; default (0.45j/0.55c0.45j/0.55c): 72%; constraint-heavy (0.40j/0.60c0.40j/0.60c): 96%. This weighting pattern is expected: MACS’s advantage lies in deterministic constraint enforcement and cross-turn state consistency rather than free-text fluency alone. The strongest findings are supported independently by deterministic metrics (brand compliance, filter compliance, drift, disclosure) and by controlled ablations for session state and SQL enforcement; the default weighting (0.45j/0.55c0.45j/0.55c) balances helpfulness and reliability without using a purely constraint-dominant metric. Importantly, the Pass@5 advantages on the two uniquely-dominant scenarios (S1 constraint accumulation, S8 exclusion reversal) are driven entirely by sconstraints_constraint: the MACS-NoSession ablation (Section 6.4) collapses both from 100% to 0%, independently of how sjudges_judge and sconstraints_constraint are combined. 7. Conclusion MACS is a hybrid multi-agent framework for reliable conversational recommendation in fixed-catalog settings. In our evaluated setting, MACS consistently outperforms catalog-bound prompt baselines on reliability-oriented metrics, especially in scenarios requiring persistent constraint tracking across turns and explicit disclosure of unsatisfied requirements. On macro Pass@5, MACS achieves the strongest score (72% vs. 56% GPT+Catalog / 52% Gemini+Catalog); mean judged response-quality scores are similar across systems (0.751 vs. 0.736). The reliability advantages are sharpest on the hardest scenarios: 100% Pass@5 on exclusion reversal (catalog-bound baselines: 20% and 0%), 100% Pass@5 on constraint accumulation (GPT+Catalog: 60%, Gemini+Catalog: 40%), and a disclosure score 0.338 points above the nearest baseline on catalog-impossible queries. Controlled ablations confirm the contribution of each component: removing the persistent slot dictionary drops multi-turn macro Pass@5 from 72% to 52% (constraint-accumulation and budget-overwrite scenarios collapse from 100% to 0%), and removing SQL enforcement drops brand compliance from 1.000 to 0.684 on brand-constrained single-turn queries. These ablation results reduce dependence on the judge-weighted composite when interpreting the paper’s main reliability claims. Taken together, these results suggest that in fixed-catalog conversational commerce, architectural separation between language interaction and correctness-critical enforcement is a practical and effective way to improve reliability without sacrificing response quality. Limitations. All evaluations are limited to a single consumer-electronics domain (laptops and accessories), so the results should not be interpreted as evidence of generalization to other merchant catalogs or recommendation domains. The baseline comparison uses catalog-injected generation baselines that receive MACS-prepared candidate sets; accordingly, the comparison isolates constraint-aware generation over a shared candidate pool rather than full end-to-end retrieval quality. The benchmark was authored by the research team, and the multi-turn evaluation consists of 10 scripted scenarios (50 data points at K=5), which may introduce unintentional alignment between benchmark design and system structure. Response quality is measured using an LLM judge (GPT-4o-mini) rather than human raters, and no independent human evaluation was conducted. Session-state and SQL-enforcement component ablations are reported in Section 6.4; the remaining multi-turn and cross-domain components are left for future work. Future work. Additional work includes evaluation on standardized external benchmarks such as the retail subset of τ-Bench (Yao et al. 2024), independent human validation of G-Eval scores, comparison against a tool-augmented agent with independent retrieval, expanding beyond consumer electronics, and extending component ablations. Acknowledgements. Generative AI tools were used for limited language editing in preparing this manuscript. The authors are fully responsible for all content, claims, and results. Appendix Appendix A Multi-Turn Scenario Descriptions Table 8. Multi-turn descriptions (10 scenarios, 3–5 scripted turns each). “Key challenge” identifies the conversational state-management property under test. All scenarios are recommendation-focused except S4 and S5. ID Name Turns Key challenge S1 Constraint Accumulation 4 Preserve all accumulated constraints S2 Mind-Change Pivot 3 Use-case pivot (gaming→ ) S3 Brand Excl. Persist. 4 Maintain brand exclusion across turns S4 Vague → Specific 4 Elicit & resolve underspecified query S5 Comparison + Follow-up 3 Context-free comparison S6 Price Refinement 3 Budget update (narrowing) S7 Budget Overwrite 3 Budget replacement (not accumulation) S8 Exclusion Reversal 4 Un-exclude previously excluded brand S9 Long Session 5 Constraint persistence over 5-turn session S10 Dense Constraint 4 Simultaneous multi-constraint enforcement Appendix Appendix B Query Construction Details The 225-query benchmark spans five intent classes (direct purchase, multi-constraint filtering, vague preference elicitation, service inquiries, and comparison) authored across four buyer personas, yielding 138 specified (≥ 1 hard constraint) and 87 underspecified queries across 40 interaction groups; the 140-query three-way-comparison subset follows the same distribution. References (1) Anthropic (2024) Anthropic. 2024. Model Context Protocol. https://w.anthropic.com/news/model-context-protocol. Accessed March 2026. Chen et al. (2021) Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harrison Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. 2021. Evaluating Large Language Models Trained on Code. arXiv:2107.03374 [cs.LG] https://arxiv.org/abs/2107.03374 Christakopoulou et al. (2016) Konstantina Christakopoulou, Filip Radlinski, and Katja Hofmann. 2016. Towards Conversational Recommender Systems. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, San Francisco, California, USA, 815–824. doi:10.1145/2939672.2939746 Feng et al. (2023) Yue Feng, Shuchang Liu, Zhenghai Xue, Qingpeng Cai, Lantao Hu, Peng Jiang, Kun Gai, and Fei Sun. 2023. A Large Language Model Enhanced Conversational Recommender System. arXiv:2308.06212 [cs.IR] https://arxiv.org/abs/2308.06212 Friedman et al. (2023) Luke Friedman, Sameer Ahuja, David Allen, Zhenning Tan, Hakim Sidahmed, Changbo Long, Jun Xie, Gabriel Schubiner, Ajay Patel, Harsh Lara, Brian Chu, Zexiang Chen, and Manoj Tiwari. 2023. Leveraging Large Language Models in Conversational Recommender Systems. arXiv:2305.07961 [cs.IR] https://arxiv.org/abs/2305.07961 Google (2026) Google. 2026. Universal Commerce Protocol (UCP). https://developers.google.com/merchant/ucp. Released January 2026; Accessed March 2026. Jannach et al. (2021) Dietmar Jannach, Ahtsham Manzoor, Wanling Cai, and Li Chen. 2021. A Survey on Conversational Recommender Systems. Comput. Surveys 54, 5 (2021), 1–36. doi:10.1145/3453154 Kostric et al. (2024) Ivica Kostric, Krisztian Balog, and Filip Radlinski. 2024. Generating Usage-related Questions for Preference Elicitation in Conversational Recommender Systems. ACM Transactions on Recommender Systems 2, 2, Article 12 (2024). doi:10.1145/3629981 Lei et al. (2020) Wenqiang Lei, Xiangnan He, Yisong Miao, Qingyun Wu, Richang Hong, Min-Yen Kan, and Tat-Seng Chua. 2020. Estimation-Action-Reflection: Towards Deep Interaction Between Conversational and Recommender Systems. In Proceedings of the 13th International Conference on Web Search and Data Mining. ACM, Houston, Texas, USA, 304–312. doi:10.1145/3336191.3371769 Liu et al. (2023b) Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Aohan Zeng, Zhiyuan Liu, Yuxiao Dong, and Jie Tang. 2023b. AgentBench: Evaluating LLMs as Agents. arXiv:2308.03688 [cs.AI] https://arxiv.org/abs/2308.03688 Liu et al. (2023a) Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023a. G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. ACL, Singapore, 2511–2522. https://arxiv.org/abs/2303.16634 Liu et al. (2023c) Yuanxing Liu, Weinan Zhang, Yifan Chen, Yuchi Zhang, Haopeng Bai, Fan Feng, Hengbin Cui, Yongbin Li, and Wanxiang Che. 2023c. Conversational Recommender System and Large Language Model Are Made for Each Other in E-Commerce Pre-Sales Dialogue. In Findings of the Association for Computational Linguistics: EMNLP 2023. ACL, Singapore, 9587–9605. https://arxiv.org/abs/2310.14626 OpenAI (2025) OpenAI. 2025. Agentic Commerce Protocol (ACP). https://openai.com/index/buy-it-in-chatgpt/. Released September 2025; Accessed March 2026. Sun and Zhang (2018) Yueming Sun and Yi Zhang. 2018. Conversational Recommender System. In Proceedings of the 41st International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, Ann Arbor, Michigan, USA, 235–244. doi:10.1145/3209978.3210002 Wang et al. (2019b) Hongwei Wang, Miao Zhao, Xing Xie, Wenjie Li, and Minyi Guo. 2019b. Knowledge Graph Convolutional Networks for Recommender Systems. In Proceedings of The Web Conference 2019. ACM, San Francisco, California, USA, 3307–3313. doi:10.1145/3308558.3313417 Wang et al. (2025) Jiangyuan Wang et al. 2025. ShoppingBench: A Real-World Intent-Grounded Shopping Benchmark for LLM-Based Agents. arXiv:2508.04266 [cs.IR] https://arxiv.org/abs/2508.04266 Wang et al. (2019a) Xiang Wang, Xiangnan He, Yixin Cao, Meng Liu, and Tat-Seng Chua. 2019a. KGAT: Knowledge Graph Attention Network for Recommendation. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, Anchorage, Alaska, USA, 950–958. doi:10.1145/3292500.3330989 Wang et al. (2022) Xiaolei Wang, Kun Zhou, Ji-Rong Wen, and Wayne Xin Zhao. 2022. Towards Unified Conversational Recommender Systems via Knowledge-Enhanced Prompt Learning. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, Washington, DC, USA, 1929–1937. doi:10.1145/3534678.3539382 Xi et al. (2024) Yunjia Xi et al. 2024. MemoCRS: Memory-Enhanced Sequential Conversational Recommender Systems with Large Language Models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. ACM, Boise, Idaho, USA, 10 pages. doi:10.1145/3627673.3679599 Yang et al. (2024) Dayu Yang, Fumian Chen, and Hui Fang. 2024. Behavior Alignment: A New Perspective of Evaluating LLM-based Conversational Recommender Systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, Washington, DC, USA, 1–5. https://arxiv.org/abs/2404.11773 Yang and Chen (2024) Ting Yang and Li Chen. 2024. Unleashing the Retrieval Potential of Large Language Models in Conversational Recommender Systems. In Proceedings of the 18th ACM Conference on Recommender Systems. ACM, Bari, Italy, 1–10. doi:10.1145/3640457.3688146 Yao et al. (2024) Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. 2024. τ-Bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. arXiv:2406.12045 [cs.AI] https://arxiv.org/abs/2406.12045 Zhang et al. (2016) Fuzheng Zhang, Nicholas Jing Yuan, Defu Lian, Xing Xie, and Wei-Ying Ma. 2016. Collaborative Knowledge Base Embedding for Recommender Systems. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, San Francisco, California, USA, 353–362. doi:10.1145/2939672.2939673 Zhou et al. (2023) Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. 2023. WebArena: A Realistic Web Environment for Building Autonomous Agents. arXiv:2307.13854 [cs.AI] https://arxiv.org/abs/2307.13854 Zou et al. (2024) Jie Zou, Aixin Sun, Cheng Long, and Evangelos Kanoulas. 2024. Knowledge-Enhanced Conversational Recommendation via Transformer-Based Sequential Modeling. ACM Transactions on Information Systems 42, 6 (2024), 1–27.