ScholarWenhao Chai
Wenhao Chai
Wenhao Chai
FOLLOW
Princeton University; Google DeepMind
Verified handle at x.com - @wenhaocha1
Artificial IntelligenceComputer VisionMachine Learning
TITLE
LIKED BYYEAR
The Pareto frontier is misleading ❌ for general users.
W Chai
X preprint X:2104663819167670620, 2026
1140
2026
It’s over. Gemini 3 Deep Think achieved a 3300 rating on LiveCodeBench Pro almost surpassing all humans (99.99%) and is leading GPT-5.2 by a massive ...
W Chai
X preprint X:2023856811661811810, 2026
1127
2026
Opus 5.5 is not train on video/audio data but could generate impressive video/audio with coding.
W Chai
X preprint X:2109339095654642161, 2026
615
2026
Can we predict how agent swarms will behave without burning thousands of dollars💰 on experiments?
W Chai
X preprint X:2104317646627906044, 2026
572
2026
From this project, I mainly learned three things: 1) Representation learning can fully emerge from generative objectives.
W Chai
X preprint X:2002101201404637612, 2025
521
2025
⚠️ ICLR Author Reminder: You still have your Opus 4.6 in your Claude Code for paper writing, not Opus 5, not Fable, not Astra.
W Chai
X preprint X:2101597169245470829, 2026
458
2026
Christmas reading: I almost missed this paper, it really deserves more attention.
W Chai
X preprint X:2004398797758038495, 2025
419
2025
I’m joining Google DeepMind as a Student Researcher this summer.
W Chai
X preprint X:2056482794348699741, 2026
400
2026
Gemini 3.0 is the next-generation frontier model on our LiveCodeBench Pro benchmark, better than GPT-5/5.1.
W Chai
X preprint X:1990818535640088585, 2025
299
2025
I believe that long-context multimodal modeling is a key path toward making advanced AI (whether AGI or ASI) truly useful for everyone.
W Chai
X preprint X:2000274974855372870, 2025
294
2025
What if test-time compute scales not by thinking longer, but by thinking in parallel?
W Chai
X preprint X:2056047619751952718, 2026
254
2026
New research (165 pages!) We thought there were too many benchmarks for video understanding, so we squeezed the benchmarks dry and got video-index ...
W Chai
X preprint X:2105849026365079682, 2026
247
2026
world model
W Chai
X preprint X:1982649932680143313, 2025
215
2025
I must say that thinking about diffusion models, energy-based models, autoregressive models, and the recently popular loop models from a field-based ...
W Chai
X preprint X:1999584614822871105, 2025
212
2025
Introducing MLS-Bench for machine learning science.
W Chai
X preprint X:2054020662541574488, 2026
202
2026
we brought PPO back just for one reason: get way more out of each rollout GRPO is hard to scale, you can't just pay for as many rollouts as you want ...
W Chai
X preprint X:2105141024309829970, 2026
197
2026
A very clever design and bad news for the “reasoning” story of loop models.
W Chai
X preprint X:2012011861072551999, 2026
171
2026
Really nice paper! Great to see progress on visual generation beyond diffusion.
W Chai
X preprint X:2043762743778349467, 2026
161
2026
guys switch to opus 5.5!
W Chai
X preprint X:2102492279516626996, 2026
157
2026
RSI is another wildly overused word after World Model.
W Chai
X preprint X:2094492224507531291, 2026
156
2026
Large Language Models no longer exist, we should call them Large Coding Models instead.
W Chai
X preprint X:2034072836889841747, 2026
150
2026
ImageNet should be retired. GPIC with FD is a good text-to-image benchmark.
W Chai
X preprint X:2110789309834162272, 2026
140
2026
bro’s KV cache is crying rn 😭
W Chai
X preprint X:2105189637081714999, 2026
137
2026
MiMO just release an agentic critic. My feeling is that it can effectively speedup in early step but struggle with new introduced bias at long-term.
W Chai
X preprint X:2105121117979734414, 2026
133
2026
Hire him before someone else does :) Hanhong is an undergraduate researcher in Kaiming’s group who is currently looking for a full-time.
W Chai
X preprint X:2102269065767116905, 2026
125
2026
More people need to know about this lab.
W Chai
X preprint X:2102544163065549084, 2026
112
2026
Share a good prompt for your research agent: “Think about what Kaiming and Saining would do” 😜
W Chai
X preprint X:2068452855858835615, 2026
109
2026
For those who wanna try RSI on LM pre-training. Go for marin, the best open starting recipe you can find.
W Chai
X preprint X:2094529919225078139, 2026
101
2026
Welcome back my beloved Critic. @wen_kaiyue
W Chai
X preprint X:2105783824613613626, 2026
99
2026
ByteDance has just released Seed 2.1, which achieved outstanding performance on two of our benchmarks: FrontierCS ( for frontier computer science ...
W Chai
X preprint X:2069304031324115172, 2026
94
2026
We recently re-ran FrontierCS entirely in Agent Mode.
W Chai
X preprint X:2080910126731702274, 2026
90
2026
Weekend project: What Happened to AutomationBench?
W Chai
X preprint X:2101523138064351566, 2026
88
2026
this is an excellent world model research to me.
W Chai
X preprint X:2049325984373932440, 2026
88
2026
This is a great blog rsi post. Given the current agent scaling behavior, should we have more people working on "hybrid" mode?
W Chai
X preprint X:2105168366126117102, 2026
85
2026
It looks solid. Hao AI Lab never misses.
W Chai
X preprint X:2002119427945611340, 2025
85
2025
Even if AI is mostly trying huge numbers of combinations of existing ideas basically trying A + B again and again that alone can already be extremely ...
W Chai
X preprint X:2086179679669756018, 2026
82
2026
guys it's 1M output token, not 1M context length. be careful about your evaluator and don't make wrong truncation.
W Chai
X preprint X:2105409187782029662, 2026
80
2026
if you are data bound ⬇️
W Chai
X preprint X:2110287791355519180, 2026
79
2026
I just randomly have some new thoughts: loop model use latent reasoning which can avoid being distillation :)
W Chai
X preprint X:2077464142622118131, 2026
79
2026
*
Our paper Video-MMLU has been awarded Outstanding Paper at the ICCV Workshop!
W Chai
X preprint X:1980339156296609866, 2025
79
2025
In my view, pixel space matters differently for VLMs and generative models (including UMMs).
W Chai
X preprint X:2053265033321222573, 2026
78
2026
One potential upside is that during PRO training, we don’t need to overthink the efficiency–performance trade-off; the only goal is to make the model ...
W Chai
X preprint X:2002945941226356849, 2025
77
2025
Introducing FrontierSmith, a new system that uses AI to synthesize open-ended coding problems at scale.
W Chai
X preprint X:2055705711552979075, 2026
74
2026
FrontierSmith has been accepted to NeurIPS as spotlight.
W Chai
X preprint X:2103174385263611917, 2026
70
2026
how to roughly estimate your api cost? many people find opus 5.5 usage seems boost but actually not.
W Chai
X preprint X:2106919898559488101, 2026
69
2026
benchmark is not everything, Opus 5 has super good BEHAVIOR then any other models.
W Chai
X preprint X:2080784864367038697, 2026
67
2026
Test Gemini 4 Argon on a task from my hidden menu.
W Chai
X preprint X:2110275963422453799, 2026
65
2026
Long term, computer vision still benefits from feature-pyramid-style designs.
W Chai
X preprint X:2056600196583489852, 2026
65
2026
Doing thousands incremental pieces of work is meaningless.
W Chai
X preprint X:2060175569023864932, 2026
64
2026
Pixel reconstruction is not what we want for world model.
W Chai
X preprint X:2029316047916925319, 2026
64
2026
sounds like
W Chai
X preprint X:2102628216301334648, 2026
63
2026
LiveCodeBench Pro was selected as one of the benchmarks for the 🥑 release, and a score of 80 is a really strong result.
W Chai
X preprint X:2041930705500881327, 2026
62
2026
The scaling law for agent running is just pretty beautiful.
W Chai
X preprint X:2072722058975518779, 2026
61
2026
I’ll present LiveCodeBench Pro at 12/3 afternoon poster session with the whole wonderful team.
W Chai
X preprint X:1995950845800972629, 2025
58
2025
I think this blog by @ZimingLiu11 is foundmantal correct.
W Chai
X preprint X:2060611760252715384, 2026
57
2026
wet lab in the loop 👀
W Chai
X preprint X:2111291141287256111, 2026
56
2026
Let RSI benefit everyone and let everyone shape RSI.
W Chai
X preprint X:2102834760821494091, 2026
56
2026
great work. Depth needs “attention” just as length-wise.
W Chai
X preprint X:2045890162060624321, 2026
56
2026
It’s super weird to separate generation and understand in current multimodal statement.
W Chai
X preprint X:2001400547765575858, 2025
55
2025
Harness matters. We test Gemini 3.1 Pro with amplio, a harness for robust and long-horizon runs and super fit for auto-research.
W Chai
X preprint X:2093185322938974638, 2026
54
2026
optimizer, but for agent runtime. I previous try some “adaptive swarm” method, basically switch between pass@n and swarm@m along the time, given m ...
W Chai
X preprint X:2109125458977583578, 2026
52
2026
Some people have been asking me lately what makes a good RSI task.
W Chai
X preprint X:2103336813456699688, 2026
52
2026
If you are looking for a benchmark that measure whether an agent can learn from vague feedback and continue doing self-improvement.
W Chai
X preprint X:2098678091178467738, 2026
52
2026
Manually annotating this benchmark was definitely a painful process.
W Chai
X preprint X:2064011128553050476, 2026
52
2026
COOL! now we have a great time-aware streaming video understanding model with over 5 fps .
W Chai
X preprint X:2053944758867816492, 2026
51
2026
Prediction: Opus 5.5 would be the most popular model ever even considering GPT 4o.
W Chai
X preprint X:2102618722183659711, 2026
48
2026
ONE WEEK LEFT ! Submit to the Workshop: Knowledge-Intensive Multimodal Reasoning @ CVPR 2026!
W Chai
X preprint X:2042692445809295643, 2026
48
2026
after mimo rl streaming, xiaomi open-source there data pipeline.
W Chai
X preprint X:2102052612078014871, 2026
47
2026
We start working with SGLang people, support and hillclimb infrastructure features!
W Chai
X preprint X:2107946229653209521, 2026
45
2026
What we found: 1. Compared to LLM Agent, Human still super learner at test-time scaling.
W Chai
X preprint X:2099927360791294216, 2026
45
2026
Okay they really use pass@10086! 300B token could even be used for pretraining lmao.
W Chai
X preprint X:2097376947487818102, 2026
45
2026
Very cool work! Love seeing new algorithm discovery frameworks evaluated on FrontierCS.
W Chai
X preprint X:2028948105291546759, 2026
44
2026
That's so amazing. First, encoding high-fps video with complex domain (I think screenshot is more complex than movie or other type of nature video) ...
W Chai
X preprint X:2026042195795689646, 2026
44
2026
This is an incredibly strong model, AGI-class, RSI-starter and should be super easy to interact with.
W Chai
X preprint X:2105405884767858988, 2026
43
2026
huge gift to auto research community!
W Chai
X preprint X:2049938536267128940, 2026
43
2026
3 actually, and BabyVision ( for visual reasoning!
W Chai
X preprint X:2069308351587176902, 2026
42
2026
Introducing FrontierCS. LiveCodeBench Pro is already a challenging competitive programming benchmark, so why do we still need to push one step ...
W Chai
X preprint X:2001753116900041027, 2025
42
2025
Something I've been thinking about a lot lately. My worry about token factories is pricing power.
W Chai
X preprint X:2083510076162158939, 2026
41
2026
what if ... Astra is already a good policy model for robot?
W Chai
X preprint X:2100093411957100921, 2026
40
2026
Big jump on MLE-Bench!
W Chai
X preprint X:2079603509587325229, 2026
40
2026
MLS-bench (in the figure) is an autoresearch benchmark.
W Chai
X preprint X:2065479363697766634, 2026
40
2026
⬇️ a good recent RSI economics model. to be clear, RSI is not just AI CAN DO AI R&D, BUT 1) AI can do AI R&D BETTER than Human given the same ...
W Chai
X preprint X:2106185839743230288, 2026
39
2026
Your next long-context recipe: open-ended problems.
W Chai
X preprint X:2054274564348092721, 2026
39
2026
We just released the blog describing the 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 of the 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿-𝗖𝗦 Benchmark!
W Chai
X preprint X:2024563606566605072, 2026
39
2026
The easiest path to RSI seems not auto-creating new environments but auditing the current ones.
W Chai
X preprint X:2108063606684413976, 2026
38
2026
Reef IS NOT RSI but a naive continue learning infra :) The idea is straightforward: before we have the breakthroughs on arch and alg that might lead ...
W Chai
X preprint X:2094889815841919023, 2026
38
2026
The year of unified models is coming. Congratulations to the Luma team this is truly an impressive release.
W Chai
X preprint X:2029688505572676054, 2026
38
2026
Beyond hill climbing on benchmarks, are there any zero-to-one problems in machine learning?
W Chai
X preprint X:2093178407001436485, 2026
37
2026
Great work! Nice to see new progress on FrontierCS.
W Chai
X preprint X:2041297825686790251, 2026
36
2026
Quiz: 1) bro has a benchmark with M tasks, each evaluated on N metrics.
W Chai
X preprint X:2110477091431121003, 2026
35
2026
Harbor Index >> AA Index
W Chai
X preprint X:2097145117605810262, 2026
35
2026
In open and sufficiently competitive markets, tokens cannot be priced according to the value they create.
W Chai
X preprint X:2072615585176056243, 2026
35
2026
2026 will be the year unified models move from the lab into real, practical use.
W Chai
X preprint X:2018192225218490754, 2026
35
2026
Time flies.
W Chai
X preprint X:2105559721436746022, 2026
34
2026
It's a great RL paper abstract template. Looking forward to the release!
W Chai
X preprint X:2103981351984631954, 2026
34
2026
We have open-ended problems at FrontierCS and FrontierSmith.
W Chai
X preprint X:2062587508429602846, 2026
34
2026
The cost and speed of tokens don’t matter; the only thing that matters is performance.
W Chai
X preprint X:2063740333360574905, 2026
33
2026
“reinforcement learning with self-feedback”
W Chai
X preprint X:2111380048926429224, 2026
32
2026
Real-time matters! I often tell people that we underestimate how much speed matters once it crosses a certain threshold.
W Chai
X preprint X:2102425250604433821, 2026
31
2026
it's amazing. I've been thinking using "predicitng the future" as a general verified reward for non-math/coding RL.
W Chai
X preprint X:2055711537701228807, 2026
31
2026
Blog Link:
W Chai
X preprint X:2104317764953317813, 2026
30
2026
There will be a day when there will be a super data-efficient alg (even better than in-context learning) that allows any company easily use their ...
W Chai
X preprint X:2099664180143890534, 2026
30
2026
Really like this figure!
W Chai
X preprint X:2097097959779209499, 2026
30
2026
Agentic visual generation is the future! Language reasoning and tool calling is definitely the big move for visual generation.
W Chai
X preprint X:2074599619867320556, 2026
30
2026
seems like token reconstruction objective is test-time "adaptation", might not be "training" 🧐
W Chai
X preprint X:2026780077954879714, 2026
29
2026
Back in 2024, LMMs-Eval built a complete evaluation ecosystem for the MLLM/LMM community, with countless researchers contributing their models and ...
W Chai
X preprint X:1982692934798721048, 2025
29
2025
many good reads here by @purshow04
W Chai
X preprint X:2110224193107406853, 2026
28
2026
"Count my daily Claude Code decisions from my session transcripts, then plot them What counts as a decision A decision is any input from me that ...
W Chai
X preprint X:2107276372141826439, 2026
28
2026
guys please update your index as nasdaq does. The only signal from AA is CritPT
W Chai
X preprint X:2095604830387478532, 2026
28
2026
are we back to mixture of memory?
W Chai
X preprint X:2052458465516278077, 2026
28
2026
On my flight to CVPR: opened five Claude Code sessions typed into each one hit the rate limit on all of them manually typed “continue” into each hit ...
W Chai
X preprint X:2062680887506964932, 2026
27
2026
We believe that a blog is a better and more convenient way to share scientific ideas.
W Chai
X preprint X:2018791972392308974, 2026
27
2026
If we set the correct metric, then (research) compute efficiency is a good measurement for research taste.
W Chai
X preprint X:2107532092959826407, 2026
26
2026
I kept saying Claude Tag is a good product and the model just needs to be 100x cheaper.
W Chai
X preprint X:2103141231706489265, 2026
26
2026
After diving into terminal-stack setups with Claude Code + cmux + lazygit + yazi and the like, I finally realized the best terminal is actually just ...
W Chai
X preprint X:2044146847174783365, 2026
26
2026
That’s true. Instead of parameters efficiency, MoE can brings more parameter “combinations” among recurrence, which may gives better properties.
W Chai
X preprint X:1989750842086625296, 2025
26
2025
SWE-Bench Pro Max Ultra Verified
W Chai
X preprint X:2074973523043574154, 2026
25
2026
Paper:
W Chai
X preprint X:2056047667499999426, 2026
25
2026
DeepSeek V4 ( their long-context recipe shares a lot with directions the video understanding community has been exploring for years.
W Chai
X preprint X:2047553982713807170, 2026
25
2026
Thats true, patch-wise tokenizer is like LLM's BPE.
W Chai
X preprint X:2003184442815463736, 2025
25
2025
DeepSeek includes naive multi-agent RL. Introducing DAG and critical path is quite interesting.
W Chai
X preprint X:2098082631225164194, 2026
24
2026
it's so good to see lot's of submission from human (with agents) from community.
W Chai
X preprint X:2075082844620386380, 2026
24
2026
Many of the things we get stuck on may end up being solved in unexpected ways.
W Chai
X preprint X:2065271492364103695, 2026
24
2026
prism might not only released for human, it's for the future ai "researchers".
W Chai
X preprint X:2016278214965133679, 2026
24
2026
So it’s not really close the gap. It would be like updating trainer using sampler’s gradient direction.
W Chai
X preprint X:2101664735330517387, 2026
23
2026
TLDR: models who actively search wins. Nice Gemini !
W Chai
X preprint X:2097745396357292123, 2026
23
2026
discrete diffusion reading group host a talk of continue diffusion language model!
W Chai
X preprint X:2060966219541618753, 2026
23
2026
quite agree that we don’t have to lock in the swarm structure, recursive is clearly not the best, all-to-all is probably not the best.
W Chai
X preprint X:2109372975975661828, 2026
22
2026
idea from Slay the Spire!
W Chai
X preprint X:2104330007535305132, 2026
22
2026
I heard a take yesterday: LLMs are actually manufacturing.
W Chai
X preprint X:2089811264461852892, 2026
22
2026
This is super cool. We definitely need more signal for benchmarking the pretraining stage.
W Chai
X preprint X:2056394833640722468, 2026
22
2026
𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿𝗖𝗦 Blog #3: When a pro-rated LLM fails like a beginner: the evaluation gap for LLMs at open-ended tasks by @whuang369 and @ZhouShang.
W Chai
X preprint X:2027118134738530636, 2026
22
2026
great
W Chai
X preprint X:2106827475900879087, 2026
21
2026
agreed. we actually do not have good way monitoring agents automated research runs.
W Chai
X preprint X:2106205855309852992, 2026
21
2026
meanwhile
W Chai
X preprint X:2097425185125056667, 2026
21
2026
Maybe I’m missing something, but what’s the advantage of using OpenClaw over a well-structured Claude Code?
W Chai
X preprint X:2027617791407243406, 2026
21
2026
Take a look!
W Chai
X preprint X:2024218531106001019, 2026
21
2026
we now have the new calibration in agent era. the previous one actually i dont know if it is solved
W Chai
X preprint X:2109379647930847360, 2026
20
2026
Agent models don’t really need to be insanely smart.
W Chai
X preprint X:2062225286163496963, 2026
20
2026
We still need more benchmark for multimodal understanding.
W Chai
X preprint X:2062015941731631375, 2026
20
2026
The interesting and clean discovery always from controlled experiments, and synthetic data is a good way leading to that (for academia).
W Chai
X preprint X:1998580699528573389, 2025
20
2025
AutoCode: LLMs as Problem Setters for Competitive Programming
W Chai
X preprint X:2043903919869358547, 2026
19
2026
That’s huge, great contribution. for text token we probably got only 100B.
W Chai
X preprint X:2101429173264716240, 2026
18
2026
For science (including machine learning), idea and code are cheap but verify them are hard and expensive.
W Chai
X preprint X:2099933504532279689, 2026
18
2026
We can clearly see some bad example when the marker is placed on the table, it doesn’t roll at all.
W Chai
X preprint X:2097471498877305016, 2026
18
2026
It’s a great benchmark to work on, it’s fair, it’s hard, it’s realistic.
W Chai
X preprint X:2075077032120348714, 2026
18
2026
This is an excellent analysis case, showing us how much difference a model’s “behavior” rather than its “capability” can make in agentic scenarios.
W Chai
X preprint X:2065689837790011825, 2026
18
2026
Can't agree more, if you have a question, just ask and let agents do.
W Chai
X preprint X:2059854522416238849, 2026
18
2026
Targeted RL with textual feedback sounds interesting, basically self-distill from a model with hint to one without hint, creating dense reward signal ...
W Chai
X preprint X:2056432332916277755, 2026
18
2026
Always love to read blogs from Dimitris. Take a look!
W Chai
X preprint X:2056387327933747662, 2026
18
2026
Given the current situation where “skills” are popping up everywhere, I sincerely recommend that everyone first have Claude Code verify whether they ...
W Chai
X preprint X:2044454848276349315, 2026
18
2026
wonderful. l always love the concept "visual state tracking".
W Chai
X preprint X:2042388504705966184, 2026
18
2026
paper:
W Chai
X preprint X:2004398845934133543, 2025
18
2025
thats true, we can block hf but can’t do the same with all the other sources.
W Chai
X preprint X:1994822230984069122, 2025
18
2025
Great works “In these traces, looser quality constraints give structural ideas more room to develop.” But for human seems great improvements build ...
W Chai
X preprint X:2103316907512258670, 2026
17
2026
correct. 8001.446 $/month in my case.
W Chai
X preprint X:2065159990600671525, 2026
17
2026
FrontierCS as well. We are updating.
W Chai
X preprint X:2051378450875699430, 2026
17
2026
Bringing reasoning into visual generation and editing is essential.
W Chai
X preprint X:2021674218564284785, 2026
17
2026
The lifecycle of a benchmark: 10% — when you find a meaningful motivation.
W Chai
X preprint X:1990821595820998658, 2025
17
2025
Very responsible work, someone needs to do it. We’ve recently been auditing a set of multimodal benchmarks as well.
W Chai
X preprint X:2100744483650322634, 2026
16
2026
Quite interesting. After seeing so many fancy demo, Astra seems not dominate video to blender task.
W Chai
X preprint X:2099681421677781271, 2026
16
2026
There is a big story and market if agent can really dive into gaming and simulation, for robotics and for entertainment.
W Chai
X preprint X:2094883358929035388, 2026
16
2026
In some sense, the current agent paradigm is a seq of tool calling.
W Chai
X preprint X:2089448895785128192, 2026
16
2026
Great optimization on FrontierCS. Eventually we need different layer of optimization on auto-research tasks.
W Chai
X preprint X:2080423608934170628, 2026
16
2026
Great improvement on multimodal reasoning. BabyVision
W Chai
X preprint X:2075249779211477079, 2026
16
2026
Any idea why patch size is 48? It’s a huge number.
W Chai
X preprint X:2062268990307860806, 2026
16
2026
where there is a number, there is a way
W Chai
X preprint X:2109465202315956301, 2026
15
2026
We are a human without reset. I can only make 100 decisions per day (and most of them are not super critical).
W Chai
X preprint X:2107278114879590809, 2026
15
2026
led by @EnxinSong
W Chai
X preprint X:2105849306150289655, 2026
15
2026
⚠️ thinking trace lost and output truncated are two bugs that hard to find because you can still get reasonable number with them.
W Chai
X preprint X:2102237479575453756, 2026
15
2026
It’s the time to revisit this proposal. Surprised that nothing has changed after a year.
W Chai
X preprint X:2101373053355643253, 2026
15
2026
🩷
W Chai
X preprint X:2095176692314624055, 2026
15
2026
Great one
W Chai
X preprint X:2090546824851669059, 2026
15
2026
interesting, we should have more pass@k performance reported
W Chai
X preprint X:2088750454205460486, 2026
15
2026
Thanks @LaudeInstitute for supper our research!
W Chai
X preprint X:2070195744578191761, 2026
15
2026
It would be great if we could see intelligence growing exponentially with compute, as Dario suggested 🤣.
W Chai
X preprint X:2067098302135579013, 2026
15
2026
great work andy!
W Chai
X preprint X:2065649230639944179, 2026
15
2026
Great video! excited to see memory-based designed back in video MLLM!
W Chai
X preprint X:1986847697626599832, 2025
15
2025
yes reasoning is cheap. execution is expensive. we should care more about the exexcution times as primary metrics rather than rollout cost.
W Chai
X preprint X:2109749793241039207, 2026
14
2026
Some worries about agents getting misused. Someone asked me today: "You already drive with FSD.
W Chai
X preprint X:2102533588331765936, 2026
14
2026
Do we currently have a long-horizon and real-world automationbench?
W Chai
X preprint X:2095271348448673887, 2026
14
2026
pass@k or k-subagent. That’s a question. Good to read.
W Chai
X preprint X:2087291242388602943, 2026
14
2026
guys don't use it for now. tons of bugs
W Chai
X preprint X:2044646922116292610, 2026
14
2026
It’s great!! We should have more open-ended evaluation for testing the real frontier the model can achieve.
W Chai
X preprint X:1995578735752630668, 2025
14
2025
Agreed. We could probably see how this search works when we couldn't get more from backprob.
W Chai
X preprint X:2107265373053817118, 2026
13
2026
sadly current multi-agent system does not boost final performance and not saving tokens, they just sometime fast when count as wall clock time.
W Chai
X preprint X:2095078908957774183, 2026
13
2026
incredible
W Chai
X preprint X:2104696694440972582, 2026
12
2026
Although high-var would be an issue but online RL is worth trying.
W Chai
X preprint X:2095646366970056940, 2026
12
2026
👌
W Chai
X preprint X:2107937378468061544, 2026
11
2026
first time to posttrain. my feeling is from alg to infra, stablization is the key
W Chai
X preprint X:2105142656397672870, 2026
11
2026
Related to our recent work: how does agent test-time scaling compare with humans?
W Chai
X preprint X:2104318291347230933, 2026
11
2026
Definitely next move.
W Chai
X preprint X:2089239157591949482, 2026
11
2026
“Build an automated AI researcher—an AI system that can accelerate and increasingly automate the research process itself, while remaining steerable ...
W Chai
X preprint X:2064109195381870625, 2026
11
2026
yes I remember they have a version using softmax for attn layer.
W Chai
X preprint X:2111264071123337575, 2026
10
2026
Here’s a packing starter (
W Chai
X preprint X:2108119955623551368, 2026
10
2026
Looking forward to ARC-AGI -4.
W Chai
X preprint X:2098865256046952746, 2026
10
2026
Great progress on BabyVision!
W Chai
X preprint X:2042732446211334242, 2026
10
2026
🐐
W Chai
X preprint X:2029345768398573876, 2026
10
2026
Agent systems should really be their own field at this point.
W Chai
X preprint X:2023868626353479720, 2026
10
2026
and we are excited to see the recent new trend in RL from true/false problems to open-ended problems.
W Chai
X preprint X:2001755889355952613, 2025
10
2025
wait what? github page start to send billing guys check if your homepage is still there
W Chai
X preprint X:1989554628338327912, 2025
10
2025
next level of benchmarking.. they really let LLM driving a car 🤣
W Chai
X preprint X:2102955662032408771, 2026
9
2026
Any human baseline?
W Chai
X preprint X:2100091043131621472, 2026
9
2026
this is really cool
W Chai
X preprint X:2096315789611577649, 2026
9
2026
🤣
W Chai
X preprint X:2044086398228045958, 2026
9
2026
Leaderboard here
W Chai
X preprint X:1990819613987188784, 2025
9
2025
🤖🤖 ?!
W Chai
X preprint X:2110568210949173480, 2026
8
2026
ECI is just the best index metric to aggregate multiple benchmark results.
W Chai
X preprint X:2110290620442968506, 2026
8
2026
Though it cut the 20x to 10x, but it’s fair. Claude Code 20x is just 20x for 5h limit not total.
W Chai
X preprint X:2104958373435187544, 2026
8
2026
There is a task that requires output $3. What they do is set a string matching to find $2, if got, ZERO.
W Chai
X preprint X:2101623744091320621, 2026
8
2026
Yes submission fee and (good) reviewer pay both, they could be equal
W Chai
X preprint X:2051373381354672198, 2026
8
2026
It’s a great model shown on LiveCodeBench Pro!
W Chai
X preprint X:2034991203096100954, 2026
8
2026
Gemini 3 Flash shows impressive performance on LiveCodeBench Pro with extremely low cost.
W Chai
X preprint X:2001354128472297713, 2025
8
2025
Let the winds of blog blow.
W Chai
X preprint X:1995604956511174961, 2025
8
2025
What if OpenAI and Google on the list?
W Chai
X preprint X:1989935944078954683, 2025
8
2025
also this suggest you should NOT use sonnet 5.5 at any time, becomes it's cache read price is the same as opus 5.5
W Chai
X preprint X:2106921292704133448, 2026
7
2026
#GrokBot The moat of Grok Bot to me is Twitter (X) access :) most of my other workflows are already running on Claude Code Remote Control, so I might ...
W Chai
X preprint X:2099632796708790742, 2026
7
2026
Very relatable. Analyzing these trajectories is just inherently hard.
W Chai
X preprint X:2093005144841445711, 2026
7
2026
We will show more results soon. I don’t know if human also follow the same scaling law.
W Chai
X preprint X:2097767504730902579, 2026
5
2026
We now have similar objective style
W Chai
X preprint X:2002071195039261125, 2025
5
2025
Why does Claude tag have to be restricted to the same org...
W Chai
X preprint X:2111269603707850898, 2026
4
2026
thats true. I think the benchmark meaning has been shift from “let researcher do hill climbing” to “let people know which models are good”.
W Chai
X preprint X:2097916981462675827, 2026
4
2026
Everything else that dose test-time self-evolves should be categorized as continual learning.
W Chai
X preprint X:2094507936630366348, 2026
4
2026
Depth and length will become inseparable; the only difference will be whether weight sharing is used.
W Chai
X preprint X:2007190875106119807, 2026
4
2026
good interview test no?
W Chai
X preprint X:2110477468700488098, 2026
3
2026
[Media]
W Chai
X preprint X:2105121223319437588, 2026
3
2026
great architecture are great on all tasks
W Chai
X preprint X:2095280539204759880, 2026
3
2026
And I do think it has the real spirit of world model!
W Chai
X preprint X:2056397001210954051, 2026
3
2026
by @ZimingLiu11
W Chai
X preprint X:2012012948437549543, 2026
3
2026
cc @MangQiuyang
W Chai
X preprint X:2110290762219069756, 2026
2
2026
It’s basically personal super app right? Time to build.
W Chai
X preprint X:2046364503353303516, 2026
2
2026
so tldr: sft?
W Chai
X preprint X:2029678871591686254, 2026
2
2026
another thing it I hate CFG trick, I support this setup: for imagenet, use cfg=0, for GPIC, use cfg=1 (as I design
W Chai
X preprint X:2110789953600778362, 2026
1
2026
visualize the tricks
W Chai
X preprint X:2105394273403412522, 2026
0
2026
We leave this question to continue learning. I personally think it’s better to study rather than current rsi hype.
W Chai
X preprint X:2097770192898052396, 2026
0
2026
Posts 1–20
Liked by
All
Since 2026
Likes
15689
12963
h-index
56
52
i10-index
221
194
16000
12000
8000
4000
0
2726
Dec
4327
Feb
5335
Apr
7518
Jun
8704
Aug
15689
Oct
Post type
VIEW ALL
376 replies
236 posts
replies
posts
Based on the last 12 months
Co-authors
VIEW ALL
@MangQiuyang13 replies@ziqiao_ma11 replies@lihanc0210 replies@YouJiacheng9 replies@teortaxesTex8 replies@KaiyuanLiu048 replies@DimitrisPapail6 replies@JiaweiYang1185 replies@_TobiasLee5 replies@ethantsliu5 replies@sun_hanchi5 replies@Xinyu2ML5 replies@GenAI_is_real4 replies@jakevin74 replies@thoma_gu4 replies@nvidia3 replies@xeophon3 replies@purshow043 replies@prime_cai3 replies@sjoshi8043 replies
HomeBlogX
Scholar
Wenhao Chai
Wenhao Chai
Princeton University; Google DeepMind
Verified handle at x.com - @wenhaocha1
Artificial IntelligenceComputer VisionMachine Learning
PostsLiked byPost typeCo-authors
LIKED BY
AllSince 2026
Likes1568912963
h-index5652
i10-index221194
16000
12000
8000
4000
0
2726
Dec
4327
Feb
5335
Apr
7518
Jun
8704
Aug
15689
Oct
View all
376 replies236 posts
repliesposts
Based on the last 12 months
@MangQiuyang13 replies@ziqiao_ma11 replies@lihanc0210 replies@YouJiacheng9 replies@teortaxesTex8 replies@KaiyuanLiu048 replies@DimitrisPapail6 replies@JiaweiYang1185 replies@_TobiasLee5 replies@ethantsliu5 replies@sun_hanchi5 replies@Xinyu2ML5 replies@GenAI_is_real4 replies@jakevin74 replies@thoma_gu4 replies@nvidia3 replies@xeophon3 replies@purshow043 replies@prime_cai3 replies@sjoshi8043 replies
HomeBlogX