Ruihan Sun

Zhenbo Xu, Jinghan Yang, Gong Huang et al.

With the rise and development of computer vision and LLMs, intelligence is everywhere, especially for people and cars. However, for tremendous food attributes (such as origin, quantity, weight, quality, sweetness, etc.), existing research still mainly focuses on the study of categories. The reason is the lack of a large and comprehensive benchmark for food. Besides, many food attributes (such as sweetness, weight, and fine-grained categories) are challenging to accurately percept solely through RGB cameras. To fulfill this gap and promote the development of intelligent food analysis, in this paper, we built the first large-scale spectral food (SFOOD) benchmark suite. We spent a lot of manpower and equipment costs to organize existing food datasets and collect hyperspectral images of hundreds of foods, and we used instruments to experimentally determine food attributes such as sweetness and weight. The resulting benchmark consists of 3,266 food categories and 2,351 k data points for 17 main food categories. Extensive evaluations find that: (i) Large-scale models are still poor at digitizing food. Compared to people and cars, food has gradually become one of the most difficult objects to study; (ii) Spectrum data are crucial for analyzing food properties (such as sweetness). Our benchmark will be open source and continuously iterated for different food analysis tasks.

Naifan Zhang, Ruihan Sun, Jinwei Su et al.

Subjective multi-turn dialogue tasks, such as emotional support, require conversational policies that adapt to evolving user states and optimize long-horizon interaction quality. However, reinforcement learning (RL) for such settings remains challenging due to the absence of reliable process supervision. Outcome-only training collapses credit assignment across turns into a single trajectory-level reward, while naïve turn-level group sampling incurs prohibitive rollout costs in interactive environments. We propose a critic-free and efficient RL algorithm named MAPO that leverages dense process feedback from a judge model and propagates long-horizon effects through Monte Carlo returns. To stabilize optimization, we introduce a mixed advantage estimator that combines turn-level normalization with batch-level normalization, enabling fine-grained yet scalable credit assignment. Across multiple subjective dialogue benchmarks, including EMPA, EmoBench, and EQ-Bench, and model scales ranging from 7B to 32B, our method consistently improves both training stability and final performance over outcome-only GRPO and single-level normalization baselines. On EMPA, we improve rates by up to 9 points and increase dialogue scores by as much as +43.2 over the 7B base model. Despite training only on EMPA-style environments, our approach generalizes well, yielding consistent improvements on unseen emotional-intelligence benchmarks, including up to +4 points on EmoBench and +3.5 on EQ-Bench. Together, these results demonstrate that dense process supervision combined with mixed-level normalization enables effective and scalable RL for subjective, open-ended multi-turn dialogue.