Skip to content

๐ŸŽ Recommender Systems

๐Ÿง  NeurIPS2026 ยท 24 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿ”ฌ ICLR2026 (24) ยท ๐Ÿ’ฌ ACL2026 (22) ยท ๐Ÿงช ICML2026 (11) ยท ๐Ÿค– AAAI2026 (27) ยท ๐Ÿง  NeurIPS2025 (24) ยท ๐Ÿงช ICML2025 (17)

๐Ÿ”ฅ Top topics: Recommendation ร—15 ยท Personalized Generation ร—4 ยท LLM ร—4 ยท Alignment/RLHF ร—3

OmniSimulator: Aligning Small Language Models for Authentic Heterogeneous Behavior Modeling

OmniSimulator establishes a structured five-dimensional CoT schema and task-aware causal distillation framework that addresses the degeneration of native thinking into shallow pattern matching and alleviates hyper-activity bias, boosting Qwen3-8B's predictive accuracy by 34.33% to rival Claude-Sonnet-4.5.

Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization

PerTIDE introduces an action-on-item shared event schema and a multi-timescale state-space architecture that routes action-gated evidence into long-, short-, and episodic traces, enabling a single reusable history-update core to serve both predictive ranking and generative headline synthesis with a constant 12 KiB per-user memory cache.

AdaM-Rec: Adaptive Modality Routing for Multimodal Recommendation

AdaM-Rec tackles query-dependent modality reliance in multimodal recommendation by constructing granularity-matched pseudo-queries over historical interactions to dynamically calibrate modality routing weights and retrieval budgets via agentic self-optimization.

An Embarrassingly Simple Graph Heuristic Reveals Shortcut-Solvable Benchmarks for Sequential Recommendation

This paper introduces an intentionally simple, training-free Transition-Graph Heuristic (TGH) that outperforms modern generative and deep sequential recommenders by retrieving few-hop transition neighbors and ranking them with frozen text similarity, uncovering widespread shortcut structures across standard recommendation benchmarks.

Bridging Textual Profiles and Latent User Embeddings for Personalization

This paper introduces BLUE, a reinforcement learning framework that optimizes a generative LLM profiler by uniting embedding-space bidirectional contrastive rewards with text-space next-item prediction signals under GRPO, creating user profiles that are both transparently interpretable and highly effective for retrieval.

Controllable and Content-Based Recommendations

CCBR introduces an editable natural-language user profile bottleneck derived directly from raw multimodal content via MLLMs, plugging into frozen collaborative filtering decoders via tag-supervised pretraining and InfoNCE alignment to deliver competitive accuracy alongside causal, text-driven and multimodal steerability.

DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules

Addressing the industrial IoT bottleneck where fault detection is automated but maintenance action selection requires scarce expert knowledge, this paper introduces DiagnosticIQโ€”the first benchmark derived from 118 real-world symbolic rules across 16 equipment types comprising 6,690 MCQA itemsโ€”and exposes through condition inversion that frontier LLMs rely heavily on superficial rule pattern matching rather than genuine logical evaluation.

FoMEMO: Towards Foundation Models for Expensive Multi-objective Optimization

FoMEMO establishes the first foundation model paradigm for expensive multi-objective optimization, pre-training a PFN architecture on hundreds of millions of synthetic GP datasets to predict preference-conditioned aggregated posterior distributions in a single forward pass, enabling zero-shot, update-free in-context optimization.

Goal-Conditioned Supervised Learning for Multi-Objective Recommendation

MOGCSL preserves multiple future session rewards as a goal vector, trains a goal-conditioned next-item predictor with ordinary cross-entropy, and selects inference goals using statistics or two CVAEs, improving purchase prediction and training cost without dominating every click metric or guaranteeing that specified goals are attainable.

Is Your LLM-as-a-Recommender Agent Trustable? LLMsโ€™ Recommendation is Easily Hacked by Biases (Preferences)

This paper introduces BiasRecBench to assess the vulnerability of the LLM-as-a-Recommender paradigm, employing an \(\epsilon\)-Bound quality-margin calibration pipeline to demonstrate that frontier LLMs are heavily susceptible to authority, bandwagon, and instruction biases even when possessing sufficient reasoning ability to identify the ground truth.

LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling

LIME proposes a decoupled attention recommendation framework using globally learned "link embeddings" combined with a bidirectional XOR attention mask, reducing user-history self-attention to linear complexity and enabling offline pre-computation of candidate interaction weights, matching heavy Transformer accuracy while delivering a 10ร— inference speedup.

LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling

LIME resolves the fundamental expressiveness-efficiency dilemma in recommendation test time scaling by introducing global learnable link embeddings as an intermediary, using decoupled XOR attention on the user side for \(O(\ell \cdot N)\) linear long-history modeling and caching offline candidate attention weights on the item side for constant-time \(O(\ell)\) scoring.

MARS: Multi-Resolution Adaptive Routing for Sequential Recommendation

Addressing temporal aliasing and the erasure of short-term intent in long-history recommender compression, MARS writes interaction sequences into recurrent state tracks anchored to distinct half-life priors and dynamically reads them via sparse Top-\(k\) routing into compact seed memories, preserving fixed-size candidate scoring while boosting accuracy and behavioral-shift robustness.

Mult-DPO: Multinomial Direct Preference Optimization for Recommender Systems

Addressing the fundamental mismatch between multi-positive unordered set-wise feedback and pairwise DPO assumptions in recommender systems, Mult-DPO introduces a multinomial surrogate likelihood that provably upper-bounds the intractable marginalized Plackett-Luce loss, enabling efficient closed-form alignment across multiple candidates and fine-grained preference levels.

Not All Is Lost: Repairing Lossy User Preference States of Personalization Encoders

This paper introduces REPAIR, a post-compression state correction framework that operates under frozen personalization encoders and task heads by mining cached per-timestep representations to construct low-rank state-relative residuals, decomposing them across multi-scale temporal bases, and selectively writing corrections back to bridge the recoverability gap.

Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for RL-based LLM Recommenders

This paper theoretically demonstrates that GRPO with binary rewards is equivalent to optimizing global AUC while beam-search negative sampling reshapes the objective into One-way Partial AUC (OPAUC); it further introduces Windowed Partial AUC (WPAUC) and an efficient Threshold-Adjusted Windowed reweighting (TAWin) method to achieve controllable, fine-grained alignment with targeted Top-K metrics.

Rethinking Personalized Generation: Test-Time Alignment via Factorized Ranking Models

The paper reframes personalized generation as selecting suitable candidates already produced by a generator, using an MLP with 2.8M parameters by default to rank frozen generator representations; it beats four generalist reward models across nine datasets but remains behind a dataset-finetuned 8B reward model, while single-trajectory decoding guidance recovers only part of the selection gain.

Rethinking Semantic ID Construction for Generative Recommendation: SimHash with Parallel Decoding and Semantic Alignment

Challenging the consensus that Locality-Sensitive Hashing (SimHash) is inherently inferior to learned quantizers for generative recommendation, this work attributes the performance gap to autoregressive decoding mismatch and discretization information loss, proposing FLASH to revitalize training-free SimHash via parallel decoding and semantic alignment for superior accuracy and cold-start generalization.

Self-EvolveRec: Self-Evolving Recommender Systems with LLM-based Directional Feedback

Self-EvolveRec establishes an LLM-driven code evolution framework for recommender systems that replaces undirected scalar-metric trial and error with directional feedback from a user simulator and a model diagnosis tool, coupled with a co-evolving diagnostic mechanism to adapt evaluation criteria dynamically.

Simple Projection-Free Algorithm for Contextual Recommendation with Logarithmic Regret and Robustness

By exploiting the positive scale-invariance of action selection in contextual recommendation, this paper eliminates the expensive Mahalanobis projection step of Online Newton Step methods, introducing a simple projection-free second-order perceptron algorithm (CoRectron) that achieves optimal \(O(d \log T)\) logarithmic regret and natural robustness to suboptimal action feedback.

TRACE: Tourism Recommendation with Accountable Citation Evidence

TRACE establishes the first conversational tourism recommendation benchmark that jointly evaluates preference matching, review-span citation attribution, and mid-dialogue rejection recovery across 10,000 dialogues and 25 metrics, exposing the "Three-Competency Gap" between LLMs and non-LLM retrievers.

Trie-Aware Transformers for Generative Recommendation

Addressing the limitation where generative recommenders flatten hierarchical semantic IDs into linear streams and disregard prefix tree (trie) topologies, TrieRec introduces trie-aware absolute positional encoding (TAE) to aggregate item-path context and topology-aware relative positional encoding (TRE) based on lowest common ancestors, delivering an average relative improvement of 8.83% across three major backbones.

UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence

UxSID introduces discrete Semantic IDs (SIDs) as high-density semantic probes to guide ultra-long user behavior sequence compression, combining an item-agnostic interest compression module with hierarchical gated dual-stage attention to achieve target-aware interest routing while maintaining constant-time online inference latency and delivering strong CTR gains.

What Gets Measured Gets Managed: Sign-aware Recommendation Needs Sign-aware Evaluation

Revealing that state-of-the-art sign-aware graph recommender systems suffer from a pervasive "valence-blindness" at the inference ranking stage, this paper proposes a family of signed evaluation metrics (SRecall, SHR, SNDCG) that penalize negative recommendations and designs an orthogonal auxiliary loss to decouple valence from relevance.