We are glad to announce many new “just accepted” articles at ACM TORS. They cover important themes shaping the next decade of recommender-systems research. We are particularly proud to see contributions from leading authorities in the RecSys community. In the past six months alone this includes articles by Alan Said, Barry Smyth, Dietmar Jannach, Michael Ekstrand, Robin Burke, Francesco Ricci, Joeran Beel, Maria Soldedad Pera, Markus Zanker, Marko Tkalcic, Li Chen, and many more.

The following includes only a selection of our latest ‘just accepted’ articles. For a full list and download, visit https://dl.acm.org/journal/tors.

A field taking stock—and setting new standards

Three decades of recommender-systems work are mapped with unusual breadth and statistical rigor in Barry Smyth’s “People Who Liked This Also Liked … A Publication Analysis of Three Decades of Recommender Systems Research.” Beyond historiography, the article quantifies how a “core” RS community differs from the wider, RS-adjacent literature: outside venues publish roughly three times as many RS-related papers with about five times as many authors, yet the core RS papers accrue about 40 citations per article versus 12 outside and are cited earlier (≈0.6 vs. 1.1 years to first citation). Less than 13% of core papers receive zero citations compared to about 30% outside the core. Smyth also reports topic momentum, leading venues (e.g., RecSys with ~1,700 RS papers at ≈44 citations per article; KDD papers exceeding 100 citations/article), and a ranked list of the 25 most-influential recent papers (2014–2023). These numbers offer editors, authors, and program chairs a baseline for assessing impact and disciplinary diffusion (Smyth 2025).

Sustainability as a first-class objective

“Green Recommender Systems: Understanding and Minimizing the Carbon Footprint of AI-Powered Personalization” turns environmental impact from an afterthought into a design constraint. Alan Said, Joeran Beel and their PhD students Lukas Wegmeth and Tobias Vente instrument typical RS pipelines from RecSys 2013 and 2023, measuring energy via hardware meters and reporting emissions in CO₂e. A representative 2023-style pipeline with modern deep models consumes ≈6,100 kWh for experiments plus ≈500 kWh for ancillary tasks, yielding ~447 kg CO₂e under a Swedish-mix assumption; the paper details when and why transformer components dominate energy, and shows how curating datasets and models reduces cost without hurting accuracy. The work closes with concrete guidance (e.g., energy-aware benchmarking, carbon accounting) and a call to action to adopt green RS practices in both academia and industry (Wegmeth et al. 2025).
A complementary sustainability lens appears in tourism. Merinov & Ricci introduce an offline, data-driven protocol to simulate how the salience of sustainable recommendations (e.g., less-crowded POIs) shifts behavior and experience quality when classic train/test splits and even A/B tests are impractical. Their design connects causal inference with practical evaluation for sustainability-oriented RS deployments (Merinov & Ricci 2025).

Responsible and inclusive RS

“Recommending With, Not For: Co-Designing Recommender Systems for Social Good” argues for participatory methods that rebalance power between designers and affected communities—bridging fairness, evaluation, and HCI traditions. The paper consolidates user-centric evaluation practices and positions co-design as a method to align RS with public-interest goals (Pera et al. 2025).
“Inclusive Recommender Systems: Addressing Needs of Users with ADHD” operationalizes inclusion. Through three-stage, participatory sessions (six adult participants; ~18 total hours), the authors derive design needs around self-regulation, time loss, and hyper-focus triggered by recommendation streams, and translate coping strategies into concrete interface features and co-design guides. This adds a rarely documented neurodiversity perspective to RS design (Khan et al. 2025).

Industrial-scale efficiency and engineering

At production scale, engineering choices dominate both cost and recency. “SplitE: Enhancing Knowledge Graph Embedding Precision with Entity Split and Contextualization” introduces a two-stage embedding pipeline with entity splitting and contextualization plus an aggregation function for frequently updated nodes (e.g., job openings, associates). On Walmart’s People.AI KG (≈1.6 M nodes, 83 M edges), monthly retraining plus on-the-fly aggregation reduces cloud GPU cost by >90% relative to naive frequent retraining (eight A100s × 22 h per run; ~$201,480/year on-demand) while improving Hit@10 across benchmarks; SplitE outperforms strong baselines particularly on symmetry- and polysemy-heavy datasets like WN18 (26,896 polysemous words) (Wang et al. 2025).

LLMs for lifelong sequences—toward principled gains

“Full-Stack Optimized Large Language Models for Lifelong Sequential Behavior Comprehension in Recommendation” identifies a failure mode: LLMs’ CTR performance peaks around short sequences (e.g., K≈15 on MovieLens-1M) and degrades as sequences lengthen. The ReLLaX framework addresses this at three levels: data (semantic user-behavior retrieval, SUBR), prompts (soft-prompt augmentation, SPA), and parameters (a component-fully-interactive LoRA, CFLoRA). Across BookCrossing, ML-1M, and ML-25M, ReLLaX improves AUC and log-loss over LLM baselines; ablations show the CFLoRA module and SUBR each provide measurable gains (e.g., better AUC/log-loss than TALLRec/iLoRA under matched settings). The paper also provides a theoretical view in which prior LoRA-based LLM4Rec variants emerge as constrained special cases of CFLoRA (Shan et al. 2025).

Cross-cutting implications

Taken together, this set raises the bar on (i) measurement—from carbon accounting and causal-ish evaluation protocols to community-level bibliometrics, (ii) responsibility—from co-design with affected groups to design patterns for neurodiverse users, and (iii) scalability—from KG embeddings that cut cloud cost by an order of magnitude to LLM adaptations that recover accuracy on long behavioral histories. Notably, Green RecSys (Wegmeth et al. 2025) surfaces energy as a first-order metric to be reported alongside accuracy, and Smyth (2025) supplies field-level indicators of influence and diffusion that editors and reviewers can adopt. The presence of renowned authors—Barry Smyth on the field’s history; Alan Said and Joeran Beel on sustainability; Francesco Ricci on tourism RS; and Weinan Zhang’s team on LLM4Rec—underlines both scholarly continuity and methodological innovation across this TORS cohort.

---

References

Be sure to consult the final published versions for authoritative pagination and metadata.

• Khan, S., et al. (2025). Inclusive Recommender Systems: Addressing Needs of Users with ADHD. ACM TORS. (Six participants; ~18 h; participatory co-design and guides.)
• Merinov, P., & Ricci, F. (2025). Shaping Sustainable Tourist Experience: Simulating the Impact of Recommendation Salience. ACM TORS. (Offline, data-driven protocol for behavior-change evaluation.)
• Pera, M. S., et al. (2025). Recommending With, Not For: Co-Designing Recommender Systems for Social Good. ACM TORS. (Participatory frameworks for RS for social good.)
• Shan, R., Zhu, J., Lin, J., Zhang, W., et al. (2025). Full-Stack Optimized Large Language Models for Lifelong Sequential Behavior Comprehension in Recommendation. ACM TORS. (ReLLaX; SUBR, SPA, CFLoRA; AUC/log-loss gains.)
• Smyth, B. (2025). People Who Liked This Also Liked … A Publication Analysis of Three Decades of Recommender Systems Research. ACM TORS. (Core vs. outside metrics; venue/topic influence; top-25 list.)
• Wang, Q., et al. (2025). SplitE: Enhancing Knowledge Graph Embedding Precision with Entity Split and Contextualization. ACM TORS. (Walmart People.AI; >90% cloud-cost reduction; better Hit@10/MRR.)
• Wegmeth, L., Vente, T., Said, A., & Beel, J. (2025). Green Recommender Systems: Understanding and Minimizing the Carbon Footprint of AI-Powered Personalization. ACM TORS. (≈6,100 kWh+; ~447 kg CO₂e; mitigation guidelines.)

Disclaimer: Joeran Beel is both author of this blog post and and author of one featured article.

Information on AI: This post was written by AI, based on the PDFs of all articles recently accepted by ACM TORS. The AI choose, which authors and articles to summarize. This is not to say, that missing articles were of lesser relevance to the recsys community.