PaperlayerView exact revision on arXiv

Source-linked AI summary

CARA: Cognitive Adaptive Recommendation Agent

Weijun Gao, Jinyang Dong, Chuanru Ren, Hengxiao Li

arXiv:2608.16919v1cs.IRcs.AI

TL;DR

Existing LLM-based recommendation agents often lack explicit modeling of how user preferences become decisions. CARA addresses this with coarse-to-fine candidate filtering and affective–rational judgment, outperforming representative baselines on most metrics across three Amazon Reviews domains.

  • Problem

    Existing LLM-based recommendation agents often rely on semantic matching and single reasoning chains without explicitly modeling the recommendation decision process.

  • Method

    CARA uses coarse-to-fine candidate filtering, affective and rational judgment, and boundary-aware KTO to structure recommendation decisions.

  • Results

    CARA outperforms representative baselines on most metrics across three Amazon Reviews domains.

  • Takeaways & Limitations

    The results support structured dual-perspective recommendation as effective under the evaluated sparse-interaction settings.

  • Takeaways & Limitations

    The evaluation is limited to three product domains, a relatively small user population, and sampled candidate sets.

Abstract

from arXiv · show

Recent advances in large language models and agent-based recommendation frameworks have introduced new opportunities for more flexible and context-aware recommendation. However, existing methods still largely rely on semantic matching, end-to-end generation, or loosely structured agent workflows, without explicitly modeling how user preferences are processed and translated into final decisions. To address this limitation, we propose CARA, a cognitively inspired recommendation framework that formulates recommendation as a structured decision-making process. The core intuition of CARA is that user decisions are jointly shaped by two complementary mechanisms: intuitive affective preference and deliberate rational evaluation. Accordingly, CARA organizes recommendation into two coordinated stages: candidate filtering, which narrows the search space based on coarse-grained preference constraints, and dual-perspective decision modeling, which captures recommendation decisions through affective and rational judgment. We further introduce a boundary-aware KTO strategy that prioritizes instructions the model can solve occasionally but not consistently, thereby increasing the density of informative preference signals. Extensive experiments on three Amazon Reviews domains show that CARA achieves the best performance on most evaluation metrics, with relative improvements of up to 10.15% over the baseline.

Introduction

CARA frames recommendation as a coarse-to-fine cognitive decision process that separates candidate filtering from affective and rational evaluation. Its boundary-aware KTO training targets informative intermediate-solvability instructions, while experiments and analyses show strong performance and improved stability on complex samples.

  • Motivation: Existing LLM-based recommendation methods often rely on semantic matching or a single reasoning chain that directly generates rankings without explicitly modeling recommendation decisions.These workflows can mix decision evidence of different natures.
  • Framework: CARA organizes recommendation as a coarse-to-fine structured decision process with grounded candidate filtering followed by affective and rational judgment.Affective judgment covers categories, features, styles, and usage scenarios, while rational judgment evaluates price, budget, quality, and utility.
  • Training: Boundary-aware KTO selects instructions with intermediate empirical solvability to concentrate optimization on informative preference signals.CARA is post-trained with SFT followed by KTO, with solvability estimated from format validity, decision correctness, and factual consistency.
  • Evaluation: CARA outperforms representative baselines on most evaluation metrics across the CDs, Office, and Beauty domains of Amazon Reviews.The introduction reports strong performance across all three domains without specifying individual metric values here.
  • Analysis: Ablation studies support candidate filtering, affective judgment, and rational judgment, while boundary-aware KTO improves judgment stability on complex samples and reduces hallucination rates.Training-stage analysis attributes basic task capability gains to SFT and further stability improvements to boundary-aware KTO.

Related Work

Related work spans classical personalized recommendation, LLM-based agents, and the application of LLMs to recommender systems. These directions develop preference modeling, agent capabilities, explicit reasoning, and two main LLM recommendation routes.

  • Classical Recommendation: Classical recommender systems model personalized preferences through user-item interactions, latent user and item representations, popularity, or textual relevance.Collaborative filtering and matrix factorization perform personalized ranking, while Pop and BM25 use popularity and textual relevance as signals.
  • LLM-Based Agents: LLM-based agents support natural-language interaction, tool use, memory, and multi-step reasoning, with capabilities enhanced by reasoning-action synergy and experiential feedback.Personalized memory is another reported mechanism for improving agent capabilities.
  • Multi-Agent and Deliberate Reasoning: Multi-agent frameworks improve problem solving through role specialization, collaborative planning, and conversational coordination.These frameworks extend agent-based approaches beyond individual agents.
  • Multi-Agent and Deliberate Reasoning: Deliberate reasoning methods use explicit intermediate reasoning, including chain-of-thought, tree-structured, and graph-structured reasoning, to improve complex-task solving.The passage identifies these methods as approaches for strengthening LLM reasoning on complex tasks.
  • LLM-Based Recommendation: LLM applications in recommender systems mainly follow direct end-to-end recommendation or LLM-enhanced traditional recommendation models.These routes expand recommendation’s technical boundaries and address problems such as data sparsity and cold start.

Methodology · Problem Definition

CARA formulates recommendation as an agent-based candidate-ranking task. Given a user’s temporal interaction history, candidate items, and item information, multiple collaborative agents generate a ranking that places the ground-truth item as highly as possible.

  • Problem Definition: CARA models recommendation as an agent-based candidate ranking task over users, items, and available item information.The formulation defines U as the user set, I as the item set, and X as available item information.
  • Problem Definition: For each user, the method represents interactions as a temporally ordered history H_t^u.The history is indexed by time and user.
  • Problem Definition: CARA considers a candidate set C_t^u that is a subset of the item set I.The candidate set is defined for each user and time step.
  • Problem Definition: Multiple collaborative agents analyze user preferences using the interaction history and candidate-item information.Their analysis produces a recommendation ranking for the user.
  • Problem Definition: The agents generate a ranking list R̂_t^u for the candidate items.The ranking is produced from the available candidate-item information and preference analysis.
  • Problem Definition: The objective is to rank the ground-truth item i+^u as highly as possible in R̂_t^u.This objective defines the target of the agent-based candidate-ranking task.

Structure Overview

CARA structures recommendation as a coarse-to-fine, dual-process decision pipeline that combines candidate filtering with affective and rational judgments. It maintains natural-language cognitive memories and uses error-driven reflection to selectively revise them during sequential interactions.

  • CARA operationalizes intuitive preference formation and deliberative evaluation as affective and rational judgment within a coarse-to-fine recommendation architecture.
  • Candidate filtering: The candidate filtering agent conservatively removes items violating high-confidence constraints, while retaining ambiguous candidates for downstream evaluation.It outputs PASS or REJECT decisions with concise rationales, and passes retained candidates to both judgment modules.
  • Cognitive memory: CARA equips each user with natural-language cognitive memory that summarizes historical preferences, separating affective preferences from rational requirements such as price and quality.Affective memory covers categories, features, and styles, while rational memory supports budget compatibility and reliability judgments.
  • Dual-perspective decision-making: Affective judgment models subjective preference, whereas rational judgment assesses objective constraints, feasibility, and practical utility using separate cognitive memories and item information.The modules produce judgment scores and confidence values, whose evidence reliability can vary across item and user states.
  • Feedback-driven memory update: Error-driven reflection attributes ranking failures to filtering or judgment stages and selectively revises the corresponding memory without updating the underlying large-model parameters.The conservative update strategy changes only memory fragments conflicting with true feedback or historical evidence, preserving stable long-term preferences.

CARA Post-Training

CARA’s post-training combines supervised fine-tuning for structured recommendation behavior with reinforcement learning through preference optimization. Its boundary-aware KTO strategy concentrates learning on unstable decision-boundary instructions to improve output stability, grounding, and task accuracy.

  • Supervised Fine-Tuning: Supervised fine-tuning establishes task understanding, format stability, and initial judgment accuracy for parseable structured recommendation decisions.SFT teaches the agent to understand recommendation context and follow the required output format.
  • Boundary-Aware KTO: Focusing KTO on unstable behaviors near the decision boundary reduces format drift, factual inconsistency, and wrong judgments while improving output stability, grounding ability, and task accuracy.This approach avoids the naive strategy’s failure to consistently reduce hallucinations across datasets.
  • Boundary-Aware KTO: Boundary-aware KTO prioritizes instructions whose empirical solvability indicates correct but unstable model behavior rather than optimizing uniformly across all instructions.Instructions with solvability that is too high offer limited benefit, while those with solvability that is too low produce noisy feedback.
  • Boundary-Aware KTO: For each retained instruction, KTO uses responses evaluated for format validity, decision correctness, and factual consistency to construct positive and negative feedback samples.Two desirable and two undesirable responses are selected for each retained instruction.
  • Boundary-Aware KTO: The boundary-aware construction increases effective preference-signal density by excluding simple mastered samples and unreliable low-solvability samples.The selected boundary set is defined as Qbd = {qn | τmin ≤ˆpn ≤τmax}.

Experiments · Experimental Setup

CARA is evaluated across three Amazon Reviews domains using chronologically organized user interactions and fixed candidate sets containing one positive and nine negative items. The study compares CARA with traditional, sequential, and LLM-based recommenders under specified training, evaluation, and implementation settings.

  • Datasets: CARA is evaluated on CDs, Office, and Beauty, using 100 users per domain with the first 10 interactions for cognitive memory evolution and the remainder for testing.The interactions are organized chronologically.
  • Evaluation Metrics: Each test interaction uses one ground-truth item and nine unseen items sampled from a fixed candidate pool shared across methods.This setup ensures all methods are evaluated on identical candidates.
  • Baseline Methods: CARA is compared with Pop, BPR, and BM25; GRU4Rec and SASRec; and LLMRank, AgentCF, PromptRec, and TaxRec.These baselines represent traditional retrieval, sequential recommendation, and LLM-based recommendation categories.
  • Baseline Methods: The Full and Sample suffixes indicate training on all users versus the same 100-user subset used for CARA.This distinguishes the training-data scale of the baseline variants.
  • Implementation Details: CARA uses Qwen3-1.7B as its recommendation agent and Qwen3-32B as both teacher and judge models.The implementation also sets the KTO boundary interval to [0.35, 0.65] and selects two desirable and two undesirable responses per retained instruction.

Overall Performance

CARA delivers consistently competitive performance across three domains, with its clearest advantages on metrics assessing entire ranking lists. Its structured decision modeling is less dependent on sparse collaborative signals and improves deeper-cutoff ranking quality.

  • Overall comparison: CARA achieves consistently competitive performance across the three domains, with particularly clear advantages on metrics evaluating the quality of entire ranking lists.The results highlight the benefit of explicitly modeling user decisions under sparse interaction settings.
  • Overall comparison: GRU4Rec and SASRec show no stable advantages because limited histories and small user populations provide insufficient collaborative signals in the Sample setting.CARA is less dependent on these signals, instead reasoning over user memory, item descriptions, and statistical information.
  • Overall comparison: BM25, AgentCF, and TaxRec generally outperform purely interaction-based sequential models by using item descriptions and language-model knowledge to mitigate data sparsity.These methods remain limited by semantic matching or holistic ranking without explicitly distinguishing factors underlying user decisions.
  • Ranking quality: CARA’s advantage is strongest in overall ranking quality rather than top-1 accuracy, with larger deeper-cutoff gains on Office and CDs.These gains suggest CARA more consistently moves the target item toward the top of the candidate list.
  • Ranking quality: On Beauty, BM25 and BPRFull remain competitive on several top-rank metrics, while CARA performs better on NDCG@5 and NDCG@10.Lexical matching and collaborative signals capture some dominant preference patterns in this domain.

Model analysis

Ablation results show that candidate filtering and both affective and rational judgments are necessary for CARA’s recommendation performance. Training-stage and KTO analyses further indicate that structured supervision and boundary-aware preference optimization improve task grounding while focusing learning on informative, unstable instructions.

  • Module ablation: All three ablations underperform full CARA across datasets and metrics, confirming contributions from candidate filtering, rational judgment, and affective judgment.Removing filtering weakens preliminary candidate-space constraints, while removing either judgment disrupts the dual-perspective decision structure.
  • Training-stage analysis: SFT provides stable gains over the base model by teaching candidate-judgment formatting and basic task logic, while subsequent KTO further improves the aligned agent.Other CARA modules and experimental settings remain unchanged, so the comparison primarily reflects the agent alignment stage.
  • Training-stage analysis: KTO reduces hallucination rates, improving both recommendation accuracy and grounding because candidate rationales provide context for subsequent affective and rational judgments.Factually inconsistent rationales can propagate erroneous evidence through the recommendation pipeline.
  • Boundary-aware KTO: Boundary-aware KTO achieves lower hallucination rates than vanilla KTO by optimizing instructions with intermediate empirical solvability rather than all candidate instructions.These boundary instructions are solvable occasionally but not consistently, reallocating optimization toward learnable yet unstable behaviors and increasing informative preference-signal density.

Conclusion

CARA formulates recommendation as a structured decision process combining candidate filtering with affective and rational judgment. Experiments across three Amazon Reviews domains show that CARA outperforms representative baselines on most metrics, while its core modules and boundary-aware KTO improve recommendation behavior.

  • Framework: CARA formulates recommendation as a structured decision process with grounded candidate filtering, affective and rational judgment, confidence-adaptive ranking, and separate memory updates.These components enable continuous preference calibration without modifying the underlying language model.
  • Framework: CARA enables continuous preference calibration through error-driven updates of separate affective and rational memories without modifying the underlying language model.
  • Experimental findings: CARA outperforms representative baselines on most metrics across three Amazon Reviews domains under sparse interaction settings.Ablation and post-training analyses verify contributions from its core modules.
  • Experimental findings: Boundary-aware KTO improves judgment stability and reduces unsupported generations.
Loading 2608.16919v1…