
PhD student @PrincetonCS. Penn '25. Interested in all things empirical deep learning.
Princeton, NJ
Joined November 2024
- Tweets25
- Following239
- Followers110
- Likes135
Boya Zeng retweeted
How do you find the best agentic reasoning strategy for a given task?
We introduce Strategist, an agent that evolves better llm reasoning strategies as Python programs, storing each one in a library so the next task can start from it.
Yes, chain-of-thought was invented by human, but the next one might not be!
Details in thread 🧵
I really like controlled analyses like this. For almost any verifiable task with a benchmark, there are myriad ways to max it out, and in the near future agents will likely be able to hillclimb most verifiable metrics.
But getting to a better understanding (e.g., through empirical studies on highly controlled data) will always be valuable, and is much harder to automate (after all, “understanding” is not something verifiable, as we’ve also seen in math).
Excited to release our NeurIPS 2026 paper, “Do Diffusion Models Learn to Generalize Basic Visual Skills?”
We asked if diffusion models can generalize skills beyond what they were trained on, or do they mostly interpolate between examples in their training data?
This is hard to test on today's text-to-image models, because we don't know what is in their training data. We trained small diffusion models from scratch on synthetic images where we control every value the model sees.
We found that they interpolate well but rarely extrapolate.
Congrats to @amish_sethi04 on the release! This project took quite some time🥲, but we put a lot of effort into getting all the details right
2
94
Boya Zeng retweeted
Excited to release our NeurIPS 2026 paper, “Do Diffusion Models Learn to Generalize Basic Visual Skills?”
We asked if diffusion models can generalize skills beyond what they were trained on, or do they mostly interpolate between examples in their training data?
This is hard to test on today's text-to-image models, because we don't know what is in their training data. We trained small diffusion models from scratch on synthetic images where we control every value the model sees.
We found that they interpolate well but rarely extrapolate.
Boya Zeng retweeted
LLM pretraining has run on single-epoch text, bits per byte and iso-FLOP curves for years. visual generation still memorizes ImageNet for tens to hundreds of epochs and pushes FID down with CFG, with no strong signal those tricks transfer to real text-to-image.
new blog: 🎨 scaling generative models on GPIC, 100M permissive text-image pairs.
training pixel-space JiT from 3.6e18 to 2.3e20 FLOPs:
1) held-out loss is our main metric, and compute-optimal TPP sits at about 50 image tokens per image parameter
2) EMA is not necessary: a proper learning-rate schedule beats EMA on held-out loss at every budget by 0.4% to 1.0%
3) we use CFG = 1 (only the conditional branch) for every result, and call on everyone to drop CFG: it saves compute and leaves no room to hack the metric
all code, configs and sweep data are open 🧵
blog: wenhaochai.com/blogs/scaling…
Boya Zeng retweeted
Will your ICLR paper get accepted? 👀
Introducing PaperScorer, our NeurIPS '26 project that predicts acceptance from anonymized submissions.
It beats GPT, Gemini, and Jev, with well-calibrated probabilities.
Try it on your paper today: paperscorer-ai.com
Blog: sachinkonan.github.io/PaperS…
Details in thread 🧵
Boya Zeng retweeted
Interactive software built on demand is becoming a dominant way we interact with AI.
But we often end up as its testers: we find a bug, tell Claude, wait for a fix, and repeat.
SWEeper-Bench 🧹 measures whether agents can sweep those bugs away before we ever see them. Surprisingly, Grok 4.6 does it best 🧵
Boya Zeng retweeted
Is model scaling the only source of agent improvement?
We @henryqin1997 @YaxinLu1997 @VITAGroupUT @VictorKaiWang1 are glad to share our work: We reach 95.3% Raw Accuracy, or a $15 Frontier Run (88.8% with DeepSeek-v4-Flash, matching GPT-5.6 Sol Max), on Terminal-Bench 2.1 via Harness Scaling.
Human effort -> Non scalable ❌
Agent direction? Local optimization ❌
Human direction + agent action = better workflow control✅
StateM:
1. Open-source model + StateM -> Close-source frontier, everyone can try!
2. Training free, significantly improve frontier agents.
3. Zero-transfer-cost on same family of agents. [1/8]🧵
📄arxiv.org/abs/2608.15089
🌐 henryqin1997.github.io/state…
Excited to share i1! In T2I research, arch & recipe & data often change together, which makes it hard to isolate what actually matters. With i1, we tried to make this more controlled thru many modeling & data ablations.
My main takeaway from the paper is about evaluation:🧵
2. We should move quickly when a standard evaluation stops being informative
GenEval is a very popular and useful benchmark for T2I, and we started our project using it as the main signal for controlled experiments. But as we looked more closely, we found that the signal could be quite noisy- many failures were evaluator errors rather than generation errors, and the benchmark covers only a fixed set of objects and rules.
This is not a criticism of GenEval specifically. It is what naturally happens when a popular benchmark becomes widely adopted and models improve around it. At some point, the benchmark can no longer be the best target for driving the next stage of progress.
The same pattern also showed up in class-conditional ImageNet generation. With the ImageNet FID lower bound (i.e., FID of the ImageNet validation set) already broken by VAR in 2024, ImageNet FID became less useful as the central metric for image generation progress. Works such as MiniT2I (@kevinxbwang2007) and T2I-ImageNet (@lucasdegeorge) also suggest that T2I is feasible at ImageNet-scale resources.
To me, this means text-to-image generation should increasingly become the natural next step after ImageNet for evaluating visual generative models: it is harder, more open-ended, and closer to how people actually use these models.
Boya Zeng retweeted
Can a small academic team build a strong text-to-image model using only public datasets?
Introducing i1: a simple, fully open recipe for strong text-to-image models
Boya Zeng retweeted
Today, agents execute isolated tasks. Tomorrow, agents will steer complex decisions across long horizons.
Introducing CEO-Bench, a first step to measure "Steering Intelligence." In CEO-Bench, agents are asked to run a simulated startup for 500 days.
ceobench.com
Boya Zeng retweeted
T2I systems look complex only because we keep adding components. How about removing them?
Introducing our latest work MiniT2I 🪄: purely flow matching objective on pixels, no auxiliary losses, all open-source data and academia-affordable compute. Everything is just designed so simple and minimal. ✨
Boya Zeng retweeted
Zeng et al., "i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models"
A fully reproducible recipe & code & weights and everything for a truly open text-to-image model. A LOT of interesting findings.
Boya Zeng retweeted
Today’s vision benchmarks suggest VLMs are nearing saturation, but real-world visual understanding is far from solved.
Introducing WorldBench: 2,000 hand-written, human-verified VQA questions focused on visual diversity and designed to be challenging for frontier models. Gemini-3.1-Pro leads with just 64.0% accuracy. (1/10)
Boya Zeng retweeted
Static math benchmarks saturate. We built one that doesn't.
Announcing MathDuels, the first self-play math benchmark.
Every frontier LLM writes problems for the others, and is graded on the ones written for it. As models improve, so does the benchmark.
Boya Zeng retweeted
One static model does not fit all😭
We just dropped our latest work: Functional Neural Memory. Instead of static models, we generate custom "parameters" for every single input.
✅Prompt your model anytime
✅Instant personalization
✅Better instruction following
✅Flexible & dynamic memory (w/o memory bank✌️)
(🧵1/6)
Boya Zeng retweeted
Do diffusion models produce novel data points, or reproduce the training data they see? What factors drive this?
In a new paper, we study this problem in the 3D generation context.
We propose an evaluation framework to quantify memorization in 3D shape generative models. We evaluate existing 3D generators and run controlled experiments to investigate what drives memorization and how to reduce it.





















