Skip to content

SAND: Subjective-Anticipation-Augmented News Recommendation via Dual-Pathway Distillation

Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Reasoning
Keywords: Subjective Anticipation / News Recommendation / Knowledge Distillation / Agentic Collaboration / Candidate-Aware Attention

TL;DR

Addressing the fundamental gap where conventional news recommenders rely exclusively on unobserved objective article bodies, SAND introduces the concept of an "anticipated body" driven by user-specific cognitive expectations, synthesized via an offline three-LLM collaborative system and aligned using candidate-aware attention alongside dual-pathway knowledge distillation.

Background & Motivation

In online personalized news recommender systems, the prevailing paradigm formulates news recommendation as a click prediction task: inferring user preference profiles from chronological click histories and modeling candidate articles from their objective content, such as titles, categories, and full body texts. However, when readers browse modern news feeds, they encounter only highly condensed and information-sparse titles prior to making a click decision. The full news body remains entirely unobserved at the moment of choice. To bridge this information asymmetry, users naturally form a subjective anticipation of the content they expect to read upon clicking, mentally supplementing the sparse cues of the headline with their own prior domain knowledge, personal background, and current interests. For instance, encountering the headline "Apple Unveils New Product," technology enthusiasts inherently anticipate technical details regarding processor architectures or camera sensors, whereas financial investors anticipate stock volatility forecasts and revenue projections. Traditional models that condition candidate representations solely on objective body text overlook the cognitive reality that click decisions are propelled by subjective anticipation rather than post-click factual observation.

Explicitly incorporating this "anticipated body" into news recommendation exposes two formidable technical dilemmas. On the generation side, because a user's mental expectation is a qualitative cognitive state lacking ground-truth supervision, relying on unconstrained zero-shot prompts from a single Large Language Model (LLM) introduces substantial variance, hallucinations, and stylistic instability. Concurrently, different reader demographics demand distinctly calibrated generation styles while preserving individualized topical nuances. On the representation learning side, LLM-generated text systematically exhibits lower lexical diversity and distinct syntactic distributions compared to human-authored journalism. Consequently, when a neural news encoder processes both human-written historical articles and synthetic candidate anticipations, the learned embeddings suffer from structural misalignment in the latent space, severely destabilizing downstream preference matching.

To resolve these interconnected bottlenecks, this paper proposes the SAND framework, orchestrating an end-to-end synergy between generative agent simulation and geometric representation alignment. At the generation tier, SAND deploys an offline multi-agent collaborative system consisting of a Summarizer, Generator, and Controller across clustered user cohorts, iteratively evolving discrete generation policies. At the representation tier, SAND introduces candidate-aware attention to dynamically attend to historical articles congruent with current expectations, coupled with a dual-pathway knowledge distillation mechanism that anchors student representations to a frozen teacher pre-trained on authentic human journalism. Core idea: formalize the cognitive driver of news reading as a personalized "anticipated body," generating it through an offline three-LLM collaborative system and training a candidate-aware student via dual-pathway knowledge distillation to eliminate synthetic distribution shift.

Method

Overall Architecture

The SAND architecture comprises two primary operational stages: (i) an offline Agentic Collaborative Generation module that synthesizes stylistic, user-tailored anticipated bodies for each user-candidate interaction; and (ii) an Anticipated-Body-Aware Preference Learning network that performs candidate-conditioned interest aggregation alongside dual-pathway knowledge distillation.

The pipeline operates as follows: users are first grouped into \(K\) latent clusters via K-Means clustering over representations extracted by a pre-trained news recommender. Within each cluster, the Summarizer LLM distills a compact cognitive profile from click histories, the Generator LLM synthesizes anticipated news bodies adhering to a group-specific discrete policy tuple, and the Controller LLM refines the policy iteratively based on diagnostic scoring gaps from a frozen model. During model optimization, candidate news articles are encoded from their headline and synthesized anticipated body, serving as the dynamic Query in candidate-aware attention over clicked historical articles. Finally, dual-pathway knowledge distillation enforces simultaneous cosine similarity alignment across both candidate-aware user embeddings and candidate news embeddings against a frozen teacher network.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Data<br/>User Click Histories & Candidate Headline"] --> B["Agentic Collaborative Anticipated Body Generation<br/>User Clustering + Summarizer / Generator / Controller Loop"]
    B --> C["Candidate-Aware Attention Network<br/>Candidate Anticipated Embedding as Query over Historical Clicks"]
    C --> D["Dual-Pathway Knowledge Distillation Alignment<br/>Frozen Teacher Regularizing User & News Embedding Manifolds"]
    D --> E["Click Prediction Output<br/>Dot-Product Preference Scoring & Candidate Ranking"]

Key Designs

1. Agentic Collaborative Anticipated Body Generation: Absence of Ground-Truth & Cohort Heterogeneity -> Summarizer-Generator-Controller Game with Policy Evolution -> Controlled High-Quality Expectation Synthesis Because mental anticipations lack objective ground truth, single-shot LLM prompting produces erratic text. Furthermore, generating individual policies for tens of thousands of users induces prohibitive computational overhead. SAND resolves this by clustering users into \(K\) cohorts \(U_k\) via K-Means on pre-trained user vectors, assigning each cohort a specialized collaborative system of three LLMs: - Summarizer (\(f_{\text{sum}}\)): Compresses the titles of a user's clicked articles \(\{t_d : d \in \mathcal{H}_u\}\) into a dense textual profile \(s_u = f_{\text{sum}}(\{t_d\})\), articulating primary interest domains, information-seeking behavioral patterns (e.g., focusing on fiscal implications vs. architectural engineering), and reading personas. - Generator (\(f_{\text{gen}}\)): Takes the user summary \(s_u\), candidate headline \(t_{d_c}\), and cohort policy \(\pi_k\) to produce the anticipated news body \(\tilde{b}_{u,d_c} = f_{\text{gen}}(s_u, t_{d_c}, \pi_k)\) written in authentic journalistic tone. The policy \(\pi\) is formally defined as a 5-tuple \((\pi_{\text{tone}}, \pi_{\text{abs}}, \pi_{\text{spec}}, \pi_{\text{len}}, \pi_{\text{fmt}})\) over discrete spaces, governing tone (very neutral, neutral, mildly expressive), abstraction level (high level, mixed, specific), speculation count (0, 1, or 2 hypotheses), target length (short, medium, long), and structural format (narrative vs. bullet points). - Controller (\(f_{\text{ctrl}}\)): Refines policy \(\pi_k\) across training batches. At iteration step \(t\), the controller evaluates batch \(B^{(t)}\), computing the performance gap \(\delta_{u,d_c} = \hat{p}^b(u, d_c) - \hat{p}^{\tilde{b}}(u, d_c)\) between frozen recommender scores using actual bodies versus anticipated bodies. Using the mean loss \(\Delta^{(t)}\), diagnostic success/failure exemplars \(S^{(t)}\), and persistent longitudinal memory \(M^{(t)}\), it outputs an updated policy \(\pi_k^{(t+1)}\). The policy converges offline and is frozen as \(\pi_k^*\), incurring zero policy-search overhead during online serving.

2. Candidate-Aware Attention Network: Candidate-Independent Static User Representations -> Candidate Anticipation Embedding as Attention Query -> Amplifying Historical Evidence Aligning with Current Anticipation Conventional news recommenders aggregate clicked history embeddings via additive attention with a static, learnable query vector, deriving a uniform user vector \(u_u\) regardless of the candidate being evaluated. This fails to reflect that a user's active interests dynamically pivot depending on the specific headline presented.

SAND utilizes the candidate news embedding itself—derived from the title and anticipated body \(e_{d_c} = E_{\text{news}}^s(t_{d_c}, \tilde{b}_{u,d_c})\)—as the dynamic attention Query. For historical clicked articles \(d \in \mathcal{H}_u\), representations are computed from their observed actual bodies \(e_d = E_{\text{news}}^s(t_d, b_d)\). The candidate-aware user representation \(u_{u,d_c}\) is obtained via candidate-conditioned attention pooling: $\(u_{u,d_c} = \sum_{d \in \mathcal{H}_u} \alpha_{u, d_c, d} \cdot e_d, \quad \text{where} \quad \alpha_{u, d_c, d} = \frac{\exp(e_{d_c}^\top e_d)}{\sum_{d' \in \mathcal{H}_u} \exp(e_{d_c}^\top e_{d'})}\)$ This formulation guarantees that clicked articles whose authentic historical contents resonate most strongly with the reader's current subjective anticipation \(\tilde{b}_{u,d_c}\) receive higher attention weights, formalizing the past evidentiary foundation for the current click decision.

3. Dual-Pathway Knowledge Distillation: Synthetic Lexical Diversity Gap & Manifold Drift -> Frozen Actual-Body Teacher Coupled with Dual Manifold Projections -> Preserving Coherent Cross-Modal Geometric Alignment Because LLM-generated anticipated text exhibits constrained vocabulary distributions relative to human reporting, a student encoder optimized from scratch on mixed inputs tends to contract into a restricted sub-manifold biased toward synthetic phrasing. This degrades the dot-product compatibility in candidate-aware attention.

SAND introduces a teacher network identical in architecture to the student, pre-trained to convergence on authentic news bodies and held strictly frozen. The student is trained with authentic bodies for history and anticipated bodies for candidates. To prevent representational drift, distillation constraints are applied across two distinct semantic levels: - User-Embedding Alignment (Path 1): Regularizes the sequence aggregation mechanism, penalizing angular divergence between student and teacher candidate-aware user vectors: $\(\mathcal{L}_{\text{user}}^{\text{KD}} = 1 - \cos(u_{u,d_c}^s, u_{u,d_c}^t)\)$ - News-Embedding Alignment (Path 2): Directly constrains candidate news encoders, aligning the directional geometry of the student's candidate embedding with the teacher's space: $\(\mathcal{L}_{\text{news}}^{\text{KD}} = 1 - \cos(e_{d_c}^s, e_{d_c}^t)\)$ This dual-pathway design preserves structural coupling: Path 2 regularizes the constituent news representations, while Path 1 preserves the fidelity of the attention aggregation manifold.

Loss & Training

The overall training objective combines the primary click recommendation loss with the two distillation terms. The recommendation loss \(\mathcal{L}_{\text{rec}}\) is formulated as negative log-likelihood over positive clicks and \(J=4\) uniformly sampled unclicked negative candidates. The joint loss function is: $\(\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{rec}} + \lambda_1 \mathcal{L}_{\text{user}}^{\text{KD}} + \lambda_2 \mathcal{L}_{\text{news}}^{\text{KD}}\)$ where \(\lambda_1, \lambda_2 \ge 0\) modulate distillation intensity. The model is trained using the Adam optimizer with a learning rate of \(0.001\), batch size of 32, and early stopping on the validation set. Sensitivity grid searches establish \((\lambda_1, \lambda_2) = (0.5, 0.5)\) as the optimal operating point. At inference, candidate ranking requires only a single student forward pass and vector dot-product \(\hat{p}(u, d_c) = (u_{u,d_c}^s)^\top e_{d_c}^s\).

Key Experimental Results

Main Results

Evaluations were conducted on the benchmark datasets MIND-1W (English) and Adressa-1W (Norwegian) across 11 competitive baselines spanning DL-based, GCN-based, PLM-based, and agentic LLM paradigms under a strict chronological 5:1:1 split.

Category Method MIND-1W MRR MIND-1W nDCG@5 MIND-1W HIT@1 Adressa-1W MRR Adressa-1W nDCG@5 Adressa-1W HIT@1
DL-based LSTUR (An et al., 2019) 0.5742 0.6800 0.3338 0.4753 0.6015 0.2702
DL-based NAML (Wu et al., 2019a) 0.5807 0.6845 0.3544 0.5571 0.6662 0.3314
DL-based FIM (Wang et al., 2020) 0.5348 0.6496 0.2932 0.4539 0.5854 0.2338
GCN-based CNE-SUE (Mao et al., 2021) 0.5767 0.6815 0.3490 0.5384 0.6501 0.3389
GCN-based CROWN (Ko et al., 2025) 0.5762 0.6813 0.3428 0.5294 0.6432 0.3276
GCN-based LIME (Ryu et al., 2025) 0.5561 0.6661 0.3118 0.5587 0.6665 0.3462
PLM-based NAML-BERT (Wu et al., 2021) 0.5907 0.6920 0.3673 0.4956 0.6171 0.2907
Agent-based AgentCF (Zhang et al., 2024) 0.5323 0.6479 0.2791 0.4787 0.6047 0.2376
Agent-based ToolRec (Zhao et al., 2024) 0.5205 0.6387 0.2701 0.4887 0.6137 0.2482
Agent-based MACRec (Wang et al., 2024) 0.5705 0.6768 0.3376 0.5602 0.6671 0.3691
Agent-based AFL (Cai et al., 2025) 0.5858 0.6882 0.3635 0.5495 0.6601 0.3291
Ours SAND 0.5981 0.6978 0.3790 0.5826 0.6854 0.3692
Gain (%) vs. Best Baseline +1.25% +0.84% +3.18% +4.00% +2.74% +0.03%

Ablation Study

Ablations on MIND-1W decouple the contributions of the generative multi-agent system (Idea 1) and the preference distillation learning architecture (Idea 2).

Table 1: Cumulative Ablation of Agentic Generation (Idea 1) without KD

Variant Configuration MRR nDCG@5 HIT@1 Incremental Analysis
Base (One-Shot Direct Prompting) 0.5645 0.6725 0.3276 Unconstrained baseline generation
(+) Controller Optimization 0.5675 0.6747 0.3331 Diagnostic feedback guides stylistic convergence
(+) Cohort K-Means Clustering 0.5743 0.6798 0.3420 Group-specific policies yield largest MRR leap
(+) Summarizer User Profiling 0.5789 0.6832 0.3523 Full Idea 1: +7.54% cumulative HIT@1 over Base

Table 2: Isolated Ablation of Preference Learning & Distillation (Idea 2) with Full Idea 1 Inputs

Variant Configuration MRR nDCG@5 HIT@1 Key Mechanistic Insights
Teacher (Pre-trained on Actual Bodies) 0.5805 0.6845 0.3489 Baseline ceiling of authentic content modeling
Fine-tuning (Teacher tuned on Anticipations) 0.5782 0.6828 0.3441 Pure gradient adaptation collapses semantic manifold
SAND w/o KD (No Distillation) 0.5789 0.6832 0.3523 Unregularized synthetic space loses alignment
SAND w/o CAA (No Candidate-Aware Attention) 0.5895 0.6911 0.3674 Static query cannot match dynamic anticipations
SAND w/ Path 1 (User-Side KD Only) 0.5607 0.6694 0.3215 Asymmetric distillation distorts embedding space
SAND w/ Path 2 (News-Side KD Only) 0.5714 0.6777 0.3352 Unaligned sequence aggregation degrades ranking
SAND w/ Actual body (Replaced with Real Body) 0.5875 0.6899 0.3571 Proves anticipated bodies provide +6.13% net gain
SAND (Full Configuration) 0.5981 0.6978 0.3790 Dual distillation + anticipation maximizes performance

Key Findings

  1. Anticipated Bodies Drive the Accuracy Gain: Replacing the candidate's anticipated body with its authentic human-written body while holding the candidate-aware attention and dual-pathway distillation intact (SAND w/ Actual body) causes HIT@1 to drop from 0.3790 to 0.3571. While this variant still surpasses the Teacher (0.3489), confirming the efficacy of dual-pathway KD, the remaining 6.13% margin isolates the distinct advantage of representing candidates through subjective reader anticipation rather than objective content.
  2. Structural Necessity of Dual-Pathway Distillation: Implementing conventional single-pathway distillation on either the user side (Path 1, HIT@1: 0.3215) or the news side (Path 2, HIT@1: 0.3352) underperforms the unregularized model (w/o KD, HIT@1: 0.3523). Because user vectors are direct attention-weighted compositions of news vectors, regularizing only one pathway induces structural tension and manifold distortion; dual-pathway constraints are strictly required to preserve holistic geometric coherence.
  3. Inference Latency Advantage Over Agentic Baselines: On MIND, SAND achieves an inference wall-clock time of 37.7 minutes, cutting latency by 38% relative to AFL (61.0 min) and by more than 16x relative to MACRec (611.2 min). Because the LLM is queried strictly once offline per candidate rather than recursively during online ranking, SAND marries LLM reasoning power with the high-throughput inference speed of lightweight neural recommenders.

Highlights & Insights

  • Cognitive Shift from Content Engineering to Anticipation Modeling: The paper departs from decades of content-based news recommendation by recognizing that clicks are triggered by subjective, pre-reading cognitive expectations, repositioning LLMs from expensive online rerankers into offline cognitive simulators.
  • Closed-Loop Agentic Policy Evolution: By coupling Summarizer profiling, Generator parametrization, and Controller feedback driven by frozen recommender diagnostics, the framework achieves robust, self-tuning generation control without supervised reward tuning or reinforcement learning.
  • Dual-Pathway Cross-Modal Regularization: Identifying and mitigating the lexical distribution gap between synthetic LLM text and human prose via symmetric dual-pathway distillation offers a reusable architectural blueprint for integrating LLM-generated artifacts into classical retrieval pipelines.

Limitations & Future Work

  • Cold-Start Vulnerability: The Summarizer depends entirely on historical click titles to infer background knowledge; zero-history cold-start users cannot receive personalized anticipated bodies (though mapping demographic attributes to cohort clusters offers a plausible future remedy).
  • Inherited LLM Pre-training Biases: The anticipated bodies generated by GPT-4o-mini inherently reflect the societal and cultural priors embedded in the foundation model's training corpus, requiring content moderation and bias auditing before production deployment.
  • Per-Candidate Generation Overhead: Generating full textual bodies for every user-candidate pair incurs non-trivial offline API expenses ($9.86 across test sets), highlighting the need for future exploration into lightweight distilled generators, caching, and batched semantic pooling.
  • vs. Classical Discriminative Recommenders (NAML, LSTUR, FIM): Traditional architectures focus on intricate text encoders over objective articles but remain blind to pre-click anticipation; SAND adopts their efficient CNN-based backbone while fundamentally transforming candidate input semantics.
  • vs. Graph-Enhanced Recommenders (CNE-SUE, CROWN, LIME): Graph neural networks capture multi-hop collaborative structural signals and topic lifetimes, but still operate over objective article nodes; SAND demonstrates that modeling subjective cognitive anticipation outperforms heavy graph convolutions.
  • vs. Agent-Based LLM Recommenders (MACRec, AgentCF, ToolRec): Prior agentic methods prompt LLMs directly as rerankers or conversational planners, resulting in excessive inference latency and erratic outputs; SAND restricts the LLM to offline synthesis and delegates scoring to a fast neural student.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ [Pioneers the formalization of subjective anticipated bodies, bridging human cognitive psychology and recommender systems]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Evaluated against 11 diverse baselines across two languages, accompanied by full ablation, cost analysis, and 5x5 grid sensitivity studies]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Impeccably organized narrative, mathematically rigorous formulations, and lucid visualization of agentic dynamics]
  • Value: ⭐⭐⭐⭐⭐ [Presents an actionable, low-latency paradigm for integrating generative LLMs into high-throughput production recommender pipelines]