Haonan Wang

Intelligence climbs a staircase of scaling laws.

By 2023, OpenAI had made the path to general intelligence legible: capability follows compute, along a log curve — every doubling buys a roughly constant step of capability. But a logarithm is also a promise of saturation: on any single axis, the next FLOP buys less than the last, and eventually the curve goes flat.

So the frontier doesn't ride one curve — it stacks them. Each time an axis saturates a new one opens, the climb re-bases, and you can read the stages straight off the training pipeline: pretraining params & data, longer context across pretrain and midtrain, post-train RL that thinks longer, tuning-free test-time — parallel thought and interaction — and now modality. My research is a series of bets on where the next curve begins; each colored curve below is one of those bets, and the papers it produced.

Axis 01 · Context scaling 2024

Scale what the model readslonger context → more intelligence

Intelligence begins with capacity: a model that can hold more of the world in one context is simply smarter. So I worked where long context is actually made: schedule the context window during pretraining (SkyLadder), and repair the numerics that silently break RoPE at long range (AnchorAttention).

Pretraining

SkyLadder: Better & Faster Pretraining via Context Window Scheduling

Short→long context scheduling: up to +3.7% on benchmarks with up to 22% faster pretraining.

NeurIPS 2025 · co-author · Sea AI Lab

arXiv Code

Mid-training

When Precision Meets Position: BFloat16 Breaks Down RoPE in Long-Context Training

Diagnosed the BF16 × RoPE breakage; AnchorAttention restores positional coherence and cuts long-context training time.

TMLR 2025 · first author

arXiv Code

Axis 02 · Post-training · RL 2025

Scale what the model thinksmore thinking → more intelligence

Intelligence also lives in time: give the same weights longer to think, and they are smarter — trained with RL, where the bottleneck is the reward signal. VeriFree drops the verifier, carrying RLVR beyond math and code; Variational Reasoning supplies the theory — thinking traces as latent variables.

RL without verifiers

Reinforcing General Reasoning without Verifiers

Verifier-free RL matches verifier-based training on MMLU-Pro / GPQA — and extends it to unverifiable domains.

ICLR 2026 · core author · Sea AI Lab

arXiv Code OpenReview

Probabilistic view

Variational Reasoning for Language Models

Thinking traces as latent variables: ELBO → multi-trace objectives, with a forward-KL that stabilizes training.

ICLR 2026 · co-author · Sea AI Lab

arXiv Code

Axis 03 · Test-time · beyond one pass 2025

Scale beyond a single passmore minds, more turns → more intelligence

One chain of thought saturates too. So scale test time itself, without touching the weights: in width, fork K reasoning traces and merge their logits into one answer (Think in Parallel, Answer as One); in depth, treat interaction with the environment as its own scaling axis (MiroThinker).

Parallel · width

Think in Parallel, Answer as One: Logit Averaging for Open-Ended Reasoning

K traces think independently, one voice answers — +8.3 pass@1 on LiveCodeBench-hard (DeepCoder-14B).

ICLR 2026 · first author

arXiv OpenReview

Sequential · interaction

MiroThinker: Model, Context & Interactive Scaling

Open deep-research agents — interaction depth scales like parameters and context (up to 600 tool calls per task). MiroThinker-1.7 & H1 add verification for heavy-duty agents.

2025–26 tech reports · core contributor · MiroMind

arXiv v1.0 arXiv 1.7 & H1 Code

Axis 04 · Beyond text 2025 → now

Scale into the worldmore of the world → more intelligence

Text is one projection of the world — and it is running out. The modality axis scales what a model can perceive and what it can create: Next-Event Prediction turns unlabeled video into reasoning supervision; UniRL — my current chapter at Tencent Hunyuan — puts one RL loop under multimodal generation. Generation is the strictest test of understanding — what I cannot create, I do not understand.

Understand · video

Fostering Video Reasoning via Next-Event Prediction

Future segments as self-supervision (V1-33K, FutureBench): temporal reasoning from unlabeled video.

ICLR 2026 · co-first author · Sea AI Lab

arXiv Code OpenReview

Generate · now

UniRL: One RL Loop for Agentic Multimodal Generation

Open-source RL post-training across text, image and video — rollout, reward, advantage estimation, log-prob replay and distributed training, unified for heterogeneous model families.

Tencent Hunyuan · project lead · 2026 →

GitHub

Intelligence climbs a staircase of scaling laws.

Scale what the model readslonger context → more intelligence

Scale what the model thinksmore thinking → more intelligence

Scale beyond a single passmore minds, more turns → more intelligence

Scale into the worldmore of the world → more intelligence

Every plateau hides a takeoff. What is the next axis?

News

Publications

From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs

Think in Parallel, Answer as One: Logit Averaging for Open-Ended Reasoning

Fostering Video Reasoning via Next-Event Prediction

Prefix-Tuning+: Modernizing Prefix-Tuning through Attention Independent Prefix Data

Reinforcing General Reasoning without Verifiers

Variational Reasoning for Language Models

Scaling GUI Agents with Visual State Transitions

JarvisHub: An Open Harness for Canvas-Native Multimodal Creative Agents

Self-Evolving LLM Agents for Evidence-Grounded Health Community Notes

When Precision Meets Position: BFloat16 Breaks Down RoPE in Long-Context Training

SkyLadder: Better and Faster Pretraining via Context Window Scheduling

NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation

Boosting Visual-Language Models by Exploiting Hard Samples

The Stronger the Diffusion Model, the Easier the Backdoor: Data Poisoning to Induce Copyright Breaches Without Adjusting Finetuning Pipeline

Can AI Be as Creative as Humans?

Training Fair Deep Neural Networks by Balancing Influence

Can Single-Pass Contrastive Learning Work for Both Homophilic and Heterophilic Graph?

Complexity Matters: Rethinking the Latent Space for Generative Modeling

Deep Active Learning by Leveraging Training Dynamics

Understanding Programmatic Weak Supervision via Source-Aware Influence Function

A Neural Corpus Indexer for Document Retrieval

Controllable Gradient Item Retrieval

Awards & Honors

Experience

Teaching

Service