Environment generation is the new AI frontier
Environment generation is the missing scaling axis for embodied AI.
Introducing SimWorld Studio: a self-evolving factory for endless interactive 3D env where agents act, fail & learn.
Env-agent co-evolvution improves navigation success 50% → 90%.
From a prompt, our SimCoder writes code to automatically build an interactive world. Agents train inside it. And their performance shapes the next world.
The single unified Voila model achieves comparable to or exceeding SOTA models on voice question answering, ASR, and TTS, respectively. 🔥
4/5
1
136
Voila real-time voice conversations and easy customization of speaker voice and persona, thanks to a set of model designs:
- New streaming multi-level audio tokenizer
- Multi-scale Transformer to blend pretrained LLM and audio modeling
- Multi-task interleaved audio-text alignment formatting and training
3/5
1
1
162
Voice AI has shown huge potentials, evolving through several key stages:
(a) traditional pipeline systems (Siri, Alexa)
(b) ASR-LLM-TTS for greater flexibility
(c) end-to-end audio-in-audio-out models for low-latency and richer voice nuances, and
(d) the emerging generation of full-duplex, autonomous voice agents that continuously listen, reason, and respond proactively
Voila releases various models, code, and data to support research in the space:
- end-to-end voice chat model
- full-duplex autonomous model (preview)
- streaming audio tokenizer
- a library of 1M pre-built voices
HF space: huggingface.co/collections/m…
2/5
Voila! 🤗🔥
Super excited to open-source Voila -- new unified Voice-Language Foundation Models for real-time conversations, audio-in audio-out.
Voila enables to build a voice-based character-ai👩👩👧👧 instantly, with over one million voice persona!
Voila unified model supports:
👥real-time end-to-end speech conversations
👄text-to-speech
👂speech recognition
All with customized voices. You can create and plug in any new voices with just 5s of audio clips🎙️
More details👇
1/5
1/ Though R1 still needs to explore the graph to find the path, it does so **strategically**! When reaching the node scropus, it immediately outputs the full path to the target (scropus → brimpus → rempus → sterpus) without additional exploration. This suggests R1 has latently recognized the correct path.
Interestingly, recent work (arxiv.org/abs/2502.08794) by @andrew_e_cohen @tydsh et al. shows transformers can implicitly compute spectral embeddings of the line graph and use them to find the shortest path—perhaps a clue to R1’s underlying mechanism.
1
2
251
⛓️ Long CoT (O1/R1) style reasoning has gained popularity recently. Rather than directly generating the solution, it exhibits an interesting “search” pattern—e.g., phrases like “let’s think again” resemble backtracking in a search algorithm.
🔍 To clearly illustrate the search behavior of long CoT models, we analyze R1 in a synthetic graph-structured reasoning task (ProsQA from the COCONUT).
Some observations:
-💡Instead of blindly traversing the graph like BFS or DFS, R1 identifies promising nodes, explores them first, and prunes the search space strategically.
-🧐It double-checks alternative choices to ensure correctness—an interesting pattern that mirrors human problem-solving strategies.
-⏳The search trajectory remains excessively long for such a simple problem in terms of token count.
🔗 LLM Reasoners has integrated R1 and ProsQA! Explore the analysis here: shorturl.at/BdlXx
More discussion in the thread! 👇
Big thank you to Gradio @Gradio @_akhaliq , OpenHands @allhands_ai @gneubig , and LLM Reasoners for the open-source tooling and technical support!
Collaborators from @SCSatCMU @LTIatCMU @UCSanDiego @HDSIUCSD @mbzuai @samsungresearch @Samsung_RA @awscloud @allhands_ai
@mdeng34 @jinyuhou0 @frankxu2004 @yikunjwang @ZhitingHu @erranlli @ericxing @EdwardSun0909
6/6
📊Preliminary results show ReasonerAgent outperforming comparable agents by up to 124% on challenging open-web browsing tasks
In particular, success rate on flight search increases from 0% to 32.2%
4/6
ReasonerAgent plans with world model simulation instead of just linear LLM reasoning
The world model helps the agent predict the future and evaluate proposed plans, which mitigates the limitation of autoregressive LLMs for planning
3/6
🤖Thrilled to introduce _ReasonerAgent_ - A fully open source, ready-to-run agent that does research🧐 in a web browser and answers your queries
Use ReasonerAgent to help you:
✈️search for flights,
🛍️compile shopping options,
🗞️research news coverage, etc.
📘Check out more 👇
1/6
📢New Application: PRM-guided search for math reasoning!
We’ve integrated Process Reward Model (PRM)-guided search into MATH reasoning, enabling more efficient inference-time scaling with SGLang. 🔥
👩💻Try it with our script: Inference Scaling for MATH (SGL) github.com/maitrix-org/llm-r…
📰Reference: Inference-time scaling (Snell et al., 2024) arxiv.org/abs/2408.03314
1
2
518
🚀Unlock efficient inference-time scaling with LLM Reasoners - now powered by SGLang!
> 🔥100x acceleration over previous veresion of LLM Reasoners (w/ huggingface)
> With just one line of code change, you can accelerate all example applications in our library (⛓️CoT, 🌲MCTS, 🎶 RAP ...)
> Support Process Reward Model: Now you can efficiently enhance your math reasoning with PRM-guided search
Check out more 👇
🚀LLM Reasoners now supports inference-time scaling approaches for web browsing🌐!
Apply any advanced planning (MCTS, Beam Search, DFS/BFS ...) to any web browsing benches (WebArena, VisualWebArena ...) with #BrowserGym.
📺To monitor/debug the LLM web planning process, we’ve also updated online visualizer. E.g., through the viz, we can see MCTS makes smarter decisions than greedy policies when planning via interaction w/ the environment.
🛣️ More exciting next steps of LLM Reasoners:
> Improving inference-time planning infra for great speed-up!
> Integrate with #AgentLab for most comprehensive web browsing eval and scaling
> Connecting LLM Reasoners to even more reasoning/planning settings, like multi-agent control and coordination.
Stay tuned for more updates soon!












