🎁 Recommender Systems

🧪 ICML2026 · 11 paper notes

📌 Same area in other venues: 🔬 ICLR2026 (24) · 💬 ACL2026 (22) · 🤖 AAAI2026 (27) · 🧠 NeurIPS2025 (24) · 🧪 ICML2025 (17) · 💬 ACL2025 (7)

🔥 Top topics: Recommendation ×3

A Paired Testing Protocol for Batch-Conditioned Refusal Robustness in LLM Serving

This paper treats the batch condition in LLM serving as a treatment variable for safety evaluation. It proposes a testing protocol consisting of safety-capability paired comparisons, scorer/human adjudication, cross-model expansion, continuous batching composition, and batch-invariant kernel ablation. The study concludes that refusal flips are real but low-frequency, model-specific, and dependent on the specific serving stack.

Can Recommender Systems Teach Themselves? A Recursive Self-Improving Framework with Fidelity Control

RSIR enables sequential recommendation models to generate new synthetic user interaction sequences using their own predictive capabilities, train a new model, and filter out samples deviating from the user preference manifold using a rank-based "fidelity check" to prevent self-consuming model collapse. It consistently improves NDCG/Recall by 4–11% across 4 datasets and 3 mainstream backbones, theoretically proving that this process is equivalent to implicit regularization along the tangent space of the user preference manifold.

GCIB: Graph Contrastive Information Bottleneck for Multi-Behavior Recommendation

GCIB employs a dual approach of "Graph Information Bottleneck + Cross-behavior Contrastive Learning." It first prunes edges in auxiliary behavior graphs that are irrelevant to the target task at the structural level (maximizing mutual information with the target behavior and minimizing mutual information with the original auxiliary graph via HSIC surrogates). It then aligns denoised auxiliary representations with sparse target representations using InfoNCE at the feature level, achieving a 7%–40% relative improvement in HR@10 / NDCG@10 across four multi-behavior recommendation benchmarks.

Incentivized Exploration with Stochastic Covariates: A Two-Stage Mechanism Design for Recommender System

RCB integrates "exploration-exploitation" and "user incentive compatibility" into a contextual bandit problem under Dynamic Bayesian Incentive Compatibility (DBIC) constraints. It proposes a two-stage algorithm (Cold Start + IPGS), proves \(\tilde{O}(\sqrt{KdT})\) regret in stochastic user covariate scenarios, allows for the integration of any offline learning oracle, and quantifies the "incentive price" — showing that the cold start sample size grows as \(1/\epsilon^2\) as the \(\epsilon\) constraint tightens.

Learning Design Skills as Memory Policies for Agentic Photonic Inverse Design

SkillPCF reformulates the inverse design of Photonic Crystal Fibers (PCF) as a "memory policy learning" problem. A PPO-trained controller selects Top-K memory operations from an evolvable skill library for each trajectory span. An executor implements these in trajectory memory, while MEEP electromagnetic simulation rewards simultaneously optimize both the controller and the skill library. This approach achieves a superior trade-off between design success rate and simulation budget compared to multiple LLM backends and classical optimization baselines.

Position: Neglecting the Sustainability of AI is Fuelling a Global AI Arms Race

Utilizing Karl Marx's "base-superstructure" framework, this position paper argues that current "sustainable AI" discussions are dominated by environmental dimensions while neglecting economic and social ones. It calls for the simultaneous elevation of both climate awareness and resource awareness axes and proposes the CARAML five-layer action framework (Individual / Community / Industry / Government / Global) to curb the escalating "global AI arms race."

Position: Stop Preaching and Start Practising Data Frugality for Responsible Development of AI

This position paper points out that the ML community has long been "preaching without practicing" regarding "data frugality"—while verbally acknowledging that coresets save energy, almost no one actually reports energy consumption or carbon emissions. Using ImageNet-1K as a case study, the authors calculate a conservative lower bound of approximately 5.82 GWh / 2589 tCO2e for downstream training and storage, calling for data frugality to evolve from a slogan into a measurable, actionable, and rewardable engineering practice.

Prompts for Public-Sector LLMs Should Be Governed as Commons

This is a position paper: the authors argue that LLM prompts used by the public sector should be versioned, provenanced, auditable, and vetoable like open-source commons. Based on a pilot benchmark using 443 neighborhood prompts from a North American city (augmented to 3,317) across five governance states, it provides three falsifiable predictions—governed prompts change output distributions, improve auditability, and shorten fault-remediation latency.

Rethinking Contrastive Learning for Graph Collaborative Filtering: Limitations and a Simple Remedy

The authors decompose the forward prediction of LightGCN into a "sum of learnable weights of multi-hop neighbor pairs." They find that the Sampled Softmax (SSM) loss only weights based on the structural similarity of the item-side neighbors and treats all four types of neighbor pairs (UU/II/UI/IU) indiscriminately. Consequently, they propose NT-SSM, which incorporates user-side structural similarity into the gradient and calibrates weighting strategies according to neighbor pair types, consistently outperforming SSM across four datasets and various GCF backbones.

RGMem: Renormalization Group-Inspired Memory Evolution for Language Agents

RGMem draws inspiration from the Renormalization Group (RG) in statistical physics to model the long-term dialogue memory of language agents as a multi-scale system ("Event Layer → Relation Layer → Concept Layer"). It employs threshold-triggered non-linear operators to coarse-grain fragmented dialogues into stable user profiles, thereby breaking the "stability vs. plasticity" trade-off.

T-POP: Test-Time Personalization with Online Preference Feedback

T-POP integrates "test-time alignment" with "neural dueling bandits." Without modifying LLM parameters, it learns a personalized reward function online using pairwise preference feedback per round, effectively addressing the cold-start problem in personalization for new users.