SELECTED WORK / 2024—2026代表工作 / 2024—2026

Research is a chain of decisions.研究,是一连串可解释的决策。

I study how models read evidence, allocate reasoning, learn from feedback, and act over time—from token-level distillation to streaming interaction.我研究模型如何读取证据、分配推理、利用反馈并持续行动:从逐 token 蒸馏到流式交互。

01 / REASONING02 / MULTIMODAL03 / AGENTS

JUMP TO快速浏览

ACTS ↗CATER ↗SP-OPD ↗CrystalBind ↗TEAL ↗ClueFlow ↗JoyAI-LLM Flash ↗SearchAuditor ↗SVI-Bench ↗Publications ↗论文成果 ↗
01 / LONG-HORIZON SEARCHUNDER REVIEW · ICLR 2027在审 · ICLR 2027
ACTS

Design the path to evidence.先设计证据路径。

A question can require several facts yet be solved by opening one page. For an agent, the path to those facts is the real learning problem.一道题可能需要多个事实,却只需打开一个页面就能答对。对搜索智能体而言,真正的训练对象是找到事实的过程。

01

Represent entity-level facts and document navigation as linked structures, with each fact tied to its source.将实体事实与文档导航连接起来,并保留每条事实的来源。

02

Shape questions around evidence that must be discovered through distinct search actions.围绕需要通过不同检索动作才能发现的证据设计问题。

03

Filter trajectories by the actions a teacher actually took, so training examples carry search behavior rather than only final answers.依据教师实际执行的取证动作筛选轨迹,让训练样本包含搜索行为,而不只是最终答案。

Evidence provenanceSearch dataAgent SFT
02 / REASONING EFFICIENCYUNDER REVIEW · ICLR 2027在审 · ICLR 2027
CATER

Spend reasoning where it is needed.把推理预算用在需要的地方。

Competence-Adaptive Training for Efficient Reasoning

A model that already solves a problem can still spend too many tokens explaining it. CATER uses the student's current success to decide when to develop capability and when to compress reliable reasoning.模型已经会解一道题,却仍可能消耗过多推理 token。CATER 根据学生当前的解题表现,决定何时巩固能力、何时压缩已经稳定的推理。

01

Fresh groups of student responses and verifier outcomes estimate competence per problem at each update.每次更新都用新采样的学生回答和验证结果,估计每道题的当前能力。

02

All-correct groups receive concise on-policy self-distillation; mixed groups receive correctness-based reinforcement learning.全对组接受简洁的 on-policy 自蒸馏;有对有错的组接受基于正确性的强化学习。

03

All-failed groups are deferred until a useful successful trajectory appears; new rollouts continually update each problem's route.全错组暂时搁置,等待出现可学习的成功轨迹;新的采样不断更新问题所属的训练路线。

Efficient reasoningOn-policy self-distillationRL
03 / ON-POLICY DISTILLATIONUNDER REVIEW · ICLR 2027在审 · ICLR 2027
SP-OPD

When the teacher and verifier disagree.当教师与验证器意见相左。

Support-Preserving On-Policy Distillation

A student can produce a correct solution whose sampled tokens the teacher would suppress. SP-OPD changes the distillation target so verifier-accepted trajectories are supported by the learning signal.学生可能写出正确解答,但教师目标却想压低其中的采样 token。SP-OPD 调整蒸馏目标,让通过验证的轨迹获得一致的学习信号。

01

Identify teacher–verifier conflict where a teacher assigns a token on a correct trajectory less probability than the behavior policy.识别教师—验证器冲突:正确轨迹中的采样 token,在教师分布下的概率低于行为策略。

02

KL-project the teacher onto a constraint that preserves the sampled token's support, while retaining relative probabilities among alternatives.用 KL 投影约束教师目标,保留采样 token 的支持度,同时保持其余候选 token 的相对概率。

03

Use the projected target in both distillation loss terms and balance the clipped surrogate by response.让两项蒸馏损失共用投影目标,并按整条回答平衡截断替代损失。

On-policy distillationKL projectionVerifier feedback
04 / GENERATIVE MODELSUNDER REVIEW · ICLR 2027在审 · ICLR 2027
CrystalBind

When does a color become the car's?属性归属,何时定型?

A model may draw a purple car and an orange house, yet give each object the other's color. CrystalBind studies when that assignment becomes hard to change during denoising, then uses the diagnosis to design inference-time guidance.模型可能画出汽车和房子,却把两者的颜色分配反了。CrystalBind 研究属性归属在去噪过程中何时难以改写,并据此设计推理时引导。

01

Counterfactual prompt swaps under the same initial noise test whether later text can still change which object owns an attribute.在相同初始噪声下交换提示词中的属性,检验后续文本还能否改变属性归属。

02

Read–write–propagate probes separate recognition of a changed condition, its effect on the latent, and the effect of that change on the final image.以 read–write–propagate 探针区分模型是否读到条件变化、是否写入 latent,以及这种变化能否影响最终图像。

03

Guidance combines a calibrated early window, live noun-attention targets, and separate objectives for object presence and attribute ownership—without training new model weights.引导方法结合早期干预窗口、动态名词注意力目标,以及分别针对对象出现和属性归属的目标,无需训练新模型权重。

Diffusion modelsAttribute bindingInference-time guidance
05 / MULTIMODAL LEARNINGUNDER REVIEW · AAAI 2027在审 · AAAI 2027
TEAL / Qwen

Make fixed features readable.让外部特征真正可读。

A projection can fit pre-extracted text, audio, and visual features into a token backbone's hidden size. It does not show that the pretrained model can use them. TEAL treats this as an interface and measurement problem.把预提取的文本、音频和视觉特征投影到 token 骨干的隐藏维度,并不代表预训练模型能用好这些信息。TEAL 将其作为接口与测量问题来研究。

01

Direct feature readers, no-backbone models, and random-initialized controls help distinguish available information from token-interface and routing effects.以直接读取原始特征的模型、移除骨干和随机初始化对照,区分上游信息量、token 接口与路由的影响。

02

Typed evidence slots, native-scale calibration, and learned missing tokens make modality identity, input bandwidth, and absent evidence explicit.类型化证据槽、原生尺度校准与可学习的缺失标记,明确表达模态身份、输入带宽和证据缺失。

03

Frozen and LoRA-adapted Qwen readers test what the new representation makes readable and where backbone routing still matters.通过冻结与 LoRA 适配的 Qwen 读取器,检验新表征能让模型读到什么,以及骨干路由在哪些环节仍然关键。

Multimodal featuresToken interfacesControlled evaluation
06 / AGENT STATECO-FIRST AUTHOR · ICLR 2027 UNDER REVIEW共同第一作者 · ICLR 2027 在审
ClueFlow

Give the agent a working memory it can explain.让智能体拥有可解释的工作记忆。

Long searches bury useful facts in a growing conversation. ClueFlow separates the short-term dialogue window from persistent evidence and candidate state.长程搜索会把关键事实埋进越来越长的对话。ClueFlow 将短期对话窗口与持久的证据、候选状态分开。

01

Evidence records capture source-backed facts; candidate records track what is supported, missing, or contradicted.证据记录保存有来源的事实;候选记录追踪已满足、缺失和矛盾的约束。

02

Browser-action provenance connects what was seen to what entered state and what shaped the final answer.浏览动作的来源链连接“看到什么”“记住什么”和“最终如何作答”。

03

Hybrid SFT and turn-level credit allocation reuse valid research steps, including those inside unsuccessful runs.Hybrid SFT 与逐轮信用分配保留有效取证动作,包括失败轨迹中做对的部分。

Persistent stateHybrid SFTTurn-level RL
07 / FOUNDATION MODELSTECHNICAL REPORT · 2026
JoyAI-LLM Flash

Make model training traceable.让模型训练有迹可循。

I am a named author on the JoyAI-LLM Flash technical report and contributed to its open release. The model is a 48B-total-parameter MoE with roughly 2.7B active parameters.我是 JoyAI-LLM Flash 技术报告署名作者,并参与模型开源发布。该模型为总参数 48B、激活参数约 2.7B 的 MoE。

01

At JD Explore Academy, my post-training work spans long-context full-parameter SFT and preference optimization for tool-using models.在京东探索研究院,我参与长上下文全参数 SFT、偏好优化等后训练工作。

02

I focus on the contracts that connect raw conversations, supervision masks, packed sequences, checkpoints, and agent evaluations.我重点核对原始对话、监督掩码、序列打包、模型权重与智能体评测之间的契约。

Post-trainingLong contextTraining systems
08 / FAILURE ANALYSISTHIRD AUTHOR · AAAI 2027 UNDER REVIEW第三作者 · AAAI 2027 在审
SearchAuditor

Find the first mistake that mattered.找到真正改变结果的第一处错误。

A failed search can contain dozens of plausible actions before a wrong answer appears. SearchAuditor traces the earliest consequential deviation and turns it into a concrete repair instruction.一次失败搜索可能包含数十个看似合理的动作。SearchAuditor 追踪最早产生实质影响的偏离,并给出可执行的修复建议。

01

Global review, backward constraint checking, and forward timeline tracing offer three complementary views of a run.全局审查、反向约束核对与正向时间线追踪,提供三种互补的轨迹视角。

02

Evidence-anchored adjudication ties each diagnosis to a specific message window and source.以证据锚定裁决,把每项诊断对应到具体消息窗口和来源。

Trace auditingFailure attributionOpen source
09 / STREAMING INTERACTIONCO-FIRST AUTHOR · ICLR 2027 UNDER REVIEW共同第一作者 · ICLR 2027 在审
SVI-Bench / InteractFlow

Evaluate the moment, not only the answer.评估开口的时机,也评估回答。

A streaming video assistant must decide when to speak, when to stay silent, and what to remember. SVI-Bench evaluates the interaction trajectory people actually experience.流式视频助手需要判断何时开口、何时沉默、记住什么。SVI-Bench 评估用户实际感受到的整段交互轨迹。

01

The benchmark spans 75 expert-authored items in nine scenarios and scores five user-observable dimensions, from proactive triggering to long-horizon memory.基准包含九类场景中的 75 个专家编写条目,评估从主动触发到长程记忆的五个用户可感知维度。

02

Speaking and silence are coupled within each item, so the score reflects whether the assistant made the right intervention decision in that situation.每个条目内联合评估开口与沉默,使分数反映助手在该情境中是否做出了正确的介入决策。

03

InteractFlow aligns each response with the visual state and time at which it occurred, preserving a replayable trace for inspection and rescoring.InteractFlow 将每次回答与当时可见的画面和时间同步,保留可复看、可重新评分的轨迹。

Video agentsTrajectory evaluationOpen-source pipeline

RESEARCH RECORD研究成果

In review & in print.在审稿件与公开成果。

A current record of manuscripts under review, peer-reviewed papers, and public technical reports.这里汇集目前在审的研究稿件、已发表论文与公开技术报告。

01 / MANUSCRIPTS UNDER REVIEW01 / 在审稿件6 × ICLR 2027 · 2 × AAAI 2027
ICLR 2027
UNDER REVIEW在审
ACTS: Controlling Evidence Acquisition Before Writing Questions for Search AgentsEvidence-first data synthesis for long-horizon search面向长程搜索的证据优先数据合成
ICLR 2027
UNDER REVIEW在审
CATER: Competence-Adaptive Training for Efficient ReasoningOnline competence routing, reinforcement learning, and concise self-distillation在线能力路由、强化学习与简洁自蒸馏
ICLR 2027
UNDER REVIEW在审
When Teachers Conflict with Verifiers: Support-Preserving On-Policy DistillationA verifier-constrained target distribution for on-policy distillation面向在线蒸馏的验证器约束目标分布
ICLR 2027
UNDER REVIEW在审
CrystalBindDiffusion attribute-binding diagnostics and inference-time guidance扩散模型属性绑定诊断与推理时引导
ICLR 2027
UNDER REVIEW在审
ClueFlowCo-first author · Persistent evidence, candidate state, and turn-level credit for search agents共同第一作者 · 搜索智能体的持久证据、候选状态与逐轮归因
ICLR 2027
UNDER REVIEW在审
SVI-Bench: Evaluating Human-Perceived Interaction Trajectories in Streaming Video SystemsCo-first author · Streaming-video benchmark and InteractFlow共同第一作者 · 流式视频基准与 InteractFlow 评测流程
AAAI 2027
UNDER REVIEW在审
TEALReadable interfaces for fixed multimodal features and token backbones面向固定多模态特征与 token 骨干的可读接口
AAAI 2027
UNDER REVIEW在审
SearchAuditor: Auditing and Attributing Failures in Long-Horizon Search AgentsThird author · Failure benchmark and evidence-grounded auditor第三作者 · 失败轨迹基准与证据锚定审计器
02 / PUBLISHED PAPERS & REPORTS02 / 已发表论文与公开报告2024—2026