Yichi Zhang
Publications
Kimi K2.5: Visual Agentic Intelligence
We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.
Small-Margin Preferences Still Matter-If You Train Them Right
Preference optimization methods such as DPO align large language models (LLMs) using paired comparisons, but their effectiveness can be highly sensitive to the quality and difficulty of preference pairs. A common heuristic treats small-margin (ambiguous) pairs as noisy and filters them out. In this paper, we revisit this assumption and show that pair difficulty interacts strongly with the optimization objective: when trained with preference-based losses, difficult pairs can destabilize training and harm alignment, yet these same pairs still contain useful supervision signals when optimized with supervised fine-tuning (SFT). Motivated by this observation, we propose MixDPO, a simple yet effective difficulty-aware training strategy that (i) orders preference data from easy to hard (a curriculum over margin-defined difficulty), and (ii) routes difficult pairs to an SFT objective while applying a preference loss to easy pairs. This hybrid design provides a practical mechanism to leverage ambiguous pairs without incurring the optimization failures often associated with preference losses on low-margin data. Across three LLM-judge benchmarks, MixDPO consistently improves alignment over DPO and a range of widely-used variants, with particularly strong gains on AlpacaEval~2 length-controlled (LC) win rate.
RLPO: Residual Listwise Preference Optimization for Long-Context Review Ranking
Review ranking is pivotal in e-commerce for prioritizing diagnostic and authentic feedback from the deluge of user-generated content. While large language models have improved semantic assessment, existing ranking paradigms face a persistent trade-off in long-context settings. Pointwise scoring is efficient but often fails to account for list-level interactions, leading to miscalibrated top-$k$ rankings. Listwise approaches can leverage global context, yet they are computationally expensive and become unstable as candidate lists grow. To address this, we propose Residual Listwise Preference Optimization (RLPO), which formulates ranking as listwise representation-level residual correction over a strong pointwise LLM scorer. RLPO first produces calibrated pointwise scores and item representations, then applies a lightweight encoder over the representations to predict listwise score residuals, avoiding full token-level listwise processing. We also introduce a large-scale benchmark for long-context review ranking with human verification. Experiments show RLPO improves NDCG@k over strong pointwise and listwise baselines and remains robust as list length increases.