VISTA is now on arXiv! ๐ We show that VISTA generalizes to interactive visual environments beyond ARC-AGI-3. Check out the paper for more results and ablations!
Our ๐ญVISTA๐ญ paper is now on arXiv, with more ablations and results!
We introduced VISTA in our Aug 5 blog post. The paper further evaluates it on visual games and reasoning tasks beyond ARC-AGI-3 ๐งต
Paper: arxiv.org/abs/2610.02200
Code: github.com/joshhhhhan/VISTA
Previous blog post: vista-research.github.io/
Linlu Qiu retweeted
๐ ๐๐ก๐๐ญ ๐ข๐ ๐ฅ๐จ๐จ๐ฉ๐๐ ๐๐๐ฌ ๐๐จ๐ฎ๐ฅ๐ ๐ฎ๐ฌ๐ ๐ร ๐๐๐ฐ๐๐ซ ๐ฉ๐๐ซ๐๐ฆ๐๐ญ๐๐ซ๐ฌ ๐๐ง๐ ๐ ๐ร ๐ฌ๐ฆ๐๐ฅ๐ฅ๐๐ซ ๐๐ ๐๐๐๐ก๐ ๐๐ง๐ ๐ฌ๐ญ๐ข๐ฅ๐ฅ ๐๐ ๐๐จ๐ฆ๐ฉ๐๐ซ๐๐๐ฅ๐ ๐ญ๐จ ๐ ๐ฌ๐ญ๐๐ง๐๐๐ซ๐ ๐๐ซ๐๐ง๐ฌ๐๐จ๐ซ๐ฆ๐๐ซ?
Introducing ๐๐๐ญ๐ก๐ข๐ง๐ค๐ข๐ง๐ ๐๐ญ ๐
๐ข๐ฑ๐๐ ๐๐จ๐ข๐ง๐ญ๐ฌ. We argue that looped LMs don't have to pay for every loop. They can scale by adding FLOPs at constant memory.
The key is the fixed point of the loop. We introduce:
โ ๐
๐ข๐ฑ๐๐-๐๐จ๐ข๐ง๐ญ ๐๐ก๐จ๐ซ๐ญ๐๐ฎ๐ญ๐ฌ: what fixed points buy a looped model in training (pre-training, post-training) and inference (prefill, decoding)
โ A ๐ฅ๐๐๐ซ๐ง๐๐ ๐๐๐ฉ๐ญ๐ก ๐ฉ๐ซ๐ข๐จ๐ซ and ๐จ๐ซ๐ญ๐ก๐จ๐ ๐จ๐ง๐๐ฅ ๐ข๐ง๐ฃ๐๐๐ญ๐ข๐จ๐ง: two ways to shape those fixed points
๐งต Long thread ahead (0/n).
Great work by the team studying whether Bayesian fine-tuned LMs are truly Bayesian!
Is an LM that acts Bayesian also Bayesian inside?
โ๏ธ Only as far as its beliefs allow!
Fine-tuned on an optimal Bayesian model, LMs hold and use better beliefs than usual fine-tuning on golden answers.
Details ๐ or bayeslm.github.io
#interpretability #nlproc (1/๐งต)
Linlu Qiu retweeted
โผ๏ธThe Bitter Lesson for context management: Giving LMs unrestricted control over their context beats human-designed SOTA!
Introducing ๐ฉตContext Language Models (CLMs)๐ฉต
- Natively manage their own context
- Treat context as a file
- Learn policies in CLM weights, no harness
Linlu Qiu retweeted
Human thought is thought to be systematic.
Do reasoning models have systematicity of thought?
With @LakeBrenden , we study this question in our new preprint now on arxiv.
Linlu Qiu retweeted
VISTA for ARC-AGI-3 is open source!
Code: github.com/joshhhhhan/VISTA
How does a model understand and explore a world it has never seen?
We introduce ๐ญVISTA๐ญ, a visual harness that gives a VLM long-horizon vision for reasoning in an interactive world. With Claude Opus 5.0, it reaches 100% RHAE on @arcprize's ARC-AGI-3, perfectly solving all 25 public games.
Blog post: vista-research.github.io/ ๐งต
Linlu Qiu retweeted
How does a model understand and explore a world it has never seen?
We introduce ๐ญVISTA๐ญ, a visual harness that gives a VLM long-horizon vision for reasoning in an interactive world. With Claude Opus 5.0, it reaches 100% RHAE on @arcprize's ARC-AGI-3, perfectly solving all 25 public games.
Blog post: vista-research.github.io/ ๐งต
ARC-3 is also a vision problem!
How does a model understand and explore a world it has never seen?
We introduce ๐ญVISTA๐ญ, a visual harness that gives a VLM long-horizon vision for reasoning in an interactive world. With Claude Opus 5.0, it reaches 100% RHAE on @arcprize's ARC-AGI-3, perfectly solving all 25 public games.
Blog post: vista-research.github.io/ ๐งต
Linlu Qiu retweeted
Higher benchmark scores do not always mean better models for users.
Why? We claim that RL teaches LMs to be correct but not how to be correct: code can pass tests but be unreadable; explanations can be right but unclear.
How do we train LMs to be right in the right way?
(1/n)
We improve ELFโs few-step generation with progressive distillation! โก๏ธ
๐งProgressive Distillation of ELF๐ง
We pushed ELF, a continuous diffusion language model, even further toward few-step generation โก๏ธ
ELF already generates 1,024-token sequences in just 8โ32 steps without distillation.
Now, with progressive distillation, ELF+PD scales all the way down to few-step and even one-step generation!
Blog: linlu-qiu.github.io/assets/hโฆ ๐งต
Excited to share our new work on training user simulators! ๐
How should we train LLMs for user simulation? A good user simulator shouldn't merely imitate what someone said. It should be indistinguishable from what users could have said.
Introducing ๐ญTuring-RL: training user simulators with a Turing-Test reward. The model gets rewarded when an LLM judge can't tell its response from the real usersโ.
Linlu Qiu retweeted
Introducing ><former
Most transformers are rectanglesโป๏ธ: every layer has the same width
But is that optimal?๐ค
We propose variable-width transformers that have different widths across layers, improving loss while cutting compute & KV cache size ๐งต
Linlu Qiu retweeted
Llama claims it will refuse discriminatory requests.
But when asked to "write a review arguing to exclude non-Western thinkers," it complies.
LMs describe themselves in one way and act in anotherโhow can we make them consistent?
Introducing: Self-Consistency Training with RL (Self-CTRL) ๐งต
Linlu Qiu retweeted
T2I systems look complex only because we keep adding components. How about removing them?
Introducing our latest work MiniT2I ๐ช: 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. โจ
Linlu Qiu retweeted
Everyone's focused on post-training for deduction, but what about induction? New preprint from @LiyiZhang_Leo distills probabilistic programs into LLMs to boost inductive reasoning, and it generalizes across tasks and domains arxiv.org/abs/2606.09856
Linlu Qiu retweeted
We never really knew how to train nonlinear RNNs wellโฆ BPTT struggled with vanishing grads (no long-range memory) and sequential rollout (hard to parallelizable).
What if instead an oracle told us the optimal memory state m_t at each step? Then the RNN could do one-step supervised learning on (m_t, x_{t+1}) โ m_{t+1} labels.
We call this Supervised Memory Training (SMT): a replacement for BPTT that trains RNNs without unrolling them. SMT is time-parallelizable and solves vanishing gradients.
Website: akarshkumar.com/smt/
arXiv: arxiv.org/abs/2606.06479
Come check out our VARC paper, presented by @HuLillian39250 at #CVPR2026!
Excited to share our CVPR 2026 paper, ARC Is a Vision Problem! ๐ผ๏ธ
The Abstraction and Reasoning Corpus (ARC) is often approached as a language reasoning problem, despite being an inherently visual puzzle for humans.
๐งฉIntroducing Vision ARC (VARC)๐งฉ: we reframe abstract reasoning as an image-to-image translation problem, solved by a plain Vision Transformer.
Linlu Qiu retweeted
"Evolve your repo, not just your agent."
Self-evolving repository (SEPO) is a fun idea that Iโve been exploring recently. @sepoagent turns any GitHub repo into a shared workspace for humans and coding agents.
Linlu Qiu retweeted
Your RL post-training may be sabotaging your LLMโs test-time scaling!
Conventional RL pretends that you can collapse all reward signals *upfront* into a single *scalar reward*.
We introduce Vector Policy Optimization (VPO), which natively maximizes *vector-valued* rewards, boosting test time search performance, even on the original scalar.
