Yueqin Yin (殷月琴)

Autocomplete models trained on editing logs learn to guess, because the logged completion often depends on information not yet visible at the cursor. We trim each SFT target to its longest context-inferable prefix, keep empty targets as silence labels, and in RL place deterministic abstention and grounding gates before a graded LLM-judge reward, which avoids the silence collapse of a judge-only reward. The deployed 4B dense model raised live acceptance from 17.8% to 26.35% over the previous 30B-A3B MoE while cutting median latency by 71%.

When a prompt contains several instructions and the reward counts how many are satisfied, RL keeps improving the easy instructions while the hard ones barely move. We propose two metrics that measure this exploration bias, then reduce it in two stages: Behavioral Bootstrapping, a short rejection-sampling fine-tuning that activates hard instructions before RL, and Scarcity-Aware Rewards, which weight each instruction by how rarely it is satisfied. Beats the cumulative-reward baseline by up to 9.2 points on IFBench.

Checks each generated sentence against the source document while reading the preceding sentences to resolve pronouns, references, and ellipsis, removing the separate decontextualization step of prior pipelines. Earlier context is used only to disambiguate; evidence comes only from the source. Gains 10+ Macro-F1 points on three new context-dependent benchmarks and matches or beats 8B-scale verifiers on 14 standard single-sentence datasets.

A single reward per response is too sparse for RLHF, and a single token is too small to carry meaning. We split responses into semantically complete segments using the SFT model's entropy, learn segment-level rewards from ordinary sequence-level preferences, and stabilize PPO with location-aware reward normalization and within-segment reward interpolation. Outperforms sequence-, sentence-, and token-level reward variants on AlpacaEval 2.0, Arena-Hard, and MT-Bench.

A synthetic coding dataset of 447K question–solution–test triplets from 12 sources. Each solution is verified by executing its unit tests, and hard problems get extra attempts so they are not filtered out. Models fine-tuned on KodCode reach state-of-the-art results on HumanEval(+), MBPP(+), BigCodeBench, and LiveCodeBench, and the data also supports RL with verifiable rewards.

DPO only compares two responses to the same prompt. RPO also contrasts responses across semantically related prompts, weighting each pair by prompt similarity, so it can learn from both paired and unpaired preference data. Outperforms DPO, IPO, and KTO on Anthropic-HH, OpenAI Summarization, and AlpacaEval 2.0.

Preference alignment without paired preference data or an external reward model. SAPO builds its own negatives by regenerating segments of chosen responses, using an EMA model and a replay buffer to keep training pairs fresh as the policy improves. Matches or exceeds DPO, ORPO, and SPIN on the Open LLM Leaderboard, IFEval, MT-Bench, and AlpacaEval 2.0.

A Transformer-based dual-space GAN that lets two latent spaces interact through a cross-space interaction mechanism, so the structure (e.g., pose) and texture (e.g., style) of a face can be edited independently. Improves controllability and identity preservation on CelebA-HQ and FFHQ.