Skip to content

Illuminating the Path: Reachability-Aware Supervision Curation for LLM-Based Recommendation

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/venzino-han/GEARec
Area: Recommender Systems
Keywords: Recommender Systems, Sequential Recommendation, Large Language Models, Context-Utilization Bias, Reachability Routing

TL;DR

GEARec overcomes context-utilization bias and severe exploration collapse in LLM-based sequential recommendation through a reachability-aware data curation pipeline—extracting dual-perspective anchors, filtering unreachable reasoning trajectories via full-pool retrieval, and aligning with dynamic fine-tuning—achieving state-of-the-art accuracy across full-item candidate pools and boosting long-tail recommendation.

Background & Motivation

Integrating Large Language Models (LLMs) into sequential recommender systems offers unprecedented potential to decipher rich semantic intentions and open-world user preferences. However, real-world user interaction sequences are inherently non-stationary and noisy, filled with spontaneous clicks, exploration detours, and transient interactions. When exposed to such extensive histories, LLMs consistently fall prey to context-utilization bias—a pervasive cognitive tendency to artificially force logical connections across all provided input context. Instead of isolating the decisive trigger behind a purchase, the model attempts to synthesize every stray historical interaction into its rationale, producing convoluted, post-hoc rationalizations that sound persuasive yet are entirely disconnected from the user's authentic shopping motivation.

In traditional point-wise recommendation formulations such as rating prediction (e.g., predicting scores from 1 to 5), this vulnerability remains concealed because the artificially constrained action space guarantees a high probability of hitting the target, enabling reinforcement learning (RL) or rejection sampling to proceed smoothly. Conversely, in realistic sequential recommendation where the model must navigate a massive global candidate pool exceeding tens of thousands of items, relying on autonomous LLM exploration triggers catastrophic reward sparsity. Compounded by context-utilization bias, models generate diffuse reasoning trajectories that almost never retrieve the true target item, causing RL exploration to collapse and starving supervised fine-tuning (SFT) of valid training trajectories. Meanwhile, naive SFT on unverified, unreachable paths introduces severe error cascades and suffers from distorted reward structures inherent in standard teacher-forcing gradients.

Addressing this data curation bottleneck requires stepping upstream of downstream query fusion or distillation architectures to confront the fundamental criterion of supervision quality: which LLM-generated reasoning trajectories are empirically trustworthy enough to serve as alignment data? The key insight is that genuine user consumption is anchored in specific decisive past interactions, and a valid reasoning trajectory must demonstrate empirical reachability by successfully retrieving the target item from the full candidate pool. Core idea: formulate an offline reachability-aware supervision curation framework (GEARec) that extracts historical anchor items via dual semantic and collaborative filtering perspectives, strictly filters out unreachable reasoning paths through full-pool retrieval gating, and optimizes the model via dynamic fine-tuning and contrastive projection, empowering compact LLMs to perform highly accurate, bias-resistant recommendations in a single inference pass.

Method

Overall Architecture

GEARec (Grounded REasoning Curation for Anchor-based Recommendation) decouples complex multi-view trajectory exploration and ground-truth verification into an offline supervision curation phase, while executing a streamlined, single-pass inference at deployment time. The offline pipeline consists of three tightly coupled stages: first, Dual-Perspective Anchor Extraction and Reasoning Generation combines target-masked semantic deduction from the LLM with collaborative filtering (SASRec) statistical signals to identify pivotal anchors and generate diverse candidate paths; second, Reachability-Aware Trajectory Routing subjects generated trajectories to full-pool dense retrieval, discarding any reasoning that fails to retrieve the target within empirical thresholds; third, Alignment via Curated Trajectories optimizes the LLM via token-level Dynamic Fine-Tuning (DFT) and projects semantic query embeddings into a shared collaborative space using InfoNCE contrastive learning. At test time, the fine-tuned LLM autonomously identifies the anchor and emits the intent description in a single pass without ground-truth guidance or multi-candidate search.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["User Interaction History<br/>Massive Global Item Pool"] --> B["Dual-Perspective Anchor Extraction<br/>Prefix-Enforced Semantic LLM + CF Fallback"]
    B --> C["Reachability-Aware Trajectory Routing<br/>Full-Pool Top-K Retrieval Gating"]
    C --> D["Smooth Optimization via DFT<br/>Token-Level Inverse Probability Reweighting"]
    D --> E["Semantic-to-Collaborative Alignment<br/>Shared Projection & InfoNCE Contrastive Loss"]
    E --> F["Deployment-Time Single-Pass Inference<br/>Autonomous Anchor Generation & Vector Search"]

Key Designs

1. Dual-Perspective Anchor Extraction: Countering Context-Utilization Bias via Constrained Grounding

To prevent the LLM from indiscriminately weaving historical noise into hallucinated causal chains, GEARec enforces that the reasoning trajectory must originate from a single, high-relevance anchor item \(a^*\) in the user's history \(\mathcal{H}_u\). Recognizing the complementary strengths of text semantics and behavioral co-occurrences, the framework establishes two distinct candidate generation channels. The primary channel, the target-masked semantic path, strictly conceals the ground-truth target \(i^+\) from the prompt. For the \(N\) most recent items in \(\mathcal{H}_u\), it applies a prefix-enforcement constraint by hard-coding the output start token as <key_item>[c_k]</key_item>, compelling the frozen LLM to reason strictly from candidate \(c_k\). A pretrained dense retriever then evaluates each candidate trajectory, selecting the anchor \(a_{\text{LLM}}\) that yields the optimal retrieval rank for \(i^+\), followed by Best-of-\(N\) sampling to construct the candidate set \(\mathcal{T}(a_{\text{LLM}})\). The auxiliary channel, the CF fallback path, extracts an anchor \(a_{\text{CF}}\) maximizing the cosine similarity in the SASRec latent space: $\(a_{\text{CF}} = \arg\max_{j \in \mathcal{H}_u} \langle e_j^{\text{CF}}, e_{i^+}^{\text{CF}} \rangle\)$ Both \(a_{\text{CF}}\) and \(i^+\) are explicitly exposed in the prompt to prompt the LLM to articulate the implicit statistical association into a structured textual narrative \(\tau_{\text{CF}}\), acting as a privileged teacher during offline curation without any inference-time leakage.

2. Reachability-Aware Trajectory Routing: Halting Error Cascades from Unreachable Hallucinations

A foundational flaw in conventional data generation is treating plausible-sounding rationales as valid training data even when they fail to retrieve the actual target item in large item pools. GEARec introduces an empirical full-pool retrieval test as an explicit gating function. Utilizing a pretrained text encoder to convert reasoning trajectory \(\tau\) into dense query embedding \(q \in \mathbb{R}^d\), the routing mechanism establishes a strict hierarchical decision policy: $\(\tau_{\text{final}} = \begin{cases} \tau_{\text{LLM}}, & \text{if } i^+ \in \text{Top-}K(q_{\text{LLM}}) \\ \tau_{\text{CF}}, & \text{else if } i^+ \in \text{Top-}M(e_{a_{\text{CF}}}^{\text{CF}}) \\ \text{Discard}, & \text{otherwise} \end{cases}\)$ The model prioritizes the autonomous semantic trajectory \(\tau_{\text{LLM}}\) whenever its query embedding \(q_{\text{LLM}}\) ranks the target \(i^+\) within the Top-\(K\) items (e.g., \(K=20\)), thereby preserving the model's intrinsic deduction. Only when the autonomous path fails does the system fall back to the CF-guided trajectory \(\tau_{\text{CF}}\), and crucially, this fallback is gated by a collaborative prerequisite verifying whether \(i^+\) falls within the Top-\(M\) nearest neighbors of the native CF anchor embedding \(e_{a_{\text{CF}}}^{\text{CF}}\) (e.g., \(M=100\)). Any instance failing both criteria is discarded, preventing unreachable paths from corrupting the final alignment dataset \(\mathcal{D}_{\text{SFT}}\).

3. Smooth Optimization via DFT: Mitigating SFT Reward Distortion without RL Overhead

Standard teacher-forcing supervised fine-tuning inherently embeds an unbalanced, skewed reward structure across token sequences, inducing premature certainty, penalizing exploration, and triggering catastrophic forgetting of pretrained general capabilities. Instead of resorting to unstable and computationally prohibitive reinforcement learning loops, GEARec adopts Dynamic Fine-Tuning (DFT). DFT acts as an importance-sampling corrective factor by dynamically reweighting token-level log-likelihood updates with the inverse of the policy probability \(p_t = \pi_\theta(y^*_t \mid y^*_{<t}, x)\) for target sequence \(y^* = \tau_{\text{final}}\): $\(\mathcal{L}_{\text{DFT}}(\theta) = \mathbb{E}_{(x, y^*) \sim \mathcal{D}_{\text{SFT}}} \left[ -\sum_{t=1}^{|y^*|} \frac{1}{\text{sg}(p_t)} \log p_t \right]\)$ where \(\text{sg}(\cdot)\) represents the stop-gradient operator. This reweighting suppresses over-confident updates on trivial tokens and amplifies gradient contributions from under-explored decision points, allowing the LLM to smoothly adapt to curated reasoning trajectories with high policy entropy and superior generalization.

4. Semantic-to-Collaborative Alignment: Bridging Textual Reasoning with Large-Scale Vector Retrieval

To bridge the modality discrepancy between the LLM's verbalized semantic space and the compact collaborative representations required for efficient candidate retrieval, GEARec establishes a shared projection framework. The curated reasoning text \(\tau_{\text{final}}\) and raw item descriptions are encoded using the same pretrained dense retriever to produce initial representations, which are subsequently mapped into a unified latent space via a shared learnable projection matrix \(W \in \mathbb{R}^{d \times d_c}\): $\(q_u = W \cdot \text{Encoder}(\tau_{\text{final}}), \quad e_i = W \cdot h_i\)$ The projection layer \(W\) is optimized independently using the InfoNCE contrastive loss over target item \(i^+\) and \(B\) in-batch negative samples \(j^-\): $\(\mathcal{L}_{\text{NCE}} = -\log \frac{\exp(\langle q_u, e_{i^+} \rangle / T)}{\exp(\langle q_u, e_{i^+} \rangle / T) + \sum_{j^-=1}^B \exp(\langle q_u, e_{j^-} \rangle / T)}\)$ where \(T\) denotes temperature. This contrastive objective guarantees that the fine-tuned LLM's generated reasoning translates directly into high-fidelity user queries capable of executing sub-50ms vector similarity searches across massive item indexes.

A Worked Example

Consider an illustrative user history from the Beauty benchmark: 1. Input Context & Noise: The user has recently purchased "Suave Shampoo", "Exfoliating Bath Towel", "Derma Roller", "Flat Top Kabuki Brush", "Vitamin C Serum", and "Detangling Brush Set". The true ground-truth next interaction is "Beauty Blender Makeup Sponges 4pc Set". Baseline RL models (like Rec-R1) generate scattered justifications attempting to connect shampoo, anti-aging derma rollers, and makeup brushes into an unwieldy aggregate; unconstrained GT baselines artificially hallucinate that derma rollers necessitate makeup sponges. 2. Dual-Perspective Selection & Routing: GEARec enforces prefix candidate generation. When conditioned on candidate "Flat Top Kabuki Brush", the model generates: "The user's recent purchases demonstrate a strong interest in makeup tools and application techniques. They previously acquired a Kabuki brush to enhance foundation application; a sponge set naturally complements this routine..." The dense retriever confirms that the target makeup sponge set falls within the Top-20 retrieved candidates under this query. Thus, the reachability condition is satisfied, and the path is routed as a verified, high-quality sample into \(\mathcal{D}_{\text{SFT}}\). 3. Deployment Inference: At test time, the fine-tuned Gemma-3-1B receives only the noisy history. In a single forward pass, it autonomously selects the Kabuki brush as the anchor and generates the focused intent query, completing the full-pool search in 44.1ms and successfully retrieving the target item at rank 1.

Loss & Training

The framework optimizes two decoupled stages: 1. Generator Fine-Tuning: Optimized on \(\mathcal{D}_{\text{SFT}}\) using DFT loss with AdamW, batch size 4, learning rate \(1\times 10^{-6}\), running for 1 epoch and capped at 3,000 steps (<30 minutes on a single NVIDIA A6000 GPU). 2. Retrieval Projection: Optimized using InfoNCE loss with batch size 128, projection dimension 1024, learning rate \(1\times 10^{-4}\), negative samples \(B=99\), running for 20 epochs.

Key Experimental Results

Main Results

All models are evaluated on four diverse benchmark datasets (Sports, Beauty, Toys, Yelp) under the strict leave-one-out protocol. To replicate authentic deployment conditions, evaluation is conducted over the entire unsampled item pool (>10,000 items) using Hit Rate (H@10) and NDCG@10 (N@10).

Paradigm Model Sports H@10 Sports N@10 Beauty H@10 Beauty N@10 Toys H@10 Toys N@10 Yelp H@10 Yelp N@10
CRM GRU4Rec 0.0204 0.0110 0.0283 0.0137 0.0176 0.0084 0.0263 0.0134
SASRec 0.0350 0.0192 0.0605 0.0318 0.0675 0.0374 0.0390 0.0194
BERT4Rec 0.0191 0.0099 0.0347 0.0170 0.0203 0.0099 0.0090 0.0045
FDSA 0.0288 0.0156 0.0407 0.0208 0.0381 0.0189 0.0544 0.0293
S3-Rec 0.0385 0.0204 0.0647 0.0327 0.0700 0.0376 0.0341 0.0168
LLM-EM AlphaRec (MLP) 0.0268 0.0135 0.0456 0.0238 0.0407 0.0221 0.0073 0.0034
AlphaRec (LGCN) 0.0328 0.0177 0.0432 0.0242 0.0179 0.0091 0.0070 0.0031
LEARN 0.0188 0.0098 0.0286 0.0136 0.0349 0.0181 0.0096 0.0042
SLMRec 0.0425 0.0209 0.0757 0.0390 0.0759 0.0399 0.0572 0.0354
DLLM2Rec 0.0292 0.0143 0.0496 0.0242 0.0589 0.0315 0.0224 0.0112
LLMEmb 0.0389 0.0205 0.0754 0.0398 0.0838 0.0458 0.0205 0.0103
LLM-SRec 0.0429 0.0171 0.0767 0.0358 0.0682 0.0327 0.0588 0.0267
LLM-GM BIGRec (SFT) 0.0110 0.0059 0.0280 0.0151 0.0425 0.0241 0.0146 0.0100
SLIM (Distill) 0.0403 0.0210 0.0664 0.0354 0.0758 0.0408 0.0558 0.0437
Rec-R1 (RL) 0.0171 0.0089 0.0523 0.0291 0.0702 0.0389 0.0185 0.0103
RecZero (RL) 0.0283 0.0154 0.0585 0.0337 0.0842 0.0478 0.0156 0.0089
Ours GEARec 0.0450 0.0247 0.0833 0.0487 0.0909 0.0520 0.0755 0.0613
Improvement vs. Best Baseline +4.90% +17.62% +8.60% +22.36% +7.96% +8.79% +28.40% +40.27%

Ablation Study

The impact of each core component, architectural backbone swap, and generator scale is analyzed across datasets:

Configuration / Variant Sports N@10 Beauty N@10 Toys N@10 Yelp N@10 Core Mechanism Analysis & Observations
GEARec (Full Model) 0.0247 0.0487 0.0520 0.0613 Integrates dual-perspective anchors, full-pool verification, and DFT alignment
w.o. LLM reasoning 0.0095 0.0182 0.0215 0.0110 Direct prediction of raw metadata; suffers sharpest decline by abandoning reasoning
w.o. anchor 0.0162 0.0315 0.0340 0.0224 Unconstrained GT reasoning; induces post-hoc rationalization and severe hallucinations
w.o. LLM anchor 0.0194 0.0388 0.0412 0.0365 Relies solely on CF anchor; misses nuanced semantic intent captured by LLMs
w.o. CF anchor 0.0215 0.0421 0.0450 0.0512 Relies solely on LLM anchor; fails on non-intuitive, global co-occurrence patterns
w.o. verification 0.0185 0.0360 0.0395 0.0310 Trains on all generated trajectories; unreachable paths contaminate alignment
Backbone: S3-Rec 0.0220 0.0450 0.0513 0.0584 Replaces SASRec with S3-Rec; gains remain robust over published baselines
Backbone: Qwen3.5-0.8B 0.0246 0.0479 0.0540 0.0560 Replaces Gemma-3-1B with Qwen3.5-0.8B; confirms model-agnostic effectiveness
Scale: Qwen3-80B (Zero-Shot) 0.0194 0.0306 0.0358 — 80B model succumbs to context-utilization bias; outperformed by curated 1B model
Scale: GPT-5.1 (Frontier API) 0.0196 0.0334 0.0354 — Frontier API fails in full-pool retrieval without reachability-curated supervision

Key Findings

  • Geometric Verification of Context Bias: Geometric latent space analysis reveals that zero-shot LLMs exhibit a Bias-Escape Score (BES) near zero (0.014 on Beauty), meaning predictions linger near the historical centroid \(h_{\text{mean}}\). In contrast, GEARec achieves a BES of 0.318 and outperforms pure RL (Rec-R1) by +14% Hit@10 in the most challenging Q4 difficulty quartile (where the target is furthest from the historical centroid).
  • Curated Quality Dominates Parameter Scale: The 1B parameter GEARec model systematically outperforms massive zero-shot models (including Gemma3-27B, Qwen3-80B, and commercial GPT-5.1) across all datasets (e.g., Toys NDCG@10 of 0.0520 vs. 0.0358 for 80B), demonstrating that reachability-grounded supervision quality is far more critical than raw model scale.
  • Superior Long-Tail Item Retrieval: By anchoring on decisive past interactions rather than averaging global noise, GEARec yields substantial accuracy gains on the bottom 80% least popular items, effectively tackling the long-tail starvation problem common in traditional collaborative filtering.
  • Rigorous Trajectory Discard Rates: Full-pool reachability filtering discards an average of 17.35% of candidate trajectories across datasets (from 11.98% on Toys to 19.51% on Sports). Among retained instances, target-masked semantic reasoning accounts for the majority (81.3% on Beauty, 69.5% on Yelp), verifying that the LLM's autonomous reasoning remains the primary driver while CF serves as a reliable safety net.

Highlights & Insights

  • Reachability Gating as a Data Curation Standard: Directly bridges generative NLP with IR retrieval metrics by establishing full-pool retrieval success as an empirical filter, preventing unreachable, hallucinated reasoning paths from corrupting the training corpus.
  • Prefix-Enforced Constrained Exploration: Utilizes hard token pre-filling (<key_item>[c_k]</key_item>) to enforce conditional generation on candidate anchors, curbing open-ended exploration collapse while systematically converting latent CF co-occurrences into explicit causal language.
  • Dynamic Fine-Tuning over RL: Replaces fragile, computationally demanding reinforcement learning policy exploration with token-level inverse probability importance sampling, eliminating exploration collapse while preserving policy entropy.
  • Practical Deployment Efficiency: Eliminates test-time search, multi-candidate generation, and CF dependencies; single-pass LLM inference combined with dense index lookup operates in 44.1ms per user, establishing practical viability for production systems.

Limitations & Future Work

  • Residual LLM Inference Latency: While 44.1ms is viable for many recommender architectures, it still introduces higher latency and compute overhead compared to pure microsecond-level dual-tower or matrix factorization models.
  • Reliance on Base CF Quality: The CF fallback anchor selection depends on pretrained SASRec embeddings; under extreme cold-start or severe data sparsity where collaborative statistics fail, the fallback mechanism can degrade.
  • Dependence on Text Metadata: The framework relies on expressive item titles and textual attributes to perform semantic reasoning, which inherently limits its direct applicability to purely numerical or opaque ID-only interaction environments.
  • vs Rec-R1 / RecZero: Prior RL methods attempt to explore reasoning trajectories autonomously using retrieval rewards, but suffer from catastrophic reward sparsity in large item pools; GEARec circumvents online exploration collapse by curating reachable trajectories offline and training via smooth DFT.
  • vs BIGRec / TALLRec: Standard SFT approaches tune models to predict raw item titles or IDs directly, bypassing reasoning and leading to over-fitted memorization; GEARec establishes structured, anchor-grounded causal paths.
  • vs SLIM / DLLM2Rec: Distillation frameworks compress LLM knowledge into compact student models, but if the teacher LLM suffers from context-utilization bias, the student inherits distorted noise; GEARec operates upstream to curate a pure, reachability-verified teacher dataset.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Formulates and geometrically quantifies context-utilization bias in recommendation, pioneering reachability-aware filtering for LLM supervision.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Strict full-pool evaluation (>10K items), comprehensive baselines across CRM/LLM-EM/LLM-GM, rigorous ablations, geometric analysis, and multi-judge LLM-as-a-judge validation.
  • Writing Quality: ⭐⭐⭐⭐⭐ Exceptional clarity, compelling narrative prose, precise problem diagnosis, and informative, self-contained figures and tables.
  • Value: ⭐⭐⭐⭐⭐ Highly actionable paradigm with open-source code and reproducible pipelines, bridging the gap between LLM reasoning and real-world retrieval.