
Solving LLM recommender cold start with knowledge graph RAG
MixRAGRec (KDD 2026) searches deeper the less it knows a user: multi-agent recommendation that picks retrieval depth per query
A KDD 2026 study let an LLM recommender pick knowledge graph retrieval depth per query. Cold-start accuracy rose from 0.39 to 0.50 and latency halved.
An LLM recommender gets a user's taste right by searching a knowledge graph more deeply the less it knows about that user. MixRAGRec, a KDD 2026 paper from Hong Kong Polytechnic University, let the model pick one of four retrieval depths per query; in a cold-start setting it chose deep graph retrieval 96.6% of the time and reached 0.501 accuracy, above the prior knowledge graph RAG method at 0.392 (paper, Appendix E)[1]. In the regular setting it saved retrieval and cut latency by more than half.
The source is the MixRAGRec paper by seven researchers from Hong Kong Polytechnic University and the National University of Singapore[1]. It was accepted to KDD 2026, held in Jeju in August 2026, and the code is public[2]. We did not reproduce the results, and the figures come from the arXiv v2 text as of November 2026. It is an offline evaluation on public movie and music data, so it should be read apart from effects in a live service.
We build a product that measures how often brands appear in AI search and AI recommendations, which gives us a stake in what AI recommends from. Where the study showing AI search knows brand names but does not recommend new products looked at cold start from outside, this paper tackles cold start inside a recommender. It is also a third case of deciding "how much to look up," after the study showing agents know in advance whether they need to search and the study verifying RAG answers after generation.
Why LLM recommenders suggest products that do not exist
LLMs can read a user's history and item descriptions to infer taste, so they have started to be used for recommendation. The catch is that they are stuck with training-time knowledge. The paper gives the example of recommending a nonexistent sequel, "The Matrix 5," to a Matrix fan, and notes that titles released after the training data cannot be recommended at all (paper, Section 1)[1]. Cold start, the old recommender problem of new users and new items with almost no history, survives in a new form in the LLM era, as a 2025 survey lays out[3].
That is why knowledge graph RAG appeared, retrieving from an external knowledge graph and feeding the results to the LLM as evidence. Attaching a relation such as "The Terminator, director, James Cameron" lets the LLM choose from facts instead of inventing. K-RagRec, published at ACL 2025 by this paper's first author, is the representative earlier method[1].
Knowledge graph RAG recommenders used one fixed retrieval depth
Existing knowledge graph RAG recommenders retrieve at the same depth for every query. The authors split the resulting problem in two. Users with clear taste get unnecessary relations attached, adding noise and cost (over-retrieval), while users with vague taste miss relations several hops away and lack evidence (under-retrieval) (paper, Section 1 and Figure 1)[1]. Turning a graph into the plain text an LLM reads also loses structure and adds noise. And whether a depth was right is only known from whether the final recommendation was right, so there is no direct label to train retrieval depth on.
How to adjust RAG retrieval depth per query
MixRAGRec sets up four experts with different retrieval depths and has a router pick one per query (paper, Section 3.3)[1].
| Expert | Retrieval | When it fits |
|---|---|---|
| E1 | No retrieval | Viewing history alone makes the taste clear |
| E2 | A few relations close to the query | A few facts settle the answer |
| E3 | Several hops of neighbors around nearby entities | Several relations need to be seen together |
| E4 | A connected graph linking entities chosen by personalized PageRank at minimum cost | The whole context is needed |
Table 1. Four experts with different retrieval depths. Source: Wang et al. 2026, Section 3.3 and Figure 3[1].
The knowledge graphs were built from DBpedia to match the movie and music data. Entities and relations sit in the same sentence-embedding space and are retrieved by similarity to the query embedding. E4 builds a minimum spanning tree with edge costs that are cheaper the closer a relation is in meaning to the query.
Here is the query embedding and is the embedding of the relation (such as director or genre) that edge represents. Relations unrelated to the query cost more and drop out of the tree.
Translating the graph into text and choosing against confusable negatives
Retrieved relations are first flattened with templates, then an alignment agent rewrites them into short text keeping only what helps the recommendation. A graph tying three family films together, for instance, becomes a description like "shared traits of family audiences, light adventure and slapstick" (paper, Section 3.4 and Table 10)[1]. The recommendation agent reads that description with the viewing history and picks a candidate, trained contrastively to raise the correct item's probability against the 10 negatives most confusable with it.
One reward trains all three agents
The router, alignment agent and recommendation agent are not trained separately but share one reward. The authors call this MMAPO (paper, Section 3.6)[1].
Here is the confidence placed on the recommendation if it is correct and -0.1 if it is wrong. is the recommendation distribution with the chosen expert's retrieval, is the distribution without retrieval (E1), and the KL divergence between them is the information gained from retrieval. is a cost that grows with retrieval depth, and the experiments used and . Expensive retrieval is worth it only when it changes the recommendation enough.
Experimental setup
The data are MovieLens-1M (6,040 users, 3,883 movies), MovieLens-20M and the music dataset LastFM-1K (paper, Table 1)[1]. Each user's most recent item is the answer, the 10 items before it are the history, and the model picks from 20 items mixing the answer with 19 distractors. Backbones are LLaMA3-8B and Mistral-7B, and baselines are direct recommendation by general LLMs such as GPT-4o, the recommendation-tuned TallRec and Rec-r1, and the knowledge graph RAG methods KG-Text, KAPING, G-retriever and K-RagRec.
Accuracy and latency both improved
MixRAGRec had the highest accuracy on all three datasets, improving on the strongest baseline by 5.8% to 20.4% with LLaMA3-8B (paper, Table 2)[1]. Used directly, even GPT-4o reached only 0.420 on MovieLens-1M. KG-Text, which just pastes the graph in as text (0.234), fell below the recommendation-tuned models, evidence that preserving graph structure matters.
| Method | MovieLens-1M | MovieLens-20M | LastFM-1K |
|---|---|---|---|
| GPT-4o direct | 0.420 | 0.524 | 0.685 |
| TallRec | 0.391 | 0.521 | 0.699 |
| Rec-r1 | 0.402 | 0.594 | 0.872 |
| K-RagRec | 0.454 | 0.578 | 0.845 |
| MixRAGRec | 0.504 | 0.676 | 0.934 |
Table 2. Accuracy of picking the answer from 20 candidates (LLaMA3-8B backbone except GPT-4o). Source: Wang et al. 2026, Table 2[1].
Latency was 1.56 seconds, under half of K-RagRec (3.78 seconds). Fixing a single expert made accuracy and time rise together with depth, with E4 at 0.499 and 2.27 seconds, while MixRAGRec reached a higher 0.504 with retrieval time at the level of the shallow E2 (0.063 seconds) (paper, Table 3)[1]. The router spent expensive retrieval only on the queries that needed it. In the ablations, removing the alignment agent caused the largest drop.
How to solve LLM recommender cold start: search deeper the less you know
MixRAGRec was also best in the cold-start setting, with 0.501 accuracy and 0.778 Recall@3, ahead of K-RagRec (0.392, 0.706) and G-retriever (0.356, 0.641), while LLaMA3-8B without retrieval scored 0.083 (paper, Appendix E and Table 6)[1].
The router's choices stand out more. In cold start it picked neighborhood graphs (E3) 37.9% and connected graphs (E4) 58.7% of the time, so deep retrieval made up 96.6%. On the regular music data it spread E1, E2 and E3 at 30.2%, 28.6% and 28.2%, with E4 at only 13.0% (paper, Appendices D and E)[1]. When it knows a user well it does not look things up; the less it knows, the more it pulls in world knowledge such as director, genre and theme to infer why someone liked something. Read with the study tracing why a shop is missing from AI recommendations to documentation, it points the same way: the raw material of recommendation is structured relations.
What this means for personalized AI agent recommendation
From here on this is our interpretation. The paper does not cover a proactive agent that suggests before being asked. It is a reactive setting that picks the next item from a history, and "multi-agent" refers to three internal modules. What it shows instead is a step personalized agents need first: deciding for itself, against the cost, how much more to look up about a user. Seen against the shift of AI search from exposure toward recommendation and action paths, that judgment shapes both the quality and the cost of agent recommendations.
There is a brand implication too. The evidence this system pulls in is DBpedia relations such as director, genre and cast. If AI shopping agents work in a similar way when recommending products, the more a brand's relations (category, use case, ingredients, substitutes) are published in structured form, the more room there is to be picked up as evidence in deep retrieval. This study does not cover commercial shopping agents, so this is inference resting on the assumption that the structure is the same. In practice it has to be measured per engine, as in the post checking whether AI chatbots' product recommendations differ by tool.
What this study did not cover
The biggest gap is the definition of cold start. Appendix E does not say how the cold-start sample was split (users with fewer than some number of items, or items unseen in training), so the cold-start numbers deserve the most caution. The evaluation is also an offline setting that picks one answer among 19 distractors, a different difficulty from a live service recommending among tens of thousands of items.
Data and user information are limited as well. MovieLens and LastFM are older public datasets, and of the 3,883 movies in MovieLens-1M, 3,253 link to the knowledge graph. The only user information in the prompt is the last 10 items and a "general user" label. Backbones are limited to two open models of 8B or smaller.
In practice, choose retrieval depth instead of fixing it
The lesson for teams building LLM recommendation or RAG is simple. Do not fix retrieval depth at one level; offer several, and put how much retrieval changes the result, together with its cost, into the reward that chooses among them. When feeding a graph to an LLM, do not paste it in as text; add a separate step that keeps only what helps the recommendation. Whether users with thin history lean on external knowledge more deeply in real data too is the first thing to check before moving this structure into a service.
Frequently asked questions
How do you solve cold start in an LLM recommender system?
For users or items with little interaction history, infer taste from external knowledge instead of logs. This study retrieved relations such as director, genre and cast from a knowledge graph and gave them to the LLM as evidence, reaching 0.501 accuracy in a cold-start setting against 0.392 for the prior knowledge graph RAG method. The paper does not say how the cold-start sample was defined.
How does knowledge graph RAG for recommendation differ from ordinary RAG?
It retrieves relations such as 'The Terminator was directed by James Cameron' instead of document chunks. That lets it follow relations several hops away to find why someone likes something, but translating a graph into text adds noise. This study handed that translation to a separately trained agent, and removing that agent caused the largest drop in performance.
How can RAG retrieval depth be adjusted per query?
Offer several options of different retrieval depth and let a router pick one per query. This study used four levels, from no retrieval to connected-graph retrieval, and trained the router with a reward that adds information gained from retrieval minus its cost. Latency came to 1.56 seconds, under half of the strongest prior method (3.78 seconds).
Is this a proactive AI agent that recommends before being asked?
No. It takes a user's ten most recent movies and picks the next one from 20 candidates, and 'multi-agent' refers to three internal modules. What it shows is a system deciding for itself how much more to look up about a user.
References
- [1]Shijie Wang, Chengyi Liu, Yujuan Ding, Shanru Lin, See-Kiong Ng, Xu Xin, Wenqi Fan, "Mixture-of-Experts Knowledge Graph Retrieval-Augmented Generation for Multi-Agent LLM-based Recommendation", KDD 2026, arXiv:2605.28175v2
- [2]MixRAGRec code (Sjay-Wang/MixRAGRec)
- [3]Weizhi Zhang et al., "Cold-Start Recommendation towards the Era of Large Language Models (LLMs): A Comprehensive Survey and Roadmap", arXiv:2501.01945, 2025
Summary
- LLM recommenders are stuck with training-time knowledge, missing new releases or inventing titles, and knowledge graph RAG curbs this but retrieved the same amount for every query.
- MixRAGRec set up four retrieval experts from no retrieval to connected graphs, let a router choose one per query, and trained the alignment and recommendation agents together under one reward.
- Accuracy on MovieLens-1M reached 0.504 against 0.454 for the strongest prior method, at 1.56 seconds, under half the latency.
- In a cold-start setting it scored 0.501 against 0.392, and the router chose deep graph retrieval 96.6% of the time: the less it knows a user, the more world knowledge it pulls in to infer taste.
- The cold-start sample is undefined, and the evaluation picks one of 20 candidates on older public datasets, so real-service effects need separate checking.
Check this topic against your own brand
TRAIL Search measures how ChatGPT and Perplexity answer your customers' questions and finds and fixes where the brand is missing, turning this post into a diagnosis. Start with 10 questions, no card required.
More posts

Calibrating a marketing mix model with geo-experiments
One ROAS number per geo-experiment discards carryover and saturation. A 2026 paper estimates all three MMM parameters from the control-treatment gap.

RAG answers improve more from verification than retrieval
Across 16 combinations of four RAG features, a post-generation completeness check alone beat three retrieval features combined, at double the latency.

AI agents already know when to search the web
LLM agents know whether they need a tool like web search before answering (AUROC 0.89 to 0.96). A UCSD study used that signal to cut tool calls by 48%.
Keep reading on this topic
- AI Overviews cite different sources than Google searchGoogle Search, AI Overviews, and Gemini cite sources overlapping only 18%, and sites blocking the AI crawler get dropped. A SIGIR 2026 study of 11,500 queries.
- AI search knows your product but rarely recommends itAI recognizes a product 99% by name but surfaces it only 3% in discovery queries. A 2025 study of 112 startups found GEO scores did not predict it.
- Measuring GEO's revenue impact with a marketing mix modelTo measure GEO's revenue effect, turn how often AI answers name a brand into expected exposures per market, then feed it to an MMM. A 2026 paper explained.
This post is part of the Papers category, which collects all 44 posts on the topic. See all posts in Papers