Jianmin Wu

ByteDance Seed, Jiaze Chen, Tiantian Fan et al. · bytedance

We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.

Guanghao Zhu, Zeyu Liu, Zhitian Hou et al.

Large-scale biomedical image-text datasets extracted from scientific literature provide valuable resources for medical multimodal model training. These datasets are commonly organized as image-caption pairs; however, figure captions are often short, context-dependent, and only partially informative without the surrounding article text. At the same time, large-scale automatic extraction introduces structural noise such as missing captions, residual markup, duplicated context, and incoherent multi-paragraph figure descriptions. We revisit data construction for medical multimodal continued pretraining (CPT) and present PMC-InterCPT, a context-grounded biomedical interleaved corpus that incorporates figure-referencing body text in addition to captions. Our pipeline recovers missing captions, cleans caption and context text, reconstructs coherent interleaved image-text samples, and applies LLM-supervised medical relevance and quality classifiers to filter noisy records. We further reveal strong modality imbalance in the resulting corpus and introduce a four-bucket evidence taxonomy for modality-aware resampling. Through CPT followed by supervised fine-tuning (SFT) on Qwen3.5-4B-Base, PMC-InterCPT effectively improves medical and general multimodal performance while using fewer CPT tokens than the raw source pool. The experimental results also illustrate the complementarity between the data quality and modality for medical multimodal CPT.

Yuanyi Wang, Yanggan Gu, Su Lu et al.

Weight-space model merging is usually formulated as an algebraic operation on checkpoints, yet at LLM scale the limiting resource is often the set of expert weights that must be read. We introduce MergePipe, a budget-aware execution layer that casts LLM merging as an \emph{expert access-set} problem: given a merge operator and a checkpoint family in a shared weight coordinate system, choose which expert delta blocks to access under an explicit I/O budget. MergePipe indexes parameter blocks, builds deterministic access plans, and executes the induced budgeted merge with replayable manifests. The plan is budget-sound by construction and recovers the full-read merge at full budget; for fixed-coefficient additive operators, the omitted-update error is bounded by the norm of omitted deltas. Across Qwen and Llama merging workloads, MergePipe reduces expert-read I/O by up to an order of magnitude and achieves up to $11\times$ speedups. Representative budget sweeps show $O(10^{-3})$ parameter deviation from full-read merges and no monotonic degradation on downstream benchmarks.

Yuanyi Wang, Su Lu, Yanggan Gu et al.

On-policy distillation (OPD) trains a student on its own rollouts with token-level teacher supervision. Recent selective OPD methods exploit the non-uniformity of OPD signals by prioritizing high-entropy or high-disagreement tokens. We revisit this principle and ask: which token-level teacher signals are actually learnable? Using a fixed-context diagnostic that measures same-context teacher-student KL reduction, we show that raw KL disagreement is a coarse proxy for learning value. It conflates learnable disagreement, where the teacher assigns corrective mass to the student's top-K candidates, with incompatible disagreement, where the teacher places mass mostly off the student's current support. We formalize this local compatibility as token teachability and show that it better predicts fixed-context improvement than raw KL alone. Motivated by this finding, we propose Teachability-Aware OPD (TA-OPD), a lightweight token-position selection method that applies OPD loss to high-teachability positions without reward models or verifiers. Across Qwen2.5 and Qwen 3 teacher-student settings, TA-OPD often surpasses full-token OPD with only 5% retained tokens and improves over entropy- and divergence-based baselines. Our results reframe selective OPD as selecting learnable teacher signals rather than merely salient tokens.

Wenyu Zhang, Guoliang You, Tianlun et al.

Coding agents are increasingly used as application builders, yet many evaluations still focus on source code, repository-level tests, or intermediate traces rather than the delivered application. We introduce WebGameBench, a requirement-to-application benchmark that evaluates whether coding agents can turn a frozen Structured WebGame Specification into a browser-accessible game. Browser-native games provide a compact but behavior-dense testbed: even simple games require coordinated input handling, spatial mapping, rule execution, state transitions, terminal conditions, restart behavior, and visible feedback. In WebGameBench, each generated artifact is built, served, and exposed as a browser-accessible application under a unified deployment protocol. A runtime evaluator then interacts with the delivered game in a real browser and assigns a three-way label: EXCELLENT, USABLE, or UNUSABLE. On a human-reviewed subset, the runtime label is broadly aligned with human gameplay review under the Usable-rate criterion. Across 111 tasks, 12 coding agents, and 14 evaluation configurations, WebGameBench separates current systems: the best configuration reaches a 76.9% usable rate but only a 20.2% excellent rate. This gap shows that crossing the minimum playable-delivery threshold is still far from complete requirement satisfaction. To our knowledge, WebGameBench is the first requirement-to-application benchmark for browser-native game delivery that validates delivered-application runtime labels against independent human gameplay review under the Usable-rate criterion.

Wenjun Wang, Yanggan Gu, Shuo Cai et al.

Low-resource deployment constraints have made model quantization essential for deploying neural networks while preserving performance. Meanwhile, model merging has become an increasingly practical low-resource strategy for integrating multiple task- or domain-specialized experts into a single model without joint training or multi-model serving. Together, quantization and model merging enable an efficient low-resource deployment pipeline by integrating multiple experts into one low-bit model. We formulate this setting as Post-Merge Quantization (PMQ). We show that directly applying post-training quantization (PTQ) to a merged model is unreliable because two distinct deviations are coupled: the quantization deviation introduced by low-bit reconstruction and the expert-relative merging deviation inherited from model merging. To mitigate these deviations, we propose E-PMQ, an expert-guided PMQ framework that uses source expert weights to provide expert- guided output targets during layer-wise calibration, together with merged-weight anchoring to stabilize the calibration and preserve the integrated behavior of the merged model. On CLIP-ViT-B/32 eight-task merging, E-PMQ improves 4-bit GPTQ from 65.0% to 73.6% under Task Arithmetic and from 69.1% to 74.8% under TIES-Merging. On harder settings, E-PMQ improves GPTQ from 34.8% to 76.7% on 20-task CLIP-ViT-L/14 and from 78.26% to 83.34% on FLAN-T5- base GLUE. These results demonstrate that E-PMQ enables effective post-merge quantization and low-bit deployment.

Shupeng Li, Weipeng Lu, Linyun Liu et al.

Domain-specific enhancement of Large Language Models (LLMs) within the financial context has long been a focal point of industrial application. While previous models such as BloombergGPT and Baichuan-Finance primarily focused on knowledge enhancement, the deepening complexity of financial services has driven a growing demand for models that possess not only domain knowledge but also robust financial reasoning and agentic capabilities. In this paper, we present QianfanHuijin, a financial domain LLM, and propose a generalizable multi-stage training paradigm for industrial model enhancement. Our approach begins with Continual Pre-training (CPT) on financial corpora to consolidate the knowledge base. This is followed by a fine-grained Post-training pipeline designed with increasing specificity: starting with Financial SFT, progressing to Finance Reasoning RL and Finance Agentic RL, and culminating in General RL aligned with real-world business scenarios. Empirical results demonstrate that QianfanHuijin achieves superior performance across various authoritative financial benchmarks. Furthermore, ablation studies confirm that the targeted Reasoning RL and Agentic RL stages yield significant gains in their respective capabilities. These findings validate our motivation and suggest that this fine-grained, progressive post-training methodology is poised to become a mainstream paradigm for various industrial-enhanced LLMs.

Yanggan Gu, Shuo Cai, Zihao Wang et al.

Model merging combines task experts into one model and avoids joint training, retraining, or deploying many expert models, but the merged model often still underperforms task experts. We study this performance gap through feature drift, the difference between features produced by the merged model and by the expert on the same input. Our theory decomposes this drift into upstream propagation and local mismatch, tracks how it propagates and combines through later layers in forward order, and links final feature drift to output drift. This view motivates FeatCal, which uses a small calibration set to calibrate the merged model weights layer by layer in forward order, reducing feature drift while staying close to merged weights and preserving the benefits of model merging. FeatCal uses an efficient closed-form solution to update model weights, with no gradient descent, iterative optimization, or extra modules. On the main CLIP and GLUE benchmarks, FeatCal beats Surgery and ProbSurgery, the closest post-merging calibration baselines: 85.5% vs. 77.0%/78.8% on CLIP-ViT-B/32 Task Arithmetic (TA) and 85.2% vs. 83.7%/82.2% on FLAN-T5-base GLUE. On CLIP-ViT-B/32, 8 examples per task reach 82.9%, and 256 examples per task take 53 seconds, about 4x faster than both baselines, showing better sample efficiency and lower calibration cost.

Daxiang Dong, Mingming Zheng, Dong Xu et al.

We present Qianfan-OCR, a 4B-parameter end-to-end vision-language model that unifies document parsing, layout analysis, and document understanding within a single architecture. It performs direct image-to-Markdown conversion and supports diverse prompt-driven tasks including table extraction, chart understanding, document QA, and key information extraction. To address the loss of explicit layout analysis in end-to-end OCR, we propose Layout-as-Thought, an optional thinking phase triggered by special think tokens that generates structured layout representations -- bounding boxes, element types, and reading order -- before producing final outputs, recovering layout grounding capabilities while improving accuracy on complex layouts. Qianfan-OCR ranks first among end-to-end models on OmniDocBench v1.5 (93.12) and OlmOCR Bench (79.8), achieves competitive results on OCRBench, CCOCR, DocVQA, and ChartQA against general VLMs of comparable scale, and attains the highest average score on public key information extraction benchmarks, surpassing Gemini-3.1-Pro, Seed-2.0, and Qwen3-VL-235B. The model is publicly accessible via the Baidu AI Cloud Qianfan platform.

Yuanyi Wang, Yifan Yang, Su Lu et al.

Continual post-training aims to extend large language models (LLMs) with new knowledge, skills, and behaviors, yet it remains unclear when sequential updates enable capability transfer and when they cause catastrophic forgetting. Existing methods mitigate forgetting through sequential fine-tuning, replay, regularization, or model merging, but offer limited criteria for determining when incorporating new updates is beneficial or harmful. In this work, we study LLM continual post-training through three questions: What drives forgetting? When do sequentially acquired capabilities transfer or interfere? How can compatibility be used to control update integration? We address these questions through task geometry: we represent each post-training task by its parameter update and study the covariance geometry induced by the update. Our central finding is that: forgetting can be considered as a state-relative update-integration failure, it arises when the covariance geometries induced by tasks misalign with the geometry of the evolving model state. Sequential updates transfer when they remain compatible with the model state shaped by previous updates, and interfere when state-relative geometry conflict becomes high. Motivated by this finding, we propose Geometry-Conflict Wasserstein Merging (GCWM), a data-free update-integration method that constructs a shared Wasserstein metric via Gaussian Wasserstein barycenters and uses geometry conflict to gate geometry-aware correction. Across Qwen3 0.6B--14B on domain-continual and capability-continual settings, GCWM consistently outperforms data-free baselines, improving retention and final performance without replay data. These results identify geometry conflict as both an explanatory signal for forgetting and a practical control signal for LLM continual post-training.

Yuanyi Wang, Yanggan Gu, Zihao Wang et al.

Large language model (LLM) merging has become a key technique in modern LLM development pipelines, enabling the integration of multiple task- or domain-specific expert models without retraining. However, as the number of experts grows, existing merging implementations treat model parameters as unstructured files and execute merges in a stateless, one-shot manner, leading to excessive disk I/O, redundant parameter scans, and poor scalability. In this paper, we present \textbf{MergePipe}, a parameter management system for scalable LLM merging. MergePipe is the first system that treats LLM merging as a data management and execution problem, and introduces a catalog-driven abstraction over model parameters, merge plans, and execution lineage. At its core, MergePipe employs a cost-aware planner that explicitly models expert parameter I/O and enforces user-specified I/O budgets, followed by a streaming execution engine that materializes merged models under transactional guarantees. Our key insight is that while base model reads and output writes are unavoidable, expert parameter reads dominate merge cost and constitute the primary optimization target. By making expert access budget-aware throughout planning and execution, MergePipe mitigates the $O(K)$ I/O growth of naive pipelines and achieves predictable scaling behavior. Experiments show that MergePipe reduces total I/O by up to an order of magnitude and delivers up to $11\times$ end-to-end speedups (up to 90\% wall-time reduction) over state-of-the-art LLM merging pipelines.

Daxiang Dong, Mingming Zheng, Dong Xu et al.

We present Qianfan-VL, a series of multimodal large language models ranging from 3B to 70B parameters, achieving state-of-the-art performance through innovative domain enhancement techniques. Our approach employs multi-stage progressive training and high-precision data synthesis pipelines, which prove to be critical technologies for enhancing domain-specific capabilities while maintaining strong general performance. Qianfan-VL achieves comparable results to leading open-source models on general benchmarks, with state-of-the-art performance on benchmarks such as CCBench, SEEDBench IMG, ScienceQA, and MMStar. The domain enhancement strategy delivers significant advantages in OCR and document understanding, validated on both public benchmarks (OCRBench 873, DocVQA 94.75%) and in-house evaluations. Notably, Qianfan-VL-8B and 70B variants incorporate long chain-of-thought capabilities, demonstrating superior performance on mathematical reasoning (MathVista 78.6%) and logical inference tasks. All models are trained entirely on Baidu's Kunlun P800 chips, validating the capability of large-scale AI infrastructure to train SOTA-level multimodal models with over 90% scaling efficiency on 5000 chips for a single task. This work establishes an effective methodology for developing domain-enhanced multimodal models suitable for diverse enterprise deployment scenarios.

Jinhui Ren, Huaiming Li, Yabin Liu et al.

High-fidelity vehicle drag evaluation is constrained less by solver runtime than by workflow friction: geometry cleanup, meshing retries, queue contention, and reproducibility failures across teams. We present a contract-centric blueprint for self-evolving coding agents that discover executable surrogate pipelines for predicting drag coefficient $C_d$ under industrial constraints. The method formulates surrogate discovery as constrained optimization over programs, not static model instances, and combines Famou-Agent-style evaluator feedback with population-based island evolution, structured mutations (data, model, loss, and split policies), and multi-objective selection balancing ranking quality, stability, and cost. A hard evaluation contract enforces leakage prevention, deterministic replay, multi-seed robustness, and resource budgets before any candidate is admitted. Across eight anonymized evolutionary operators, the best system reaches a Combined Score of 0.9335 with sign-accuracy 0.9180, while trajectory and ablation analyses show that adaptive sampling and island migration are primary drivers of convergence quality. The deployment model is explicitly ``screen-and-escalate'': surrogates provide high-throughput ranking for design exploration, but low-confidence or out-of-distribution cases are automatically escalated to high-fidelity CFD. The resulting contribution is an auditable, reusable workflow for accelerating aerodynamic design iteration while preserving decision-grade reliability, governance traceability, and safety boundaries.

Tianshu Zhu, Wenyu Zhang, Xiaoying Zuo et al.

SWE-bench-style agentic reinforcement learning relies on expensive stateful trajectories, yet substantial compute is wasted on sampled rollout groups with skewed pass rates, where binary rewards provide a weak contrastive signal. We frame this inefficiency as a pass-rate control problem and show that a 50% pass rate is the most informative operating point: it maximizes reward entropy, the probability of surviving group filtering, RLOO advantage energy under GRPO, and success--failure contrastive structure. Guided by this principle, we propose Prefix Sampling (PS), which replays trajectory prefixes to steer skewed groups toward this regime: successful prefixes serve as head starts for mostly failing groups, while failing prefixes serve as handicaps for mostly passing groups. In stateful agent environments, prefix states are reconstructed through replay while replayed tokens are excluded from the loss, restricting optimization to continuations generated by the current policy. On SWE-bench-style agentic RL, PS delivers end-to-end wall-clock speedups of 2.01x on Qwen3-14B and 1.55x on Qwen3-32B while preserving or improving final verified performance. For 14B, the SWE-bench Verified peak rises from the baseline peak of 0.273 to 0.295 under PS. Additional mathematical reasoning experiments on AIME 2025 show the same pass-rate control pattern and decompose the gains into replay, bidirectional coverage, and adaptive control.

Annan Li, Chufan Wu, Zengle Ge et al.

Large language models (LLMs) are catalyzing the development of autonomous AI research agents for scientific and engineering discovery. We present FM Agent, a novel and general-purpose multi-agent framework that leverages a synergistic combination of LLM-based reasoning and large-scale evolutionary search to address complex real-world challenges. The core of FM Agent integrates several key innovations: 1) a cold-start initialization phase incorporating expert guidance, 2) a novel evolutionary sampling strategy for iterative optimization, 3) domain-specific evaluators that combine correctness, effectiveness, and LLM-supervised feedback, and 4) a distributed, asynchronous execution infrastructure built on Ray. Demonstrating broad applicability, our system has been evaluated across diverse domains, including operations research, machine learning, GPU kernel optimization, and classical mathematical problems. FM Agent reaches state-of-the-art results autonomously, without human interpretation or tuning -- 1976.3 on ALE-Bench (+5.2\%), 43.56\% on MLE-Bench (+4.0pp), up to 20x speedups on KernelBench, and establishes new state-of-the-art(SOTA) results on several classical mathematical problems. Beyond academic benchmarks, FM Agent shows considerable promise for both large-scale enterprise R\&D workflows and fundamental scientific research, where it can accelerate innovation, automate complex discovery processes, and deliver substantial engineering and scientific advances with broader societal impact.

Yifan Yang, Jinjia Li, Kunxi Li et al.

The rapid advancement of large language models (LLMs) demands increasingly reliable evaluation, yet current centralized evaluation suffers from opacity, overfitting, and hardware-induced variance. Our empirical analysis reveals an alarming inconsistency in existing evaluations: the standard deviation across ten repeated runs of a single model on HumanEval (1.67) actually exceeds the performance gap among the top-10 models on the official leaderboard (0.91), rendering current rankings statistically precarious. To mitigate these instabilities, we propose a decentralized evaluation framework that enables hardware and parameter diversity through large-scale benchmarking across heterogeneous compute nodes. By leveraging the blockchain-based protocol, the framework incentivizes global contributors to act as independent validators, using a robust reward system to ensure evaluation integrity and discourage dishonest participation. This collective verification transforms evaluation from a "centralized black box" into a "decentralized endorsement" where multi-party consensus and diverse inference environments yield a more stable, representative metric. Experimental results demonstrate that the decentralized evaluation framework reduces the standard deviation across ten runs on the same model to 0.28. This significant improvement over conventional frameworks ensures higher statistical confidence in model rankings. We have completely implemented this platform and will soon release it to the community.

Haotian Zhao, Yuxin Zhang, Songlin Zhou et al.

Reinforcement learning (RL) has significantly advanced the ability of large language model (LLM) agents to interact with environments and solve multi-turn tasks. Yet effective training remains challenging, as sparse, outcome-only rewards make it difficult to assign credit to individual steps in an agent's action trajectory. A common remedy is to introduce dense intermediate supervision, such as process reward models or auxiliary self-supervised signals, but this increases supervision and tuning complexity and often generalizes poorly across tasks and domains. This paper presents AEM, a supervision-free credit assignment method that adaptively modulates entropy dynamics during RL training to achieve a more effective exploration-exploitation trade-off. Theoretically, we elevate entropy analysis from the token level to the response level to reduce token sampling variance and show that entropy drift under natural gradients is intrinsically governed by the product of the advantage and the relative response surprisal. Specifically, we derive a practical proxy to reshape training dynamics, enabling a natural transition from exploration to exploitation. Extensive experiments across various benchmarks and models ranging from 1.5B to 32B parameters demonstrate the effectiveness of AEM, including a notable 1.4 percent gain when integrated into a state-of-the-art baseline on the highly challenging SWE-bench-Verified benchmark.

Zhou Chen, Zhiqiang Wei, Yuqi Bai et al.

Model routing allocates queries to the suitable model, improving system performance while reducing costs. However, existing routing methods face practical limitations that hinder scalability in large-scale applications and struggle to keep up with the rapid growth of the large language model (LLM) ecosystem. To tackle these challenges, we propose TagRouter, a training-free model routing method designed to optimize the synergy among multiple LLMs for open-domain text generation tasks. Experimental results demonstrate that TagRouter outperforms 13 baseline methods, increasing the accept rate of system by 6.15% and reducing costs by 17.20%, achieving optimal cost-efficiency. Our findings provides the LLM community with an efficient and scalable solution for model ensembling, offering users an evolvable "super model."

Yanhao Zhang, Pan Pan, Yun Zheng et al.

Visual search plays an essential role for E-commerce. To meet the search demands of users and promote shopping experience at Alibaba, visual search relevance of real-shot images is becoming the bottleneck. Traditional visual search paradigm is usually based upon supervised learning with labeled data. However, large-scale categorical labels are required with expensive human annotations, which limits its applicability and also usually fails in distinguishing the real-shot images. In this paper, we propose to discover Virtual ID from user click behavior to improve visual search relevance at Alibaba. As a totally click-data driven approach, we collect various types of click data for training deep networks without any human annotations at all. In particular, Virtual ID are learned as classification supervision with co-click embedding, which explores image relationship from user co-click behaviors to guide category prediction and feature learning. Concretely, we deploy Virtual ID Category Network by integrating first-clicks and switch-clicks as regularizer. Incorporating triplets and list constraints, Virtual ID Feature Network is trained in a joint classification and ranking manner. Benefiting from exploration of user click data, our networks are more effective to encode richer supervision and better distinguish real-shot images in terms of category and feature. To validate our method for visual search relevance, we conduct an extensive set of offline and online experiments on the collected real-shot images. We consistently achieve better experimental results across all components, compared with alternative and state-of-the-art methods.

Weitsung Lin, Tinghsuan Chao, Jianmin Wu et al.

As emojis are widely used in social media, people not only use an emoji to express their emotions or mention things but also extend its usage to represent complicate emotions, concepts or activities by combining multiple emojis. In this work, we study how emoji combination, a consecutive emoji sequence, is used like a new language. We propose a novel algorithm called Retrieval Strategy to predict what emoji combination follows given a short text as context. Our algorithm treats emoji combinations as phrase in language, ranking sets of emoji combinations like retrieving words from dictionary. We show that our algorithm largely improves the F1 score from 0.141 to 0.204 on emoji combination prediction task.