IntuRec: Intuition-Guided LLM Recommender

• IntuRec is a sequential recommendation framework that anchors latent reasoning with a candidate-derived intuition embedding aligned to user preferences.
• Its two-stage architecture comprises Intuition Source Extraction (ISE) to generate candidate sets and Intuition Representation Injection (IRI) to integrate intuition into the latent state.
• Empirical evaluations show that IntuRec enhances recommendation accuracy and efficiency by aligning latent trajectories with target items, outperforming comparable models.

IntuRec is a two-stage framework for LLM-based sequential recommendation that anchors latent reasoning with recommendation intuition, a preference-aligned latent prior derived from a top-$K$ candidate set generated from user history. In the formulation introduced in "Intuition-Guided Latent Reasoning for LLM-Based Recommendation" (Liu et al., 26 Jun 2026), IntuRec addresses a specific weakness of prior latent-reasoning recommenders: they begin from an unconstrained hidden state that may be misaligned with the target item embedding manifold. The name should be distinguished from IntRec, an interactive open-vocabulary object retrieval framework rather than a recommender (Shamsolmoali et al., 19 Feb 2026). It also differs from broader intent-aware recommendation lines such as session-intent prediction, multimodal intent alignment, and reasoning-augmented large recommender models (Oh et al., 2024).

1. Definition, scope, and naming

IntuRec is defined for sequential recommendation with training instances $(u,h,y)$, where $u$ is a user, $h$ is the user’s interaction history, and $y$ is the next item the user actually interacts with. Items in both $h$ and $y$ are represented by text such as titles and descriptions, the history $h$ is converted into a prompt $x$, and an LLM serves as the backbone recommender (Liu et al., 26 Jun 2026).

The paper contrasts a naïve direct-generation recommender,

$x \xrightarrow{\text{LLM}(x)} \hat{y},$

with latent reasoning, where recommendation proceeds through a sequence of continuous latent states $(u,h,y)$0: $(u,h,y)$1 with autoregressive latent states

$(u,h,y)$2

Here, $(u,h,y)$3 denotes the last-position hidden state (Liu et al., 26 Jun 2026).

Within current arXiv usage, IntuRec most directly denotes this recommendation framework. Closely related names identify different systems: IntentRec is a hierarchical multi-task model for session-intent prediction on Netflix user engagement data (Oh et al., 2024), while IntRec is an interactive object retrieval system with an Intent State over positive and negative cues (Shamsolmoali et al., 19 Feb 2026). The overlap in nomenclature reflects a shared emphasis on intent, but the technical settings differ.

2. Problem formulation and motivation

The central claim of IntuRec is that the start point of latent reasoning, $(u,h,y)$4, is crucial because it determines which region of the embedding space the trajectory explores. Existing latent reasoning methods such as ReaRec, LARES, and LatentR$(u,h,y)$5 simply take the last hidden state produced by the LLM for the prompt $(u,h,y)$6. According to the paper, that state is not explicitly aligned with the target item embedding, can be far from the desired region of the item space, and therefore leads to suboptimal trajectories (Liu et al., 26 Jun 2026).

The paper frames this as a mismatch between the hidden reasoning representation and the item embedding manifold. In token-level or CoT-style reasoning, intermediate steps are expressed in discrete tokens and must be generated and interpreted, which the paper characterizes as slow and noisy for recommendation. By contrast, latent reasoning stays in hidden vector space and can use many fewer steps (Liu et al., 26 Jun 2026).

The motivating prior is recommendation intuition, inspired by the claim that human multi-step reasoning is guided by intuition as a latent prior. IntuRec operationalizes this by first extracting a structured candidate set of likely items and then encoding that set into a single intuition embedding $(u,h,y)$7 that replaces the original $(u,h,y)$8. The resulting intuition embedding is described as a continuous vector that encodes structured information from a user’s top-$(u,h,y)$9 candidate items, is aligned with user preference and target-item semantics, and is used as the initial latent state of the LLM’s reasoning trajectory (Liu et al., 26 Jun 2026).

A plausible implication is that IntuRec redefines latent reasoning from unconstrained exploration to constrained traversal of a preference-aligned subspace. That interpretation is consistent with the paper’s own description of intuition as a latent prior that narrows search toward promising directions.

3. Two-stage architecture

IntuRec is organized into Intuition Source Extraction (ISE) and Intuition Representation Injection (IRI) (Liu et al., 26 Jun 2026).

Stage	Main operation	Output
ISE	Train an LLM recommender and generate top-$u$0 candidates with beam search	$u$1, $u$2
IRI	Encode candidates with IDAE and inject intuition into latent reasoning	$u$3, guided reasoning states

In ISE, the LLM is trained on sequential recommendation using the next-token prediction loss

$u$4

After convergence, beam search with constrained decoding produces a high-probability candidate list

$u$5

and a negative candidate list

$u$6

where the negatives are randomly sampled items (Liu et al., 26 Jun 2026).

A key detail is Target-Aware Candidate Balancing (TACB). The issue is that the top-$u$7 list may contain the target item $u$8 quite often, creating a shortcut. TACB therefore generates top-$u$9 candidates, keeps or removes $h$0 with a calibrated retention probability, and sets

$h$1

where $h$2 is the fraction of training instances where top-$h$3 contains $h$4 and $h$5 is the fraction of validation instances where top-$h$6 contains $h$7 (Liu et al., 26 Jun 2026).

In IRI, the original latent reasoning is modified as

$h$8

The function $h$9 is implemented by the Intuition Dual-Attention Encoder (IDAE). For each candidate item $y$0, token embeddings are passed through self-attention: $y$1 and then the initial latent state attends to the candidate embeddings: $y$2 The resulting $y$3 replaces $y$4 as the actual reasoning start state (Liu et al., 26 Jun 2026).

The paper implements the backbone with Qwen2.5-1.5B. User and item texts are embedded via its standard token embeddings, and the final decoding follows the BIGRec/D$y$5 scheme, where the final hidden representation is scored against item embeddings via dot product rather than unconstrained textual decoding (Liu et al., 26 Jun 2026).

4. Objectives and reasoning dynamics

The IRI stage optimizes both next-item prediction and intuition–target alignment. The recommendation loss is

$y$6

where $y$7 (Liu et al., 26 Jun 2026).

To align the intuition embedding with the target item embedding, the paper defines a BPR-style contrastive loss. First, the target item embedding is obtained by

$y$8

Negative latent states $y$9 are built from negative candidates $h$0 using the same pipeline, and the intuition loss is

$h$1

where $h$2 is cosine similarity (Liu et al., 26 Jun 2026).

The overall IRI objective is

$h$3

Both the LLM parameters and the IDAE parameters are trained end-to-end in this stage (Liu et al., 26 Jun 2026).

The paper emphasizes that IntuRec follows the LatentR$h$4-style architecture after intuition injection, but empirically its best performance is at $h$5. This means that the main effect is in replacing the original $h$6 by $h$7, and that additional steps can introduce drift once the start point is already well aligned (Liu et al., 26 Jun 2026). The probability of a recommended item is then computed as

$h$8

5. Empirical performance, ablations, and efficiency

The empirical comparisons cover traditional ID-based sequential recommenders, LLM-based recommenders, and latent-reasoning recommenders, including Caser, GRU4Rec, SASRec, ReaRec, BIGRec, D$h$9, and LatentR$y$0-B/D (Liu et al., 26 Jun 2026).

On the CDs dataset, the reported Recall@5 values are:

• BIGRec: $y$1
• LatentR$y$2-B: $y$3
• IntuRec-B: $y$4
• D$y$5: $y$6
• LatentR$y$7-D: $y$8
• IntuRec-D: $y$9

The paper states that similar patterns hold for Recall@10 and NDCG metrics, and across Toys and Games. Additional datasets in the appendix, Instruments and Books, show that IntuRec-B > BIGRec and LatentR$h$0-B, which the paper presents as evidence of generality (Liu et al., 26 Jun 2026).

The ablation study on CDs attributes the gains to four components. First, candidate quality matters: removing TACB, replacing positive candidates with random negatives, or using low-popularity items all reduce performance. Second, IRI itself is necessary: the variant w/o IRI underperforms full IntuRec. Third, embedding-level injection is more effective than text injection. Fourth, both attention modules and the BPR alignment term matter: removing SelfAttn, CrossAttn, or $h$1 degrades performance, with the paper describing cross-attention as critical for integrating candidate items with user history (Liu et al., 26 Jun 2026).

The qualitative analysis reinforces the same mechanism. In cosine-distance visualizations, intuition embeddings $h$2 cluster much closer to target item embeddings than unconstrained LatentR$h$3 start states. In the case study of a CD-music user with a history of Beatles albums, the candidate list did not contain the ground-truth item “With the Beatles”, but IntuRec’s recommendation list included it, ranked second. The paper interprets this as evidence that the candidate list still encodes the correct semantic pattern and that the intuition embedding guides the LLM toward the target (Liu et al., 26 Jun 2026).

The efficiency profile is also explicit. For IntuRec-B, the paper reports:

• Training time per iteration: $h$4 s/it
• Training memory: $h$5 GB
• Inference time per iteration: $h$6 s/it
• Inference memory: $h$7 GB

For comparison, BIGRec uses $h$8 s/it training time and $h$9 GB training memory, while LatentR$x$0-B uses $x$1 s/it and $x$2 GB. At inference, IntuRec is described as slightly faster than LatentR$x$3 because it performs intuition aggregation only once at the first step, whereas LatentR$x$4 applies its latent reasoning module at every autoregressive step (Liu et al., 26 Jun 2026). The paper also reports an optimal region around $x$5; larger $x$6 does not help and may hurt.

6. Position within intent-aware and reasoning-aware recommendation

IntuRec belongs to a broader family of methods that make intent, state, or reasoning explicit in recommendation, but its mechanism is distinct.

IRLLRec constructs multimodal intents from interaction graphs and LLM-generated textual summaries, aligns text and interaction spaces with pairwise and translation alignment, and uses momentum distillation for interaction-text matching (Wang et al., 5 Feb 2025). IntentRec predicts session intent through a hierarchical multi-task architecture that uses action type, genre, movie/show, and time-since-release as intent proxies and then feeds the resulting intent embedding into a next-item predictor (Oh et al., 2024). InDiRec clusters sequence representations into latent intent prototypes and uses those prototypes to guide a conditional diffusion model that generates intent-aligned augmented views for contrastive learning (Qu et al., 22 Apr 2025).

A second adjacent line emphasizes reasoning more directly. R$x$7ec is a unified large recommender model that interleaves reasoning tokens and recommendation within a single autoregressive process and optimizes both via RecPO, a reinforcement learning framework using recommendation labels only (You et al., 22 May 2025). By contrast, IntuRec keeps reasoning latent rather than tokenized, and its central intervention is the preference-aligned initialization of the latent trajectory (Liu et al., 26 Jun 2026).

A third neighboring literature concerns mental-state inference in conversational recommendation. "RecToM: A Benchmark for Evaluating Machine Theory of Mind in LLM-based Conversational Recommender Systems" defines Cognitive Inference and Behavioral Prediction tasks for recommendation dialogues and uses “IntuRec” as a conceptual label for a recommender that tracks desires, intentions, and beliefs to choose appropriate next strategies (Li et al., 27 Nov 2025). This suggests a conceptual parallel—both usages emphasize a latent state aligned with user preference—but the formal algorithm named IntuRec is the latent-reasoning framework of (Liu et al., 26 Jun 2026), not a conversational ToM model.

Finally, the name must be distinguished from IntRec, which is an interactive object retrieval framework built on an Intent State over positive anchors and negative constraints for open-vocabulary object localization in cluttered scenes (Shamsolmoali et al., 19 Feb 2026). The similarity is terminological rather than architectural.

7. Limitations and future directions

The IntuRec paper states three main limitations. First, there is dependence on top-$x$8 candidates: if the base LLM recommender is poor, the induced intuition may be weak. The authors note that using SASRec-generated candidates improves over BIGRec but remains worse than using LLM-generated candidates, indicating both robustness and dependence on alignment between the candidate generator and the LLM (Liu et al., 26 Jun 2026).

Second, the experiments are conducted on relatively small Amazon categories—CDs, Toys, Games, with Instruments and Books in the appendix—rather than large-scale industrial settings (Liu et al., 26 Jun 2026). Third, the method has hyperparameter sensitivity. The paper explicitly reports that $x$9 that is too small under-represents intuition, $x \xrightarrow{\text{LLM}(x)} \hat{y},$0 that is too large introduces bias and degradation, and $x \xrightarrow{\text{LLM}(x)} \hat{y},$1 does not necessarily help once the start point is good (Liu et al., 26 Jun 2026).

The architecture also assumes a backbone that supports access to hidden states, injection of continuous latent tokens, and self-attention over item text. The implementation uses Qwen2.5-1.5B, and the paper describes porting to other LLMs as conceptually straightforward but engineering-heavy (Liu et al., 26 Jun 2026).

The future directions named in the paper are threefold: scaling IntuRec to larger-scale recommendation datasets, exploring alternative ways to generate recommendation intuition that are more efficient or more semantically rich, and combining IntuRec with other advanced reasoning techniques such as reinforcement learning or dynamic reasoning-step selection (Liu et al., 26 Jun 2026). This suggests that IntuRec is best understood not as a closed architecture, but as a template for grounding latent reasoning in candidate-derived priors.