Searching agents have achieved significant progress in NLP, yet their multimodal counterparts are constrained by a lack of suitable evaluation benchmarks requiring reasoning over continuous video, audio, and images. We propose long-form TV dramas as a natural solution and introduce DramaSearch, encompassing 3,600 riddles across 9 dramas totaling 239 hours. Our dataset is generated via a graph-constrained pipeline that processes enriched captions into structured events and multi-granularity evidence views. By employing character-to-view and entity-to-view indexing, we ensure controllable clue disclosure and strictly unique answers. The riddles are categorized into multi-clue, multi-hop, and compositional types, requiring agents to aggregate isolated clues and navigate complex dependency chains. Experiments demonstrate that a Qwen3-8B agent, trained on a held-out set using SFT and RL, achieves an accuracy of 48.3%. This result establishes a strong baseline that performs favorably against zero-shot GPT-5 (37.0%), while indicating significant room for future research. All data and code will be publicly released.

Adapting Large Language Models (LLMs) for recommendation requires careful consideration of the decoding process, given the inherent differences between generating items and natural language. Existing approaches often directly apply LLMs’ original decoding methods. However, we find these methods encounter significant challenges: 1) amplification bias—where standard length normalization inflates scores for items containing tokens with generation probabilities close to 1 (termed ghost tokens), and 2) homogeneity issue—generating multiple similar or repetitive items for a user. To tackle these challenges, we introduce a new decoding approach named Debiasing-Diversifying Decoding (D³). D³ disables length normalization for ghost tokens to alleviate amplification bias, and it incorporates a text-free assistant model to encourage tokens less frequently generated by LLMs for counteracting recommendation homogeneity. Extensive experiments on real-world datasets demonstrate the method’s effectiveness in enhancing accuracy and diversity.