Chenghao Yang @chrome1996

[🛫ICML] We will be presenting our paper about *base-aligned model collaboration* on Tuesday morning and also Saturday (HAI Co-Creativity workshop). Happy to chat! PS: We've updated our paper, code, and reading list to include the latest results and resources. Check it out!

Yichen (Zach) Wang @ICML @YichenZW

7 months ago

Lack of diversity in your LLM generation? (also noted by Artificial Hivemind, best paper @NeurIPSConf) Time to bring your base model back! An inference-time, token-level collaboration between a base and an aligned model can optimize and control diversity and quality!

2 16 53 13K 34

Excited to contribute to this work! 👁️📝 As someone relatively new to the multimodal world, I've always wanted to explore how we can better ground LLMs. But before diving deep into modeling, I believe we need a better understanding of what our benchmarks are actually evaluating. Ever since Saining's "Is Vision Good Enough for Language?" (Thanks @sainingxie !) and recent 7-hour podcast deep dives, I've been fascinated by a core question: do recent popular VQA benchmarks really measure visual dependence? Have new SOTA VLMs fixed it? And does scaling up the model fix this? Turns out, the answer is clearly "NO." In our testing, we found that many widely used benchmarks barely need the image. In many cases, replacing the image with a short text caption yields equal or better performance. Counterintuitively, as you scale up the model, visual dependence actually decreases—the larger LLM backbone just answers more from its text pretraining priors! This research made me realize exactly why spatial intelligence and world models are so critical right now. The benchmarks that explicitly test these spatial dimensions are the ones that demand real multimodal capability and cannot be overshadowed by a caption baseline. We also explore some fascinating quirks, like how "self-captioning" (a kind of visual thinking) compares to other-model-captioning. Highly recommend the blog to anyone trying to learn about the realities of multimodal AI like me! 👇 🔗 harvey-fin.github.io/seeing-is-not-…

Harvey Yiyun Fu @harveyiyun

3 weeks ago

Seeing👀 is not reasoning🤔? Introducing our new blog post. We tested 8 open VLMs on 9 VQA benchmarks, and found that a lot of "visual" accuracy isn't visual at all. Many benchmarks can be solved without ever looking at the image, and some VLMs reason better from text than from

2 28 110 26K 73

Introducing VaSE: Value-Aware Stochastic KV Cache Eviction. Reasoning models think in CoT, bloating the KV cache. Eviction caps memory but suffers capability drop. VaSE is a training-free recipe that cuts that cost: keep large-magnitude value states, evict stochastically.

What happens when we build an "AI society" with LLMs? Turns out, it's far more stereotyped than we thought. 🤖📉 Excited to share our new work on Persona Collapse—where distinct AI agents regress into narrow behavioral modes. Surprisingly, the models with the highest per-persona fidelity produce the most stereotyped populations! Huge kudos to the amazing team! @lrzneedresearch @JentseHuang, Weihao Xuan! Congrats to amazing Vivienne for her first paper! (She will be applying for PhD positions, please check her out!) Thanks for the support from the 2077AI foundation! Check out our paper & toolkit here: algoroxyolo.github.io/projects/chame…

Lorenzo Xiao @lrzneedresearch

2 months ago

We have some concerns about the current state of LLM-based social simulation. We benchmarked 10 LLMs on persona simulation. Every model collapses. The "best" ones are the worst offenders. And RLHF actively makes it worse. arxiv.org/pdf/2604.24698

4 16 83 9K 61

On-policy RL has driven the biggest leaps in training coding agents. Extending it to machine learning engineering agents should be a natural next step. But it almost never works. What I mean is, the recipe is right there — standard trajectory-wise GRPO, the same that worked for SWE. However, the problem is that one rollout step on an MLE task may take hours because the agent has to actually train a model on a real dataset at every step (preprocessing, fitting, inference, scoring). So even with the N rollouts in a group running in parallel, a single GRPO run may still take days. Every MLE agent paper I've read has retreated to SFT or offline proxy rewards for exactly this reason, giving up the exploration benefits of on-policy learning. That's why I'm excited about our new paper, SandMLE, which fixes this with a move that sounds almost too reckless to work. The instinct when on-policy RL is too slow is to engineer around it — async rollouts so the trainer doesn't sit idle waiting for slow environments, off-policy or step-wise proxies to avoid running full trajectories at all. But when we profiled where the time was going, the bottleneck had nothing to do with the algorithm. Unlike SWE where execution latency comes from compilation and test logic, MLE latency is overwhelmingly driven by the size of the dataset the ML pipeline has to chew through. Therefore, rather than downsampling existing data (which corrupts evaluation), we built a multi-agent pipeline that procedurally generates diverse synthetic MLE environments from a small seed set. Specifically, we extract the structural DNA of seed tasks (modality, label cardinality, distribution shape), mutate them into new domains (e.g., repurposing animal classification into road damage detection), inject realistic noise, embed deterministic hidden rules connecting features to labels, and construct full evaluation sandboxes with progressive milestone thresholds. Each task is constrained to only 50–200 training samples. The execution speedup is dramatic — average per-step latency drops over 13×, which makes trajectory-wise GRPO go from infeasible to routine. We also designed a dense, milestone-based reward to address the sparse credit assignment problem in long-horizon MLE. The ablation shows this matters — under a sparse reward, the 30B model's medal rate drops from 27.3% to 13.6% and valid submission collapses from 100% to 86.4%. Results across Qwen3-8B, 14B, and 30B-A3B on MLE-bench are consistently strong — 66.9% better performance in medal rate over SFT baselines. It is worth noting that the SFT baselines are not weak— we trained them on high-quality Claude-4.5-Sonnet trajectories. But SandMLE still delivers much larger gains, suggesting that direct environment interaction does teach capabilities that imitation alone does not (as expected). The most convincing evidence to me that the model's intrinsic performance gets improved is the framework-agnostic generalization. We trained exclusively with ReAct but the gains transfer to AIDE, AIRA, and MLE-Agent scaffolds at evaluation time — up to 32.4% better performance in HumanRank on MLE-Dojo. The SFT models, by contrast, are brittle when moved to unfamiliar scaffolds. The 30B SFT model collapses to 17.7% valid submission rate on MLE-Dojo with MLE-Agent, while the 30B SandMLE model achieves 83.9%. SandMLE is teaching genuine engineering reasoning, not scaffold-specific patterns. What I find most interesting beyond the specific result is that none of the hard parts of RL changed here. The algorithm is the same. The reward is conventional. We just shrunk the environment until on-policy learning became affordable. The field has largely treated environment design and RL algorithm design as separate concerns. SandMLE is a concrete case that the environment is itself the lever. When training is too expensive, the instinct is to build cleverer algorithms to tolerate it. However, often the better move is to reshape the environment so the simple algorithm just works. Paper: arxiv.org/pdf/2604.04872

Nice share! I agree with most of that -- taste, judgment, and those higher-level capabilities would be more important. It would be pretty good if researchers could finally be set free from doing a laundry list of "experiments to make reviewer #2 happy". I have some follow-up thoughts on long-context eval to share (forgive my nerdiness :-) ): 1) There have been debates on long-context eval -- basically, whether those metrics, combined with the eval datasets, really measure the model's long-context capability. I think there is a potential confounder that some prepended long context may actually not be that useful, as the topic can shift (e.g., multi-round interactions, storywriting). This may bring us to the "data-or-model" dilemma when debugging. 2) Predicting w/ and w/o context reminds me of the old days when people tried to pretrain the retriever using inverse-cloze tasks (aclanthology.org/P19-1612/). I previously experimented with similar ideas years ago, when we only had BART (direct.mit.edu/tacl/article/d…). It could win a lot on retrieval-focused eval (most current long-context evals fall in this category), but the dataset curation could be tricky, as we, in the end want some downstream utilities. Yes, in general, predicting performance using stats unrelated to applications would be wonderful. But as we move towards more grounded applications, like agentic tasks, those "dirty details" ("scaffold", "harness",..) are hard to ignore, and in many cases, this is sth we could win users and impact.

Glad to know that and thanks for sharing! Yeah, I believe the RL-ed model does not have that much space remaining to tune. We have made some initial exploration on this and developed BF as a unified explanatory framework. But further research and reports are still needed and are more than welcome!

Thanks to my friends at @OpenAI @GoogleDeepMind , and all the other passionate readers (and my kind interviewers lol) for the great questions! Here is a quick FAQ: 1) Why should I care about BF dynamics? Isn't it just about lexical-level diversity? How does it relate to application-grounded diversity (semantic uncertainty, artificial hivemind, etc.)? Great question! At first glance, BF shares similarities with lexical diversity, as it captures the length-averaged entropy for the whole space. However: Noise: Lexical diversity is known to be confounded by vocabulary size and generation length, correlating poorly with BF and leading to noisy interpretations. The Upper Bound: BF serves as the upper bound for all application-grounded diversity. "Semantics" comes from domain-specific grouping of outputs. BF demonstrates exactly how many instances exist for you to group in the first place. Control: Model probabilities and entropies are the most direct steering factors for training and inference. Studying the structure of LLM probability helps us actually control model outputs. (P.S. I already have follow-up work on RLVR rollout design based on this. Get ready for some hardcore MLSys acceleration to boost RLVR while maintaining stability and precision! 👀) 2) Is BF influenced by data contamination / seeing the prompt during training? Yes and no—it depends on how you define "influence" and "see." In the Appendix, we show that common data contamination metrics do not correlate well with BF. The BF dynamic is fundamentally tied to the structural progress of model generation, not just memorized data. When benchmarking, we intentionally chose model-task combinations to avoid severe contamination impacts while ensuring broad evaluation coverage. That said, data contamination remains an active, open problem in the field, and we welcome more discussion on this!

BranchingFactor v1.1 just dropped! 🚀 (Yes — it’s an actively updated paper.) (arxiv.org/abs/2506.17871) As models rely more on post-training, understanding the synergy between pre-training and alignment becomes crucial. Branching Factor (BF) offers a simple way to track the remaining generative potential of a model — since entropy inevitably decreases during generation, BF measures that process. What’s new in v1.1: 1️⃣ Major rewrite We now introduce BF directly — much clearer and easier to read. 2️⃣ Theorem correction + extension Thanks to @StarLi27496427 and Yuwei for catching my misunderstanding of the AEP theorem! We fixed the derivation and extended it to variable-length LLM outputs. The good news: the main result still holds — length-avg log-likelihood can estimate length-avg entropy for sufficiently long generations, in a memory-efficient way. Useful if you want to monitor entropy during training or inference. 3️⃣ Broader evaluation Added experiments on OLMo2 and Qwen3, plus multilingual and long-context tasks. Key findings so far still holds often: 📉 BF decreases during generation ✂️ Alignment significantly reduces BF ⚖️ Interestingly, OLMo2 appears less aggressively shrunk by alignment than Qwen3/Llama3 (preliminary observation). 4️⃣ SFT vs RL analysis We started dissecting how SFT and RL affect BF. Early signals from OLMo2: 🧠 Smaller models: BF shrink mostly happens during SFT (possible memorization effect). 🏗️ Larger models: SFT and RL have comparable impact. Still very preliminary — but it raises interesting questions about how post-training should scale with model size.

@ziqiao_ma @UMengineering BIG Congrats!

@JustinLin610 Heartbroken news. Big thanks for all your work in Qwen! Qwen is nothing without its people. Wish you all the best!

I will be doing my PhD defense today! Come and learn about my Grounded Alignment works! Detailed information (w/Zoom) below: Candidate: Chenghao Yang Date: Friday, January 30, 2026 Time:  2 pm CST Location: John Crerar Library 298 Zoom:  uchicago.zoom.us/j/96014992390?… Meeting ID: 960 1499 2390 Passcode: 644684

keyboard_arrow_left Previous keyboard_arrow_right Next

@m2saxon @jxmnop @universeinanegg Thanks, Michael! This blog looks really nice! Fun Fact: When I initiated my branching factor project with @universeinanegg, I was actually thinking about persona collapsing. Later, we decided to generalize our findings, and that's when Branching Factor came out!

@zhuokaiz Exciting! Glad we all thought about model collaboration! My collaborator @YichenZW has a work collaborating based and aligned models, achieving a better diversity-quality trade-off with user-defined routers. Check it out! (He is looking for interns!):

Yichen (Zach) Wang @ICML @YichenZW

7 months ago

Lack of diversity in your LLM generation? (also noted by Artificial Hivemind, best paper @NeurIPSConf) Time to bring your base model back! An inference-time, token-level collaboration between a base and an aligned model can optimize and control diversity and quality!

2 16 53 13K 34

Check out @YichenZW 's great work collaborating the base and aligned models to achieve a great diversity-quality trade-off! Yichen is an amazing collaborator with strong passion and clear communication. He is looking for a Summer 2026 research intern. Don’t miss out!

Yichen (Zach) Wang @ICML @YichenZW

7 months ago

Lack of diversity in your LLM generation? (also noted by Artificial Hivemind, best paper @NeurIPSConf) Time to bring your base model back! An inference-time, token-level collaboration between a base and an aligned model can optimize and control diversity and quality!

2 16 53 13K 34

@yufei_t Would love to meet up! Just have a follow-up work on your StoryArc work using model collaboration to achieve a good diversity-quality trade-off! Joint with Yichen (@YichenZW ) and Tenghao (@TenghaoHuang45 ) on creative generation (arxiv.org/abs/2511.05650) (A quick demo below)

@j_asminewang So excited that my must-read blogs have just come back!