Jiaxing Zhang

Jiaxing Zhang

Senior Research Scientist @ TikTok

Sitemap

A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.

Pages

AgeMem

Published:

Agent Memory 每日追踪：解读 ACL 2026 AgeMem，以及统一长期与短期记忆管理的 Agent 策略。

AgeMem

Published:

Unified long- and short-term memory management for language-model agents, with a close look at policy design and evaluation.

FARMA

Published:

追踪 LLM Agent 记忆系统的新攻击面：伪造推理记忆（FARMA）、相关安全风险，以及现有防御机制的演进。

FARMA

Published:

Forged reasoning memories expose a new attack surface in agent memory systems. A review of risks, evidence, and defenses.

GAM

Published:

追踪 ACL 2026 工作 GAM：通过事件进度图与主题关联网络，在快速适应和长期知识稳定之间取得平衡。

GAM

Published:

Event-progress graphs and thematic networks balance rapid adaptation with stable long-term knowledge.

EMO-R3

Published:

解析 EMO-R3 如何将推理形式与推理有效性共同纳入奖励设计，以应对主观视觉任务中的强化学习训练与评估难题。

EMO-R3

Published:

Rewarding both reasoning structure and effectiveness in subjective visual tasks, with implications for RL evaluation.

NEMORI

Published:

深度解析 NEMORI 如何以系统自身的预测误差作为记忆蒸馏判据，让 LLM Agent 自适应判断哪些经验值得保留。

NEMORI

Published:

Prediction error as a signal for adaptive memory distillation: deciding which experiences an agent should retain.

MAGMA

Published:

深入解析 MAGMA 如何用四张正交关系图组织 Agent 记忆，并审视其在 LoCoMo、LongMemEval 上的实验结果、贡献与局限。

MAGMA

Published:

Four complementary relation graphs organize agent memory. Evidence and limitations from LoCoMo and LongMemEval.

RnG

Published:

解析 CVPR 2026 Highlight 工作 RnG 如何以统一 Transformer 和 KV-Cache 隐式表示，从部分观测完成三维建模。

RnG

Published:

A unified Transformer uses implicit KV-cache representations to reconstruct 3D scenes from partial observations.

Agent-STAR

Published:

从五个维度解析 Agent-STAR 的长视野工具使用强化学习方案，包括课程数据、奖励设计、训练算法、泛化表现与成本权衡。

Agent-STAR

Published:

Long-horizon tool-use reinforcement learning through curriculum data, reward design, and generalization.

Experience-Following

Published:

梳理经验跟随行为的实证研究，分析记忆管理如何让 LLM Agent 的历史经验形成自我放大回路，以及更多记忆为何未必更好。

Experience-Following

Published:

How historical experience creates feedback loops in agent behavior, and why more memory does not always help.

OCR-Memory

Published:

解析 OCR-Memory 如何把长期记忆转移到视觉表示空间，以光学上下文检索缓解文本压缩中的容量、成本与信息保真矛盾。

OCR-Memory

Published:

Visual representations for long-term memory: optical context retrieval and the tradeoffs between capacity, cost, and fidelity.

SubSpec

Published:

深度解析 SubSpec 如何将目标模型自身的低比特副本用作草稿模型，为卸载式大模型提供无损、免训练的投机解码加速。

SubSpec

Published:

A low-bit copy of the target model serves as a training-free draft for lossless speculative decoding with offloaded models.

TIS

Published:

解析 ICLR 2026 Outstanding Paper 对 Transformer 表达简洁性的理论刻画，以及其相对其他网络架构的复杂度优势与成立条件。

TIS

Published:

Theoretical bounds on Transformer succinctness, architectural complexity advantages, and the conditions behind them.

MolGraphBench

Published:

梳理 MolGraphBench 对分子回归 GNN 的系统评测，揭示层类型选择并非常量，而是能够带来显著性能差异的关键超参数。

MolGraphBench

Published:

A systematic evaluation of molecular-regression GNNs reveals layer choice as a consequential hyperparameter.

GeoRA

Published:

解析 ACL 2026 Outstanding Paper GeoRA 如何依据梯度几何选择可训练参数，重新设计适用于可验证奖励强化学习的低秩适配。

GeoRA

Published:

Gradient geometry guides trainable-parameter selection for low-rank adaptation under verifiable reinforcement learning rewards.

Tempo

Published:

解析 Tempo 如何将视频压缩器设计为问题感知的小型 VLM，以相关性评分动态分配全局 token 预算并提升长视频理解效率。

Tempo

Published:

A question-aware video compressor allocates a global token budget to improve efficient long-video understanding.

EARL-BO

Published:

解析 EARL-BO 如何在高斯过程世界模型中用强化学习训练多步查询策略，使高维非短视贝叶斯优化避免显式展开场景树。

EARL-BO

Published:

Multistep query policies learned inside a Gaussian-process world model for nonmyopic Bayesian optimization.

DPLM-2.1

Published:

解析 DPLM-2.1 如何通过位级监督、几何偏置与结构表征对齐，改善多模态蛋白语言模型的三维结构学习。

DPLM-2.1

Published:

Bit-level supervision, geometric biases, and representation alignment improve multimodal protein structure learning.

HEX

Published:

解析 HEX 如何将跨块长共识引入扩散语言模型，以隐式半自回归专家实现测试时扩展，并审视其计算成本。

HEX

Published:

Cross-block consensus enables test-time scaling in diffusion language models through implicit semi-autoregressive experts.

MMedPO

Published:

解析 MMedPO 如何以病灶局部扰动和临床相关性加权，让医疗多模态偏好优化同时比较回答质量与影像证据。

MMedPO

Published:

Lesion-level perturbations and clinical relevance connect medical preference optimization to visual evidence.

DM/FM Alignment

Published:

解析奖励加权分布如何统一扩散模型与流匹配模型的对齐，在无需针对每个奖励重训模型的情况下修正采样动力学。

DM/FM Alignment

Published:

Reward-weighted distributions unify diffusion and flow-matching alignment through changes to sampling dynamics.

SimbaV2

Published:

解析 SimbaV2 如何通过超球面归一化与尺度稳定的分布式价值分类，提升深度强化学习的可扩展性与训练稳定性。

SimbaV2

Published:

Hyperspherical normalization and scale-stable distributional value prediction improve deep RL scaling and stability.

PARQ

Published:

解析 PARQ 如何以分段仿射正则和聚合近端更新取代直通估计器，让量化训练从连续权重收缩到真正的离散权重。

PARQ

Published:

Piecewise-affine regularization and proximal updates move quantization training from continuous to discrete weights.

LIMEFLDL

Published:

解析 LIMEFLDL 如何用受概率单纯形约束的归因矩阵，一次性解释结构化输出模型的完整标签分布。

LIMEFLDL

Published:

A simplex-constrained attribution matrix explains the complete label distribution of structured-output models.

KLDM

Published:

解析 KLDM 如何引入欧氏辅助速度，将周期晶体坐标上的流形扩散改写为适合生成建模的动力系统。

KLDM

Published:

Auxiliary velocities turn diffusion on periodic crystal coordinates into a dynamical system for crystal generation.

Alignment Potential

Published:

解析 Alignment Potential 如何依据奖励模型与当前策略之间的教师—学生差距，为 LLM 对齐选择更有价值的偏好数据。

Alignment Potential

Published:

The teacher–student gap between reward models and current policies guides preference-data selection for alignment.

ARFT / AutoResearchEval

Published:

解析 AutoResearchEval 如何从科学 Agent 的最终结果转向过程与产物诊断，并通过 800 条轨迹揭示发现缺陷却不修订的高频失败闭环。

ARFT / AutoResearchEval

Published:

Process and artifact diagnostics across 800 scientific-agent trajectories reveal failures to revise detected flaws.

VGA

Published:

解析 Visual Graph Arena 如何评测多模态模型从像素恢复图结构的能力，并审视布局鲁棒性、测试泄漏与感知—推理解耦问题。

VGA

Published:

Visual Graph Arena tests graph recovery from pixels, layout robustness, and the separation of perception from reasoning.

Q-VDiT

Published:

解析 Q-VDiT 如何通过 token-aware 低秩补偿与全视频关系蒸馏，缓解低比特视频扩散 Transformer 的质量崩溃。

Q-VDiT

Published:

Token-aware low-rank compensation and video-relation distillation mitigate low-bit video Transformer degradation.

TARL

Published:

解析 TARL 如何以动作熵最小化和原策略 KL 约束进行无奖励测试时策略适配，并审视环境偏移与实现一致性风险。

TARL

Published:

Reward-free test-time policy adaptation through action-entropy minimization and KL constraints under environment shifts.

QServe

Published:

解析 QServe 如何通过 W4A8KV4 量化与 GPU 内核协同设计，将解量化移出关键路径并提升大模型推理吞吐。

QServe

Published:

W4A8KV4 quantization and GPU kernel co-design remove dequantization bottlenecks to improve inference throughput.

DocVXQA

Published:

解析 DocVXQA 如何以充分性、最小性与 token interaction 生成文档问答解释，并审视答案条件化训练对解释忠实度的影响。

DocVXQA

Published:

Sufficiency, minimality, and token interactions for document-QA explanations, with a focus on faithfulness.

TIMING

Published:

How TIMING rewrites time-series attribution with randomized integration paths over contiguous segments, and what its metrics, completeness trade-offs, and evidence boundaries reveal.

TIMING

Published:

AVUT

Published:

How AVUT filters textual shortcuts to build a video QA benchmark that depends on audiovisual evidence—and where audio attribution remains uncertain.

AVUT

Published:

LongLive

Published:

How LongLive combines streaming long tuning, frame sinks, and KV-recache for sustainable long-video generation—and where quality, semantics, and speed still trade off.

LongLive

Published:

How LongLive combines streaming long tuning, frame sinks, and KV-recache for sustainable long-video generation—and where quality, semantics, and speed still trade off.

DreamGen

Published:

从具身微调、伪动作恢复到低数据真机泛化 DreamGen 证明视频生成可产生可执行训练信号，但 81,000 GPU-hour、learner-dependent 负迁移与统计不足仍阻止“低成本通用数据引擎”的强结论。

DreamGen

Published:

DreamGen shows that video generation can produce executable training signals. However, 81,000 GPU-hours, learner-dependent negative transfer, and insufficient statistical evidence still preclude the stronger claim of a low-cost, general-purpose data engine.

Lumos

Published:

Kineto trace × 因果执行图 × 反事实配置重放 Lumos 将平均重放误差从 14.0% 降至 3.3%，但未公开的 kernel 时延模型与缺失的 TP/OOM 建模限制了独立复现和全局规划。

Lumos

Published:

Lumos reduces mean replay error from 14.0% to 3.3%, but its unpublished kernel-latency model and missing TP/OOM modeling limit independent reproduction and global planning.

Emergent Misalignment

Published:

隐含漏洞代码 × 同输出异动机 × 跨域行为指纹 隐含漏洞代码微调使主自由问答风险增量达 27.0%，但低 StrongREJECT、教育语境归零和弱开放模型复现表明它不同于普通 jailbreak，也尚不足以证明稳定“邪恶人格”。

Emergent Misalignment

Published:

Fine-tuning on implicitly vulnerable code raises risk on the main free-response questions by 27.0%. Yet low StrongREJECT, zero main-set risk under educational framing, and weak replication in open models distinguish it from ordinary jailbreaking and do not establish a stable malicious persona.

FMG

Published:

图像语义选择 × HRG 形式护栏 × 可合成性—覆盖权衡 FMG 相对 DEG 将三项 Retro* 提高 31.6–49.1 pp，但 diversity 下降 8.6%–46.5%；它是受形式系统约束的语义启发式，不是已验证的自动化学家。

FMG

Published:

FMG improves three Retro* scores over DEG by 31.6–49.1 pp, but diversity drops by 8.6%–46.5%. It is a semantically guided heuristic constrained by a formal system, not a validated automated chemist.

ATLANTIS

Published:

弱到强概率位移 × 连续数据加权 × 理论换测度审计 ATLANTIS 在主表 23/24 配对上升、最大 +7.19 pp，但单 seed、权重方差未知与 p_r≠实际数据分布使重要性采样解释尚未闭环。

ATLANTIS

Published:

ATLANTIS improves 23/24 main-table comparisons, by up to +7.19 pp. Yet a single seed, unknown weight variance, and a mismatch between p_r and the actual data distribution leave the importance-sampling interpretation incomplete.

DEFAME

Published:

多模态 RAG × 动态工具规划 × 时间外事实核查审计 CR+ 相对直接 GPT‑4o +34.5 pp，但 VERITE T/OOC 仍低 AITR 4.3 pp；检索覆盖是主因，长推理不是万能答案。

DEFAME

Published:

CR+ improves by 34.5 pp over direct GPT-4o, but VERITE T/OOC remains 4.3 pp below AITR. Evidence coverage is the main driver; longer reasoning is not a universal remedy.

Ambient Creativity

Published:

扩散记忆 × Ambient Score Matching × 小样本创造性 CIFAR-10/300 高相似复制率 56.56%→9.64%，FID 25.10→23.91；但文本条件下单独收益很小，且无端到端隐私保证。

Ambient Creativity

Published:

On CIFAR-10/300, high-similarity copying falls from 56.56% to 9.64% and FID from 25.10 to 23.91. Standalone text-conditioned gains are small, however, and there is no end-to-end privacy guarantee.

VIKI-R

Published:

分层具身协作基准 × CoT 冷启动 × GRPO 规则奖励 7B 的 L2 OOD 从 25.62% 提升至 33.25%，但 L2 ID 下降 1.67 pp，且真实世界表格没有测试 VIKI-R。

VIKI-R

Published:

The 7B model improves L2 OOD from 25.62% to 33.25%, but L2 ID drops 1.67 pp, and the real-world table does not evaluate VIKI-R.

ScaleSearch

Published:

NVFP4 邻域定标搜索 × FP4 attention × 4.5-bit KV cache Qwen3-8B MATH-500 较普通 NVFP4 提高 15.0 pp；但 SSA 消融显示 scale 搜索本身只带来 0.0047 PPL，主要保真来自高精度首尾 KV。

ScaleSearch

Published:

Qwen3-8B MATH-500 improves 15.0 pp over ordinary NVFP4, but SSA ablation attributes only 0.0047 PPL to scale search itself; high-precision boundary KV blocks preserve most fidelity.

CoordRobustLR

Published:

Adversarial MNAR × Minimax Phase Diagram × Coordinate Coupling 在 N(0,I) 线性回归中，每列仅 η 比例的自适应缺失就足以达到任意替换的 minimax 难度，并在 η≈Θ(1/√d) 处使零估计器成为量级最优。

CoordRobustLR

Published:

In N(0,I) linear regression, adaptive missingness in only an η fraction of each column matches arbitrary replacement’s minimax difficulty; at η≈Θ(1/√d), the zero estimator becomes order-optimal.

PemNN

Published:

WLC 物理坐标 × 双支路融合 × 小样本单分子筛选 约三十条训练曲线时，PemNN 达到 79.61±5.24%，较最强基线高 11.35 pp；但物理贡献与额外 extension 输入仍未分离。

PemNN

Published:

With about thirty training curves, PemNN reaches 79.61±5.24%, beating the strongest baseline by 11.35 pp; physical knowledge remains confounded with extra extension input.

ThinkSLM

Published:

训练来源、提示策略、压缩路径与鲁棒性的六轴矩阵 8-bit 量化通常保住既有推理能力；结构化剪枝则可能让中间推理从可用直接跨入失效。

ThinkSLM

Published:

Eight-bit quantization usually preserves existing reasoning, while structured pruning can push intermediate reasoning from usable to failure.

MM-RLHF

Published:

人工细粒度偏好 × Critique Reward × 动态 β 的证据审计 Unsafety 40.2→13.9 是强证据；但外部 reward 平均仍低 GPT‑4o 12.25 分，且公开数据规模有三套冲突口径。

MM-RLHF

Published:

Unsafety 40.2→13.9 is strong evidence, but external reward scores still average 12.25 points below GPT-4o, and public dataset counts conflict across three sources.

MeanFlow

Published:

平均速度恒等式 × JVP 自举 × 1‑NFE 生成 1‑NFE FID 从 Shortcut 的 10.60 降至 3.43，但 CFG 贡献巨大、统计检验缺失，且论文存在两处内部数字冲突。

MeanFlow

Published:

One-NFE FID drops from Shortcut’s 10.60 to 3.43, but CFG contributes heavily, significance tests are absent, and two internal numerical inconsistencies remain.

Embodied-R1

Published:

Embodied-R1 深度研究报告：方法、定量结果、相关工作与局限性。

Embodied-R1

Published:

Embodied-R1 in-depth report: methods, quantitative results, related work, and limitations.

Ada-KV

Published:

全局 Top‑B 分配 × 逐头 eviction × variable‑length KV layout B=512 的 unknown-query LongBench 平均 +1.62 pp，但系统证据、统计区间与 safeguard 配置一致性仍不足。

Ada-KV

Published:

At B=512, unknown-query LongBench averages improve 1.62 pp, but system evidence, statistical intervals, and safeguard consistency remain insufficient.

Tilted Empirical Risk

Published:

负倾斜风险如何在重尾、污染与泛化之间做账 负倾斜确有强鲁棒潜力，但当前数据驱动选择仍依赖干净分布信息；理论可选与部署可选之间尚未闭合。

Tilted Empirical Risk

Published:

Negative tilting offers strong robustness, but current data-driven selection still needs clean-distribution information; theoretical and deployable selection remain distinct.

MolJO

Published:

连续坐标、离散类型与 backward correction 如何共同优化分子 MolJO 把联合梯度变成了可用的分子优化器，但 docking 代理、尺寸偏置与搜索预算意味着“分数更好”仍不是“药物更好”。

MolJO

Published:

MolJO makes joint gradients useful for molecular optimization, but docking proxies, size bias, and search budgets mean better scores are not yet better drugs.

HPS

Published:

Hard Preference Sampling 的理论、证据与安全含义 HPS 用外部排序器从多负例中抽一个“最像正例”的负例，在几乎不损害外部质量分的同时显著拉大隐式奖励间隔并节省训练时间；但 margin 指标与目标函数高度同构，安全结论仍需独立红队与单样本部署评测闭环。

HPS

Published:

HPS samples the rejected response most similar to the preferred one using an external ranker, widening implicit reward margins and saving training time while preserving external quality scores. But margin metrics closely mirror the objective; safety still needs independent red teaming and single-sample deployment evaluation.

BridgeVLA

Published:

输入—输出空间对齐的方程、证据链与工程代价 BridgeVLA 用三视图 2D 热图把 VLM 的视觉先验与 3D 动作关键点接到同一空间，在 RLBench、COLOSSEUM 与真实机器人上取得强结果；但 960 H100·小时、外部运动规划器与长程任务零成功率限定了“高效通用策略”的外推。

BridgeVLA

Published:

BridgeVLA connects VLM visual priors and 3D action keypoints in one space through three-view 2D heatmaps, achieving strong results on RLBench, COLOSSEUM, and real robots. However, 960 H100-hours, an external motion planner, and zero success on long-horizon tasks limit extrapolation to an “efficient general-purpose policy.”

Seesaw

Published:

Dynamic model re-sharding × tiered KV buffer × asynchronous transfer 同一组 GPU 在 prefill 与 decode 之间改变 TP/PP，可把 PCIe 系统平均吞吐提高 36%，但收益明确以延迟、CPU 内存和可复现性为代价。

Seesaw

Published:

Changing TP/PP between prefill and decode on the same GPUs raises average PCIe-system throughput by 36%, but explicitly trades off latency, CPU memory, and reproducibility.

Circuit Insights

Published:

自动解释不应只写出流畅标签，还要说明固定权重、输入路径与输出后果为何一致。 WeightLens 的平均 Clarity 较 MaxAct* 提升 27.05%，但 Purity 下降 20.00%；所有方法的 Faithfulness 仍约 0.03–0.04，说明“可读说明”与“因果可信”之间仍有明显断层。

Circuit Insights

Published:

WeightLens improves mean Clarity over MaxAct* by 27.05%, but reduces Purity by 20.00%. Faithfulness remains around 0.03–0.04 for every method, showing a substantial gap between readable descriptions and causal credibility.

I2Mole

Published:

显式跨分子边找到相互作用核心，离散环境重组检验预测能否跨药物迁移。 Transductive ACC 只领先最强基线 0.42–0.66 pp，但去掉 VQ 环境模块会损失 14.12 pp；最严格 Zhang Type 2 的 F1 仍仅 39.95。

I2Mole

Published:

Transductive ACC exceeds the strongest baseline by only 0.42–0.66 pp, but removing the VQ environment module costs 14.12 pp; F1 on the strictest Zhang Type 2 remains only 39.95.

RGM

Published:

共享骨干、轻量评分头、结果—伪标签一致性过滤：外部测试时扩展不再需要第二个 72B 模型。 QwQ-32B + 54M SPRM 在五项上比同基座 PRM-72B 高 0.9–1.6 pp；但完整 8 路生成的服务成本与真正过程判错能力仍未闭合。

RGM

Published:

QwQ-32B +54M SPRM exceeds a same-base PRM-72B by 0.9–1.6 pp on five tasks, but complete eight-path serving cost and actual process-error detection remain unresolved.

Spatial IDs

Published:

从线性空间 ID 到因果 steering：VLM 如何绑定对象、位置与时间 多个 VLM 在中间层形成近似低秩、可操纵的对象—位置绑定，但最强证据仍局限于简单二维关系。

Spatial IDs

Published:

Multiple VLMs develop approximately low-rank, manipulable object–position bindings in middle layers, but the strongest evidence remains limited to simple 2D relations.

DFM

Published:

把渐进生成压缩成“多尺度表示 + 多时间轴 + 单共享 DiT” ImageNet-1K 512px 的最佳 FDD 从普通 Flow Matching 的 132.5 降至 85.8；简单接口有效，但高频伪影、内部数据和非忠实公开实现仍是复现边界。

DFM

Published:

Best ImageNet-1K512px FDD drops from ordinary Flow Matching’s132.5 to 85.8. The simple interface works, but high-frequency artifacts, internal data, and a non-faithful public implementation limit reproduction.

PointMapPolicy

Published:

把稠密三维几何重新装回二维规则网格 PMP-xyz 将 RoboCasa 平均成功率从 DP3 的 27.04% 提至 49.12%，而多模态 PMP 在 CALVIN 达 4.01；几何与颜色的价值随任务语义发生反转。

PointMapPolicy

Published:

PMP-xyz raises mean RoboCasa success from DP3’s27.04% to 49.12%, while multimodal PMP reaches 4.01 on CALVIN. The value of geometry versus color reverses with task semantics.

SuperInfer

Published:

把“换出 KV”从内存补丁变成 SLO 调度原语 DuplexKV 将 16 GB 双向 KV 迁移从 1556.15 ms 压至 46.80 ms；在 GH200 高负载下，SuperInfer 的 TTFT SLO 达标率最高提高 74.7 个百分点，但硬件外推、Pie 对比与统计区间仍缺失。

SuperInfer

Published:

DuplexKV reduces16GB bidirectional KV migration from 1556.15ms to 46.80ms. Under high GH200 load, SuperInfer improves TTFT SLO attainment by up to 74.7 pp, but cross-hardware evidence, Pie comparisons, and statistical intervals remain missing.

AmongUs

Published:

把欺骗从静态问答搬进有身份、有行动后果的社会推理沙盒 18 个模型、2,054 局开放式博弈显示，推理与强化学习会更快提高欺骗能力，而非同速提高识别能力；白盒探针可跨域达到 0.94/0.99 AUROC，但显式角色上下文可把检测推至 1.000，SAE steering 仍无法可靠控制行为。

AmongUs

Published:

Open-ended games across 18 models and 2,054 matches show reasoning/RL improving deception faster than detection. White-box probes transfer at 0.94/0.99 AUROC, but explicit role context can raise detection to 1.000, while SAE steering still fails to reliably control behavior.

Pallatom

Published:

把序列—结构共设计推进到全原子联合生成 Atom14 坐标扩散与原子回收把 CO-DESIGN DES-aa 从 MultiFlow 的 62.74% 提升到 85.03%，并将结构簇从 134 增至 291；但 L=400 时被 CarbonNovo 反超 20.9 pp，高步长出现 92% 全螺旋偏置，且没有湿实验。

Pallatom

Published:

Atom14 coordinate diffusion/recycling raises CO-DESIGN DES-aa from MultiFlow62.74% to 85.03% and structural clusters from 134 to 291. Yet CarbonNovo leads by 20.9 pp at L=400, large step sizes induce92% all-helical bias, and wet-lab validation is absent.

ReActR

Published:

把错误轨迹变成可训练的反思能力 六类错误、困难度课程与分段反思奖励，使 Qwen3-4B 的五基准 Overall 从 LEMMA 的 52.4 提升到 57.3；但核心“反思”没有被直接测量，代码、数据、权重与统计区间也未公开。

ReActR

Published:

Six error types, difficulty curricula, and staged reflection rewards raise Qwen3-4B five-benchmark Overall from LEMMA52.4 to 57.3. Yet reflection is not directly measured; code, data, weights, and statistical intervals are unreleased.

AIM-CoT

Published:

AIM-CoT 深度研究报告：方法、定量结果、相关工作与局限性。

AIM-CoT

Published:

AIM-CoT in-depth report: methods, quantitative results, related work, and limitations.

Transition Matching

Published:

把“无穷小流”改造成“可学习的概率跳步” DTM 以随机有限步转移保留条件分布宽度，在同构 1.7B 对照中把 H100 采样从 10.8 秒降至 1.6 秒；但 head 计算、统计区间与可复现性仍是关键缺口。

Transition Matching

Published:

DTM preserves conditional-distribution width through stochastic finite-step transitions, reducing H100 sampling from 10.8 to 1.6 seconds in a matched 1.7B comparison. Head computation, statistical intervals, and reproducibility remain key gaps.

LRAM

Published:

速度赢了，真实机器人证据仍缺席 现代递归骨干在长历史与并行控制上很有吸引力，但收益取决于 action conditioning、kernel 与任务可观测性；它不是对 Transformer 的无条件替代。

LRAM

Published:

Modern recurrent backbones are attractive for long histories and parallel control, but gains depend on action conditioning, kernels, and observability. They are not unconditional Transformer replacements.

RSD

Published:

FLOPs 节省成立，服务级加速仍待证明 固定阈值把 7B 六项平均从 65.3 提至 66.1；72B MATH-500 以论文代理的 4.4× 更少 FLOPs 从 85.6 提至 88.0，但三服务架构、重复 PRM 前缀评分与统计缺失使真实 P99 结论仍然空白。

RSD

Published:

A fixed threshold raises the 7B six-benchmark average from 65.3 to 66.1. On 72B MATH-500, accuracy rises from 85.6 to 88.0 with 4.4× fewer FLOPs under the paper’s proxy. Yet three services, repeated PRM prefix scoring, and missing statistics leave real P99 performance unknown.

INSID3

Published:

每日 ML/AI 深度报告：解读 CVPR 2026 Oral 工作 INSID3 与 training-free in-context segmentation。

INSID3

Published:

Training-free in-context segmentation: methods, quantitative evidence, and limitations from a CVPR 2026 Oral paper.

Graph of Tokens

Published:

深度分析 Graph of Tokens 如何利用 token 间关系缓解稀疏混合专家模型的路由波动、负载均衡与扩展性问题。

Graph of Tokens

Published:

Token relationships stabilize routing and load balancing in sparse mixture-of-experts models.

GDPT

Published:

从维度相位转移视角解析神经网络由记忆走向泛化的 Grokking 现象，并梳理相关理论、实验设定与可信性问题。

GDPT

Published:

Grokking as a dimensional phase transition, connecting memorization, generalization, and experimental evidence.

Evo 2

Published:

深入解析 Evo 2 的长上下文基因组建模能力、训练与评测体系，以及统一生物基础模型对基因组理解和设计的意义。

Evo 2

Published:

Long-context genomic modeling and unified biological foundation models for genome understanding and design.

ETM

Published:

分析大模型推理轨迹在提供可解释性的同时，如何诱发解释过度、错误信任与评估偏差，并梳理相关研究方向。

ETM

Published:

How reasoning traces can encourage overinterpretation, misplaced trust, and evaluation bias.

Posts

portfolio

publications

Commit-level, Neural Vulnerability Detection and Assessment

Published in ESEC/FSE 2023, 2023

Commit-level vulnerability detection and CVSS assessment via context-aware graph learning.

Recommended citation: Yi Li, Aashish Yadavally, Jiaxing Zhang, Shaohua Wang, Tien N. Nguyen. 2023. Commit-level, Neural Vulnerability Detection and Assessment. Proceedings of the 31st ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE 2023). https://dl.acm.org/doi/10.1145/3611643.3616249

DeMinify: Neural Variable Name Recovery and Type Inference

Published in ESEC/FSE 2023, 2023

Neural recovery of variable names and type inference from minified code.

Recommended citation: Yi Li, Aashish Yadavally, Jiaxing Zhang, Shaohua Wang, Tien N. Nguyen. 2023. DeMinify: Neural Variable Name Recovery and Type Inference. Proceedings of the 31st ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE 2023). https://dl.acm.org/doi/10.1145/3611643.3616232

Interpreting Graph Neural Networks with In-Distributed Proxies

Published in ICML 2024, 2024

Generates in-distribution proxy graphs to improve faithfulness of GNN explanations.

Recommended citation: Zhuomin Chen, Jiaxing Zhang, Jingchao Ni, Xiaoting Li, Yuchen Bian, Md Mezbahul Islam, Ananda Mohan Mondal, Hua Wei, Dongsheng Luo. 2024. Interpreting Graph Neural Networks with In-Distributed Proxies. International Conference on Machine Learning (ICML 2024). https://arxiv.org/abs/2402.02036

LLMExplainer: Large Language Model based Bayesian Inference for Graph Explanation Generation

Published in AIAgent4IR 2025 (in conjunction with KDD), 2024

Uses LLM-guided Bayesian inference to mitigate learning bias in graph explanation.

Recommended citation: Jiaxing Zhang, Jiayi Liu, Dongsheng Luo, Jennifer Neville, Hua Wei. 2025. LLMExplainer: Large Language Model based Bayesian Inference for Graph Explanation Generation. AIAgent4IR 2025 Workshop (in conjunction with KDD 2025). https://arxiv.org/abs/2407.15351

RegExplainer: Generating Explanations for Graph Neural Networks in Regression Tasks

Published in Advances in Neural Information Processing Systems (NeurIPS 2024), 2024

An explanation framework for graph regression models with improved reliability under distribution shift.

Recommended citation: Jiaxing Zhang, Zhuomin Chen, Hao Mei, Longchao Da, Dongsheng Luo, Hua Wei. 2024. RegExplainer: Generating Explanations for Graph Neural Networks in Regression Tasks. Advances in Neural Information Processing Systems 37 (NeurIPS 2024), 79282-79306. https://proceedings.neurips.cc/paper_files/paper/2024/hash/909f526db5127f8bd8158af32d9e313a-Abstract-Conference.html

RB-FT: Rationale-Bootstrapped Fine-Tuning for Video Classification

Published in arXiv preprint, 2025

A two-stage self-improvement framework that uses model-generated rationales to adapt vision-language models to domain-specific video classification without additional annotations.

Recommended citation: Meilong Xu, Di Fu, Jiaxing Zhang, Gong Yu, Jiayu Zheng, Xiaoling Hu, Dongdi Zhao, Feiyang Li, Chao Chen, Yong Cao. 2025. RB-FT: Rationale-Bootstrapped Fine-Tuning for Video Classification. arXiv:2511.15923. https://arxiv.org/abs/2511.15923

Shot-Aware Frame Sampling for Video Understanding

Published in arXiv preprint, 2026

InfoShot is a training-free, shot-aware sampler that preserves both representative context and brief decision-critical events in long videos.

Recommended citation: Mengyu Zhao, Di Fu, Yongyu Xie, Jiaxing Zhang, Zhigang Yuan, Shirin Jalali, Yong Cao. 2026. Shot-Aware Frame Sampling for Video Understanding. arXiv:2603.17374. https://arxiv.org/abs/2603.17374

Addressing Structural Distribution Shift in Explanations for Graph Neural Networks

Published in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2026

A theoretical and empirical treatment of structural distribution shift in GNN explanations, using proxy graphs to preserve explanatory information while remaining in distribution.

Recommended citation: Zhuomin Chen, Hojat Allah Salehi, Esteban Schafir, Xu Zheng, Jiaxing Zhang, Hua Wei, Jingchao Ni, Farhad Shirani, Dongsheng Luo. 2026. Addressing Structural Distribution Shift in Explanations for Graph Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence. doi:10.1109/TPAMI.2026.3690304. https://doi.org/10.1109/TPAMI.2026.3690304

HP-JEPA: Hierarchical Partitioning for Multi-Resolution Graph Joint-Embedding Predictive Learning

Published in arXiv preprint, 2026

Multi-resolution latent prediction learns transferable graph representations across local and global structural scales.

Recommended citation: Ruichen Xu, Jingxiang Qu, Wenhan Gao, Jiaxing Zhang, Linsey Pang, Ravid Shwartz-Ziv, Yann LeCun, Yuefan Deng. 2026. HP-JEPA: Hierarchical Partitioning for Multi-Resolution Graph Joint-Embedding Predictive Learning. arXiv:2608.00491. https://arxiv.org/abs/2608.00491

PDE-OBS: Controlled Evaluation Across Observation Patterns

Published in arXiv preprint, 2026

A benchmark for evaluating physical-field reconstruction and forecasting as observation density and spatial layout change.

Recommended citation: Ruichen Xu, Siyao Wang, Fang Wan, Jiacheng Qiu, Wenhan Gao, Jiaxing Zhang, Linsey Pang, Ravid Shwartz-Ziv, Yann LeCun, Yuefan Deng. 2026. PDE-OBS: Controlled Evaluation Across Observation Patterns. arXiv:2609.36521. https://arxiv.org/abs/2609.36521

SCOPE: Observation-Conditioned Full-Target Prediction for Sparse PDE Inference

Published in arXiv preprint, 2026

Single-pass recovery of complete PDE fields from sparse measurements through latent prediction and physical reconstruction.

Recommended citation: Ruichen Xu, Siyao Wang, Fang Wan, Jiacheng Qiu, Wenhan Gao, Jiaxing Zhang, Linsey Pang, Ravid Shwartz-Ziv, Prakhar Mehrotra, Yann LeCun, Yuefan Deng. 2026. SCOPE: Observation-Conditioned Full-Target Prediction for Sparse PDE Inference. arXiv:2609.36527. https://arxiv.org/abs/2609.36527

Alignment in Multimodal Large Language Models: A Survey

Published in Preprints.org, 2026

A review of multimodal alignment spanning evidence use, human intent, training methods, and evaluation.

Recommended citation: Yuyao Yang, Jiaxing Zhang, Ling Yue, Linsey Pang, Zhiwei Liu, Zixian Ma, Zheng Zhao, Jialiang Yin, Mingdai Yang, Huayi Yang, Lucas Ni, Yi-Fan Zhang, Connor Lee, Jiayuan Ding, Ranjay Krishna, Bin Shi, Zenglin Xu, Yan Liu, Qingsong Wen, Philip S. Yu. 2026. Alignment in Multimodal Large Language Models: A Survey. Preprints.org, 202610.1245 (version 2). https://www.preprints.org/manuscript/202610.1245

talks

teaching

IS218: Web Application Development

Undergraduate course, New Jersey Institute of Technology, Department of Information Systems, 2024

I taught a course on web application development using HTML, Python, Docker and Flask. I held office hours, graded assignments, and led lab sessions.