The Recall Ceiling of LLM Recommendation Reranking

Large language models (LLMs) are increasingly used as rerankers in conversational recommender systems, yet measured gains depend strongly on the retrieval and inference protocol. On the ReDial conversational movie recommendation benchmark, we compare proprietary, open-weight, and fine-tuned LLM rerankers with collaborative-filtering and sequential baselines in a shared retrieve-then-rerank pipeline. We vary candidate-pool size, first-stage retriever, and decoding temperature. With a shared semantic top-250 candidate pool and strict candidate-aware scoring, the best proprietary reranker reaches NDCG@10 of 0.1497, compared with 0.0939 for the strongest non-LLM baseline. The same reranker reaches 0.2925 in zero-shot generation, showing that unconstrained scoring can yield a much larger apparent advantage than matched-pool evaluation. No evaluated open-weight LLM outperforms the tuned shallow autoencoder baseline under this protocol. For the strongest proprietary and open-weight rerankers, switching from semantic to collaborative-filtering candidates raises NDCG@10 by more than 50%, showing that measured reranker performance is highly sensitive to candidate generation. For the best proprietary reranker, raising temperature from 0 to 1.0 increases top-10 Jaccard distance from 0.0900 to 0.1240 while mean NDCG@10 changes negligibly, whereas weaker LLMs show larger degradation. These ReDial results support treating candidate generation, candidate-pool size, scoring policy, and decoding configuration as required reporting fields rather than implementation details.

Recent studies increasingly explore Large Language Models (LLMs) as a new paradigm for recommendation systems due to their scalability and world knowledge. However, existing work has three key limitations: (1) most efforts focus on retrieval and ranking, while the reranking phase, critical for refining final recommendations, is largely overlooked; (2) LLMs are typically used in zero-shot or supervised fine-tuning settings, leaving their reasoning abilities, especially those enhanced through reinforcement learning (RL) and high-quality reasoning data, underexploited; (3) items are commonly represented by non-semantic IDs, creating major scalability challenges in industrial systems with billions of identifiers. To address these gaps, we propose the Generative Reasoning Reranker (GR2), an end-to-end framework with a three-stage training pipeline tailored for reranking. First, a pretrained LLM is mid-trained on semantic IDs encoded from non-semantic IDs via a tokenizer achieving $\ge$99% uniqueness. Next, a stronger larger-scale LLM generates high-quality reasoning traces through carefully designed prompting and rejection sampling, which are used for supervised fine-tuning to impart foundational reasoning skills. Finally, we apply Decoupled Clip and Dynamic sAmpling Policy Optimization (DAPO), enabling scalable RL supervision with verifiable rewards designed specifically for reranking. Experiments on two real-world datasets demonstrate GR2's effectiveness: it surpasses the state-of-the-art OneRec-Think by 2.4% in Recall@5 and 1.3% in NDCG@5. Ablations confirm that advanced reasoning traces yield substantial gains across metrics. We further find that RL reward design is crucial in reranking: LLMs tend to exploit reward hacking by preserving item order, motivating conditional verifiable rewards to mitigate this behavior and optimize reranking performance.

Large language models (LLMs) are increasingly used as rerankers in recommender systems, with the expectation that semantic understanding will help in cold-start and long-tail regimes. We test this assumption with a five-domain benchmark that explicitly separates reranking quality from retrieval coverage. In a positive-controlled regime where the gold item is guaranteed present, calibrated LLM rerankers fail to consistently outperform strong collaborative and content baselines under natural traffic, and within-family scaling from Qwen3-8B to Qwen3-32B narrows but does not close the gap on most domains. In a retrieval-realistic regime where the gold item is not injected, the bottleneck is more severe: standard single retrievers place the gold item in a 200-item pool only 4.6-22.9% of the time, largely because 32-91% of cold-start targets are brand-new items with no training interactions. We introduce LHF, a validation-trained learned hybrid fusion layer over a multi-retriever union pool, as a retrieval-side realizability baseline. LHF is the only combiner we test that beats every single retriever on all five domains and recovers 17-61% of oracle coverage headroom on content-rich domains, but only 5-7% on collaboratively strong domains. End-to-end experiments reveal the remaining mismatch: learned non-LLM ranking exploits the LHF pool, while prompt-level LLM reranking often degrades it. LLMs exhibit pockets of semantic cold-start advantage, especially in text-rich domains when the item is already present, but this advantage is largely unreachable in current retrieve-then-rerank pipelines. We release the benchmark protocol, splits, prompts, evaluation tooling, and archived reproducibility artifacts: data at https://doi.org/10.5281/zenodo.20991039 and code at https://doi.org/10.5281/zenodo.20993306.