Loading video...

Video Failed to Load

Go Home

How well do today’s frontier models handle long-horizon, multi-step web agent tasks, such as identifying the top 25 U.S. CS PhD programs with ML/AI faculty likely accepting students and compiling the results into a structured sheet? Check out our new work on Odysseys: Benchmarking Web Agents on Realistic Long...

rsalakhu's profile picture

Russ Salakhutdinov

112,224 subscribers

22,518 views • 3 months ago •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

🚀We are excited to introduce the Tool Decathlon (Toolathlon), a benchmark for language agents on diverse, complex, and realistic tool use. ⭐️32 applications and 600+ tools based on real-world software environments ⭐️Execution-based, reliable evaluation ⭐️Realistic, covering daily and professional scenarios Toolathlon reveals significant shortcomings of SOTA LLMs in realistic tool-use tasks, where Claude Sonnet 4.5 achieves 38.6% success rate. It also indicates a clear gap between open-source and leading proprietary models. Check our blog: Github: Paper: 🧵⬇️
1:06

🚀We are excited to introduce the Tool Decathlon (Toolathlon), a benchmark for language agents on diverse, complex, and realistic tool use. ⭐️32 applications and 600+ tools based on real-world software environments ⭐️Execution-based, reliable evaluation ⭐️Realistic, covering daily and professional scenarios Toolathlon reveals significant shortcomings of SOTA LLMs in realistic tool-use tasks, where Claude Sonnet 4.5 achieves 38.6% success rate. It also indicates a clear gap between open-source and leading proprietary models. Check our blog: Github: Paper: 🧵⬇️

Junxian He

44,126 views • 9 months ago

LLM agents have demonstrated promise in their ability to automate computer tasks, but face challenges with multi-step reasoning and planning. Towards addressing this, we propose an inference-time tree search algorithm for LLM agents to explicitly perform exploration and multi-step planning in interactive web environments. It is the first tree search algorithm for LLM agents that shows effectiveness on realistic and complex web environments: on the challenging VisualWebArena benchmark, applying our search algorithm on top of a GPT-4o agent yields a 39.7% relative increase in success rate compared to the same baseline without search, setting a state-of-the-art success rate of 26.4%. On WebArena, search also yields a 28.0% relative improvement over a baseline agent, setting a competitive success rate of 19.2%.
0:20

LLM agents have demonstrated promise in their ability to automate computer tasks, but face challenges with multi-step reasoning and planning. Towards addressing this, we propose an inference-time tree search algorithm for LLM agents to explicitly perform exploration and multi-step planning in interactive web environments. It is the first tree search algorithm for LLM agents that shows effectiveness on realistic and complex web environments: on the challenging VisualWebArena benchmark, applying our search algorithm on top of a GPT-4o agent yields a 39.7% relative increase in success rate compared to the same baseline without search, setting a state-of-the-art success rate of 26.4%. On WebArena, search also yields a 28.0% relative improvement over a baseline agent, setting a competitive success rate of 19.2%.

Jing Yu Koh

124,894 views • 2 years ago

Introducing EdgeBench, a benchmark designed to study how agents learn from environments over at least 12~72-hour runs. We find that performance follows a log-sigmoid function of environment interaction time with high precision. EdgeBench is built with three ingredients: - 🌍 Real & Diverse: 134 real-world tasks across 6 task categories, spanning scientific problems, professional knowledge work, software engineering, optimization, formal math, and games. - ⏳ Ultra-Long-Horizon: Each task supports 12–72 hours of agent work. Recorded human effort averages 57.2 hours. - 🔁 Informative Feedback: Agents receive real-world feedback for continuous improvement. After 38,000 hours of agent runs on EdgeBench, a scaling law for learning from environments emerges: - 📈 As agents interact with task environments over time, their aggregate performance is precisely fit by a log-sigmoid function. - 🧠 This phenomenon can be explained by an elegant theory of graph exploration. We are releasing an initial 51 of the 134 tasks, together with the full evaluation framework, to help advance long-horizon agent research. Check our blog & paper for more findings! Blog Paper GitHub Dataset Details below 👇🧵
0:52

Introducing EdgeBench, a benchmark designed to study how agents learn from environments over at least 12~72-hour runs. We find that performance follows a log-sigmoid function of environment interaction time with high precision. EdgeBench is built with three ingredients: - 🌍 Real & Diverse: 134 real-world tasks across 6 task categories, spanning scientific problems, professional knowledge work, software engineering, optimization, formal math, and games. - ⏳ Ultra-Long-Horizon: Each task supports 12–72 hours of agent work. Recorded human effort averages 57.2 hours. - 🔁 Informative Feedback: Agents receive real-world feedback for continuous improvement. After 38,000 hours of agent runs on EdgeBench, a scaling law for learning from environments emerges: - 📈 As agents interact with task environments over time, their aggregate performance is precisely fit by a log-sigmoid function. - 🧠 This phenomenon can be explained by an elegant theory of graph exploration. We are releasing an initial 51 of the 134 tasks, together with the full evaluation framework, to help advance long-horizon agent research. Check our blog & paper for more findings! Blog Paper GitHub Dataset Details below 👇🧵

Deyao Zhu

360,827 views • 1 month ago

Super excited to share 🧠MLGym 🦾 – the first Gym environment for AI Research Agents 🤖🔬 We introduce MLGym and MLGym-Bench, a new framework and benchmark for evaluating and developing LLM agents on AI research tasks. The key contributions of our work are: 🕹️ Enables the exploration of different training algorithms for AI Research Agents such as RL 🛠️ Provides a flexible evaluation framework that can accommodate different artifacts such as models, algorithms, or predictions 🤖 Allows researchers to evaluate any model without the need to develop a custom agentic harness 🎯 Introduces 13 diverse open-ended AI Research tasks for evaluating AI Research Agents on a wide range of domains such as computer vision, natural language processing, reinforcement learning, game theory, and logical reasoning. 📈 Proposes a new evaluation metric for AI Research Agents MLGym makes it easy to: 1) Add new tasks 2) Evaluate new models 3) Integrate new agents Check out a video of the MLGym Agent to see how it performs the full pipeline of idea generation💡, implementation 👩‍💻, experimentation 👩‍🔬, and iteration 🔄 to improve on ML tasks. Huge thanks to the exceptionally talented Deepak Nathani who led this work and to all the other amazing collaborators who made this possible 🙏🫶🚀
1:32

Super excited to share 🧠MLGym 🦾 – the first Gym environment for AI Research Agents 🤖🔬 We introduce MLGym and MLGym-Bench, a new framework and benchmark for evaluating and developing LLM agents on AI research tasks. The key contributions of our work are: 🕹️ Enables the exploration of different training algorithms for AI Research Agents such as RL 🛠️ Provides a flexible evaluation framework that can accommodate different artifacts such as models, algorithms, or predictions 🤖 Allows researchers to evaluate any model without the need to develop a custom agentic harness 🎯 Introduces 13 diverse open-ended AI Research tasks for evaluating AI Research Agents on a wide range of domains such as computer vision, natural language processing, reinforcement learning, game theory, and logical reasoning. 📈 Proposes a new evaluation metric for AI Research Agents MLGym makes it easy to: 1) Add new tasks 2) Evaluate new models 3) Integrate new agents Check out a video of the MLGym Agent to see how it performs the full pipeline of idea generation💡, implementation 👩‍💻, experimentation 👩‍🔬, and iteration 🔄 to improve on ML tasks. Huge thanks to the exceptionally talented Deepak Nathani who led this work and to all the other amazing collaborators who made this possible 🙏🫶🚀

Roberta Raileanu

105,052 views • 1 year ago

JARVIS-1: Open-World Multi-task Agents with Memory-Augmented Multimodal Language Models paper page: Achieving human-like planning and control with multimodal observations in an open world is a key milestone for more functional generalist agents. Existing approaches can handle certain long-horizon tasks in an open world. However, they still struggle when the number of open-world tasks could potentially be infinite and lack the capability to progressively enhance task completion as game time progresses. We introduce JARVIS-1, an open-world agent that can perceive multimodal input (visual observations and human instructions), generate sophisticated plans, and perform embodied control, all within the popular yet challenging open-world Minecraft universe. Specifically, we develop JARVIS-1 on top of pre-trained multimodal language models, which map visual observations and textual instructions to plans. The plans will be ultimately dispatched to the goal-conditioned controllers. We outfit JARVIS-1 with a multimodal memory, which facilitates planning using both pre-trained knowledge and its actual game survival experiences. In our experiments, JARVIS-1 exhibits nearly perfect performances across over 200 varying tasks from the Minecraft Universe Benchmark, ranging from entry to intermediate levels. JARVIS-1 has achieved a completion rate of 12.5% in the long-horizon diamond pickaxe task. This represents a significant increase up to 5 times compared to previous records. Furthermore, we show that JARVIS-1 is able to self-improve following a life-long learning paradigm thanks to multimodal memory, sparking a more general intelligence and improved autonomy.
0:22

JARVIS-1: Open-World Multi-task Agents with Memory-Augmented Multimodal Language Models paper page: Achieving human-like planning and control with multimodal observations in an open world is a key milestone for more functional generalist agents. Existing approaches can handle certain long-horizon tasks in an open world. However, they still struggle when the number of open-world tasks could potentially be infinite and lack the capability to progressively enhance task completion as game time progresses. We introduce JARVIS-1, an open-world agent that can perceive multimodal input (visual observations and human instructions), generate sophisticated plans, and perform embodied control, all within the popular yet challenging open-world Minecraft universe. Specifically, we develop JARVIS-1 on top of pre-trained multimodal language models, which map visual observations and textual instructions to plans. The plans will be ultimately dispatched to the goal-conditioned controllers. We outfit JARVIS-1 with a multimodal memory, which facilitates planning using both pre-trained knowledge and its actual game survival experiences. In our experiments, JARVIS-1 exhibits nearly perfect performances across over 200 varying tasks from the Minecraft Universe Benchmark, ranging from entry to intermediate levels. JARVIS-1 has achieved a completion rate of 12.5% in the long-horizon diamond pickaxe task. This represents a significant increase up to 5 times compared to previous records. Furthermore, we show that JARVIS-1 is able to self-improve following a life-long learning paradigm thanks to multimodal memory, sparking a more general intelligence and improved autonomy.

AK

141,440 views • 2 years ago

One of the things I’m most excited about this year is building agents that can work productively for hours, days, or weeks. Coding agents are starting to become very competent at this, but what about computer use agents? Our new benchmark, Odysseys (co-led with Lawrence Jang) is a set of 200 new tasks derived from real world browsing behavior that measure long horizon web navigation capabilities (potentially up to hours of web browsing work). Interestingly, we find that frontier CUAs are already surprisingly good at working productively for up to an hour on these tasks, but there’s a lot of work to be done in making them even more efficient. Like every other AI researcher, my real dream is to open a cafe once we solve ASI. So, here’s Opus 4.6 doing some market research for me ("I want to do market research on the most popular cafes in Singapore. Analyse the menus of the top 10 cafes in Singapore (by Google reviews/ratings), and make sure we include at least 1 from the North/South/East/West/Central regions of Singapore. Keep the relevant pages of each cafe open, and summarise their pricing, menu offerings, unique selling points, making sure to reference which tab is opened for each cafe. For each cafe, also help me figure out how long it would take to get to it from Tampines MRT, and include this in your final summary."). I was very impressed to see Opus 4.6 complete this task after working for 52 mins, satisfying all 7 rubrics that corresponded to this task. It provided a very nice markdown summary at the end that gave me all the information I asked for!
1:11

One of the things I’m most excited about this year is building agents that can work productively for hours, days, or weeks. Coding agents are starting to become very competent at this, but what about computer use agents? Our new benchmark, Odysseys (co-led with Lawrence Jang) is a set of 200 new tasks derived from real world browsing behavior that measure long horizon web navigation capabilities (potentially up to hours of web browsing work). Interestingly, we find that frontier CUAs are already surprisingly good at working productively for up to an hour on these tasks, but there’s a lot of work to be done in making them even more efficient. Like every other AI researcher, my real dream is to open a cafe once we solve ASI. So, here’s Opus 4.6 doing some market research for me ("I want to do market research on the most popular cafes in Singapore. Analyse the menus of the top 10 cafes in Singapore (by Google reviews/ratings), and make sure we include at least 1 from the North/South/East/West/Central regions of Singapore. Keep the relevant pages of each cafe open, and summarise their pricing, menu offerings, unique selling points, making sure to reference which tab is opened for each cafe. For each cafe, also help me figure out how long it would take to get to it from Tampines MRT, and include this in your final summary."). I was very impressed to see Opus 4.6 complete this task after working for 52 mins, satisfying all 7 rubrics that corresponded to this task. It provided a very nice markdown summary at the end that gave me all the information I asked for!

Jing Yu Koh

49,418 views • 3 months ago

📢 Announcing one of the most exciting works from us this year on **scalable robot policy evaluation through real-to-sim transfer**, moving toward a scalable evaluation engine with structured world models that capture the appearance, geometry, and dynamics of environments involving deformable objects. 🤖 Evaluation remains one of the biggest bottlenecks in building general-purpose robots. Today, robots are still evaluated only in the real world, which is **orders of magnitude slower** than the development of language agents. We propose a new framework where simulation performance **strongly correlates** with the real world (r > 0.9), even for deformable objects. The key difference from existing work lies in the correlation between simulation and reality: if a robot model performs better in the digital world, does it also perform better in the real world? This question has long made people hesitant about simulation-based evaluation — especially for deformable objects. We are changing that. Our pipeline achieves effective real-to-sim transfer, establishing **state-of-the-art correlation** between simulation and reality for deformable object manipulation. It provides a **scalable and reproducible evaluation engine** for robot learning. 🌐
3:04

📢 Announcing one of the most exciting works from us this year on **scalable robot policy evaluation through real-to-sim transfer**, moving toward a scalable evaluation engine with structured world models that capture the appearance, geometry, and dynamics of environments involving deformable objects. 🤖 Evaluation remains one of the biggest bottlenecks in building general-purpose robots. Today, robots are still evaluated only in the real world, which is **orders of magnitude slower** than the development of language agents. We propose a new framework where simulation performance **strongly correlates** with the real world (r > 0.9), even for deformable objects. The key difference from existing work lies in the correlation between simulation and reality: if a robot model performs better in the digital world, does it also perform better in the real world? This question has long made people hesitant about simulation-based evaluation — especially for deformable objects. We are changing that. Our pipeline achieves effective real-to-sim transfer, establishing **state-of-the-art correlation** between simulation and reality for deformable object manipulation. It provides a **scalable and reproducible evaluation engine** for robot learning. 🌐

Yunzhu Li

39,960 views • 9 months ago

Arena reached a $100M annual revenue run rate just 8 months after launching our evaluation product. We started as a research project at UC Berkeley with a simple mission: measure AI progress through real-world use. As AI shifts from chatbots to agents taking on longer, higher-stakes work, the problem matters more than ever. Today, Arena measures real-world AI utility with a community of tens of millions. With Agent Arena, we’re evaluating long-running agents on complex, real-world tasks - how they use tools, adapt to feedback, recover from errors, and accomplish goals set by humans. We are excited to keep deepening our work in agentic evaluations. Here’s Anastasios Nikolas Angelopoulos on what this milestone means and where we go from here:
1:26

Arena reached a $100M annual revenue run rate just 8 months after launching our evaluation product. We started as a research project at UC Berkeley with a simple mission: measure AI progress through real-world use. As AI shifts from chatbots to agents taking on longer, higher-stakes work, the problem matters more than ever. Today, Arena measures real-world AI utility with a community of tens of millions. With Agent Arena, we’re evaluating long-running agents on complex, real-world tasks - how they use tools, adapt to feedback, recover from errors, and accomplish goals set by humans. We are excited to keep deepening our work in agentic evaluations. Here’s Anastasios Nikolas Angelopoulos on what this milestone means and where we go from here:

Arena.ai

210,182 views • 1 month ago

New Short Course: Building AI Browser Agents! Learn how to build AI agents that interact and take actions on websites in this course, created in partnership with and taught by and @namangarg0, Co-founders of AGI Inc. AI browser agents can log into websites, fill out forms, click through web pages, or even place orders online for you. They use both visual information, like screenshots, and structural data, like the HTML or Document Object Model (DOM) of a web page, to reason and take action. With the complexity of webpages and multiple possible actions at each step, it can be challenging for an AI browser agent to complete an assigned task. Because these agents run long action sequences, a single error—like clicking the wrong button or misreading a field—can lead to unexpected outcomes or errors that compound over time. In this course, you'll understand how autonomous web agents work, their current limitations, and how AgentQ enables them to improve through self-correction. In detail, you'll: - Learn what web agents are, how they automate tasks online, their architecture, key components, limitations, and an overview of their decision-making strategies. - Build a web agent that can scrape website and return course recommendations in a structured output format. - Build an autonomous web agent that can execute multiple tasks, such as finding and summarizing webpages, filling out a form, and signing up for a newsletter. - Explore AgentQ, a framework that enables agents to self-correct by combining Monte Carlo Tree Search (MCTS), a self-critique mechanism for continuous improvement, and Direct Preference Optimization (DPO). - Deep dive into MCTS, learn how it finds an effective path, illustrated by an example of Gridworld animation, and use AgentQ to complete web tasks. - Understand AI agents' current state and future directions—including key factors shaping their evolution, such as hardware, algorithm innovation, and data availability. By the end of this course, you will have hands-on experience building browser agents and a deeper understanding of how to make them more robust and reliable. Please sign up here:
2:06

New Short Course: Building AI Browser Agents! Learn how to build AI agents that interact and take actions on websites in this course, created in partnership with and taught by and @namangarg0, Co-founders of AGI Inc. AI browser agents can log into websites, fill out forms, click through web pages, or even place orders online for you. They use both visual information, like screenshots, and structural data, like the HTML or Document Object Model (DOM) of a web page, to reason and take action. With the complexity of webpages and multiple possible actions at each step, it can be challenging for an AI browser agent to complete an assigned task. Because these agents run long action sequences, a single error—like clicking the wrong button or misreading a field—can lead to unexpected outcomes or errors that compound over time. In this course, you'll understand how autonomous web agents work, their current limitations, and how AgentQ enables them to improve through self-correction. In detail, you'll: - Learn what web agents are, how they automate tasks online, their architecture, key components, limitations, and an overview of their decision-making strategies. - Build a web agent that can scrape website and return course recommendations in a structured output format. - Build an autonomous web agent that can execute multiple tasks, such as finding and summarizing webpages, filling out a form, and signing up for a newsletter. - Explore AgentQ, a framework that enables agents to self-correct by combining Monte Carlo Tree Search (MCTS), a self-critique mechanism for continuous improvement, and Direct Preference Optimization (DPO). - Deep dive into MCTS, learn how it finds an effective path, illustrated by an example of Gridworld animation, and use AgentQ to complete web tasks. - Understand AI agents' current state and future directions—including key factors shaping their evolution, such as hardware, algorithm innovation, and data availability. By the end of this course, you will have hands-on experience building browser agents and a deeper understanding of how to make them more robust and reliable. Please sign up here:

Andrew Ng

186,133 views • 1 year ago

Today, we publicly released RoboCasa365, a large-scale simulation benchmark for training and systematically evaluating generalist robot models. Built upon our original RoboCasa framework, it offers: • 2,500 realistic kitchen environments; • 365 everyday tasks (basic skills + long-horizon mobile manipulation); • Over 3,200 objects with many articulated fixtures/appliances. All are designed for fully controlled, reproducible benchmarking of robotic policies. Progress in robotic foundation models is real. But it’s still hard to answer basic questions like: How close are we to general-purpose autonomy? What factors drive generalization? What are the model/data scaling curves like? Real-world eval is slow and noisy, and existing sims (like LIBERO, which we built 3 years ago) often lack sufficient task and scene diversity. This benchmark comes with 2,200+ hours of demonstrations and 500K+ trajectories to support studies of multi-task training, pretraining, and continual learning at scale. Check it out at
1:16

Today, we publicly released RoboCasa365, a large-scale simulation benchmark for training and systematically evaluating generalist robot models. Built upon our original RoboCasa framework, it offers: • 2,500 realistic kitchen environments; • 365 everyday tasks (basic skills + long-horizon mobile manipulation); • Over 3,200 objects with many articulated fixtures/appliances. All are designed for fully controlled, reproducible benchmarking of robotic policies. Progress in robotic foundation models is real. But it’s still hard to answer basic questions like: How close are we to general-purpose autonomy? What factors drive generalization? What are the model/data scaling curves like? Real-world eval is slow and noisy, and existing sims (like LIBERO, which we built 3 years ago) often lack sufficient task and scene diversity. This benchmark comes with 2,200+ hours of demonstrations and 500K+ trajectories to support studies of multi-task training, pretraining, and continual learning at scale. Check it out at

Yuke Zhu

24,027 views • 5 months ago

Can vision-language-action (VLA) models generalize to diverse OOD tasks and align with customized objectives? 🤔 🚀 We introduce GRAPE, a plug-and-play algorithm to generalize robot policies via preference alignment. GRAPE unfolds three benefits to boost the generalizability of VLAs: 👉1. GRAPE aligns VLAs on a trajectory level and endows the model with the ability for global decision-making, instead of merely cloning behavior; 👉2. GRAPE implicitly models reward from both successful and failed trials to boost generalizability to diverse tasks; 👉3. GRAPE adopts a scalable preference synthesis algorithm to rank trajectories with preferences that align with arbitrary objectives. Our experiments on a diverse array of real-world and simulated robotic tasks reveal: 1⃣GRAPE enhances the performance of state-of-the-art VLA models, increasing success rates on in-domain and unseen manipulation tasks by 51.79% and 60.36%; 2⃣GRAPE is versatile to be aligned with diverse objectives and reduce collision rates by 44.31% or rollout length by 11.15% when aligning towards safer or more efficient manipulation policy, respectively. Check out our full project for more details: 🔥 Paper: 🔥 Project: 🔥 Code:
1:06

Can vision-language-action (VLA) models generalize to diverse OOD tasks and align with customized objectives? 🤔 🚀 We introduce GRAPE, a plug-and-play algorithm to generalize robot policies via preference alignment. GRAPE unfolds three benefits to boost the generalizability of VLAs: 👉1. GRAPE aligns VLAs on a trajectory level and endows the model with the ability for global decision-making, instead of merely cloning behavior; 👉2. GRAPE implicitly models reward from both successful and failed trials to boost generalizability to diverse tasks; 👉3. GRAPE adopts a scalable preference synthesis algorithm to rank trajectories with preferences that align with arbitrary objectives. Our experiments on a diverse array of real-world and simulated robotic tasks reveal: 1⃣GRAPE enhances the performance of state-of-the-art VLA models, increasing success rates on in-domain and unseen manipulation tasks by 51.79% and 60.36%; 2⃣GRAPE is versatile to be aligned with diverse objectives and reduce collision rates by 44.31% or rollout length by 11.15% when aligning towards safer or more efficient manipulation policy, respectively. Check out our full project for more details: 🔥 Paper: 🔥 Project: 🔥 Code:

Huaxiu Yao

19,988 views • 1 year ago

The verdict is in! Frontier models can pass the bar, yet they struggle on comprehensive legal research Today we're releasing Legal Research Bench, a benchmark that measures models’ ability to solve realistic legal research tasks across eight areas of U.S. law Instead of awarding partial credit, Legal Research Bench measures whether a model can conduct exhaustive legal analysis. We grade against a strict, all-pass rubric written by practicing lawyers. A model only receives full credit if every required legal element is correct Claude Opus 4.8 leads with 43.8% all-pass accuracy, followed by GPT 5.5 (40.4%) and Claude Sonnet 4.6 (38.5%). While top models score around 80% with partial credit, none exceed 44% when every required legal element must be correct The gap between partial and all-pass accuracy shows how difficult it remains for AI to produce complete, reliable legal research. We hope that Legal Research Bench helps better measure, and ultimately close that gap Lots of exciting work happening in Legal AI from Harvey and Crosby. Excited for the legal research benchmarks ahead!
1:33

The verdict is in! Frontier models can pass the bar, yet they struggle on comprehensive legal research Today we're releasing Legal Research Bench, a benchmark that measures models’ ability to solve realistic legal research tasks across eight areas of U.S. law Instead of awarding partial credit, Legal Research Bench measures whether a model can conduct exhaustive legal analysis. We grade against a strict, all-pass rubric written by practicing lawyers. A model only receives full credit if every required legal element is correct Claude Opus 4.8 leads with 43.8% all-pass accuracy, followed by GPT 5.5 (40.4%) and Claude Sonnet 4.6 (38.5%). While top models score around 80% with partial credit, none exceed 44% when every required legal element must be correct The gap between partial and all-pass accuracy shows how difficult it remains for AI to produce complete, reliable legal research. We hope that Legal Research Bench helps better measure, and ultimately close that gap Lots of exciting work happening in Legal AI from Harvey and Crosby. Excited for the legal research benchmarks ahead!

Vals AI

17,351 views • 1 month ago

PointOdyssey: A Large-Scale Synthetic Dataset for Long-Term Point Tracking paper page: introduce PointOdyssey, a large-scale synthetic dataset, and data generation framework, for the training and evaluation of long-term fine-grained tracking algorithms. Our goal is to advance the state-of-the-art by placing emphasis on long videos with naturalistic motion. Toward the goal of naturalism, we animate deformable characters using real-world motion capture data, we build 3D scenes to match the motion capture environments, and we render camera viewpoints using trajectories mined via structure-from-motion on real videos. We create combinatorial diversity by randomizing character appearance, motion profiles, materials, lighting, 3D assets, and atmospheric effects. Our dataset currently includes 104 videos, averaging 2,000 frames long, with orders of magnitude more correspondence annotations than prior work. We show that existing methods can be trained from scratch in our dataset and outperform the published variants. Finally, we introduce modifications to the PIPs point tracking method, greatly widening its temporal receptive field, which improves its performance on PointOdyssey as well as on two real-world benchmarks.
0:33

PointOdyssey: A Large-Scale Synthetic Dataset for Long-Term Point Tracking paper page: introduce PointOdyssey, a large-scale synthetic dataset, and data generation framework, for the training and evaluation of long-term fine-grained tracking algorithms. Our goal is to advance the state-of-the-art by placing emphasis on long videos with naturalistic motion. Toward the goal of naturalism, we animate deformable characters using real-world motion capture data, we build 3D scenes to match the motion capture environments, and we render camera viewpoints using trajectories mined via structure-from-motion on real videos. We create combinatorial diversity by randomizing character appearance, motion profiles, materials, lighting, 3D assets, and atmospheric effects. Our dataset currently includes 104 videos, averaging 2,000 frames long, with orders of magnitude more correspondence annotations than prior work. We show that existing methods can be trained from scratch in our dataset and outperform the published variants. Finally, we introduce modifications to the PIPs point tracking method, greatly widening its temporal receptive field, which improves its performance on PointOdyssey as well as on two real-world benchmarks.

AK

122,560 views • 3 years ago

We’re launching Optima. Now anyone can create a custom benchmark for their use case, leveraging Artificial Analysis’ leading research and platform Building and running benchmarks is difficult. We have distilled Artificial Analysis’ research and experience developing benchmarks into Optima, a new platform for benchmarking models on your own workloads and comparing performance, speed and cost efficiency. Optima allows you to find the best model for your task, or an equally performant alternative to your current setup at 10x lower cost or time per task. We’ve integrated Artificial Analysis' research and experience in benchmarks across the Optima workflow: ➤ Build benchmarks based on your own data and use cases: There are three ways to build a benchmark with Optima. Upload an existing evaluation dataset from your own files or Hugging Face, or import agent traces from platforms including Arize AI, Braintrust and langfuse.com. Install the Optima skill to build a benchmark using context from your coding environment and previous sessions. Or simply describe your use case and provide example inputs and outputs, and Optima will build the benchmark for you ➤ Run across the latest models: Run the same benchmark across leading models in a single click, and keep your leaderboard up to date as soon as new models are released ➤ Bring Artificial Analysis grading to your own benchmark: Evaluate responses against objective rubric criteria or using the same pairwise judging approach used for Artificial Analysis benchmarks including GDPval-AA and AA-Briefcase. For pairwise judging, select your preferred responses from a sample and Optima uses those preferences to rank models across your test set ➤ Compare performance, cost and time efficiency: Optima measures more than model performance. Cost per Task and Time per Task are tracked alongside benchmark scores, with category-level results and support for custom metrics, allowing you to compare the tradeoffs between models for your specific use case Ahead of launch, here are examples questions our beta testers answered with Optima: ➤ Which model can save me 10x the cost without a meaningful decrease in quality for my finance & accounting agent? ➤ Which model best matches the writing style of lawyers for my legal agent? ➤ Which model can best identify different elements in my custom image dataset? Optima is available today. Build your own benchmark at
2:33

We’re launching Optima. Now anyone can create a custom benchmark for their use case, leveraging Artificial Analysis’ leading research and platform Building and running benchmarks is difficult. We have distilled Artificial Analysis’ research and experience developing benchmarks into Optima, a new platform for benchmarking models on your own workloads and comparing performance, speed and cost efficiency. Optima allows you to find the best model for your task, or an equally performant alternative to your current setup at 10x lower cost or time per task. We’ve integrated Artificial Analysis' research and experience in benchmarks across the Optima workflow: ➤ Build benchmarks based on your own data and use cases: There are three ways to build a benchmark with Optima. Upload an existing evaluation dataset from your own files or Hugging Face, or import agent traces from platforms including Arize AI, Braintrust and langfuse.com. Install the Optima skill to build a benchmark using context from your coding environment and previous sessions. Or simply describe your use case and provide example inputs and outputs, and Optima will build the benchmark for you ➤ Run across the latest models: Run the same benchmark across leading models in a single click, and keep your leaderboard up to date as soon as new models are released ➤ Bring Artificial Analysis grading to your own benchmark: Evaluate responses against objective rubric criteria or using the same pairwise judging approach used for Artificial Analysis benchmarks including GDPval-AA and AA-Briefcase. For pairwise judging, select your preferred responses from a sample and Optima uses those preferences to rank models across your test set ➤ Compare performance, cost and time efficiency: Optima measures more than model performance. Cost per Task and Time per Task are tracked alongside benchmark scores, with category-level results and support for custom metrics, allowing you to compare the tradeoffs between models for your specific use case Ahead of launch, here are examples questions our beta testers answered with Optima: ➤ Which model can save me 10x the cost without a meaningful decrease in quality for my finance & accounting agent? ➤ Which model best matches the writing style of lawyers for my legal agent? ➤ Which model can best identify different elements in my custom image dataset? Optima is available today. Build your own benchmark at

Artificial Analysis

118,834 views • 3 days ago

Introducing ALE-Bench, ALE-Agent! Towards Automating Long-Horizon Algorithm Engineering for Hard Optimization Problems Blog: Paper: ALE-Bench is a coding benchmark primarily focused on hard optimization (NP-hard) problems. We developed this benchmark with AtCoder Inc., a leading coding contest platform company. What makes ALE-Bench unique is its focus on hard optimization problems that demand long-horizon and creative reasoning. It’s open-ended, in the sense that true optima are out of reach (NP-hard) and scores can continuously improve. We believe this benchmark has the potential to become one of the key benchmarks for reasoning and coding in the next generation. ALE-Agent is our end-to-end agent that we specifically designed for this challenging domain. In fact, our ALE-Agent has already built an impressive track record in the wild! In May 2025, our agent participated in a live AtCoder Heuristic Competition (AHC), alongside 1,000 other participants in real-time. AHC is considered to be one of the most challenging coding competitions in this domain. Our ALE-Agent achieved an impressive ranking of 21st out of 1,000 human participants in the competition (top 2%), marking a turning point for AI discovery of solutions to hard optimization problems with a wide spectrum of important real world applications such as logistics, routing, packing, factory production planning, power-grid balancing. We look forward to applying this technology to real industrial optimization opportunities. Building on the insights from this study, Sakana AI will continue to tackle the challenge of developing AI with even greater algorithm engineering capabilities. ALE-Bench Dataset: ALE-Bench Code: This research was conducted in collaboration with AtCoder Inc. (AtCoder). We are deeply grateful for their outstanding expertise and contributions in optimization and algorithms, which were invaluable in providing data, analyzing results, and enabling our AI agent’s participation in their contests.
0:54

Introducing ALE-Bench, ALE-Agent! Towards Automating Long-Horizon Algorithm Engineering for Hard Optimization Problems Blog: Paper: ALE-Bench is a coding benchmark primarily focused on hard optimization (NP-hard) problems. We developed this benchmark with AtCoder Inc., a leading coding contest platform company. What makes ALE-Bench unique is its focus on hard optimization problems that demand long-horizon and creative reasoning. It’s open-ended, in the sense that true optima are out of reach (NP-hard) and scores can continuously improve. We believe this benchmark has the potential to become one of the key benchmarks for reasoning and coding in the next generation. ALE-Agent is our end-to-end agent that we specifically designed for this challenging domain. In fact, our ALE-Agent has already built an impressive track record in the wild! In May 2025, our agent participated in a live AtCoder Heuristic Competition (AHC), alongside 1,000 other participants in real-time. AHC is considered to be one of the most challenging coding competitions in this domain. Our ALE-Agent achieved an impressive ranking of 21st out of 1,000 human participants in the competition (top 2%), marking a turning point for AI discovery of solutions to hard optimization problems with a wide spectrum of important real world applications such as logistics, routing, packing, factory production planning, power-grid balancing. We look forward to applying this technology to real industrial optimization opportunities. Building on the insights from this study, Sakana AI will continue to tackle the challenge of developing AI with even greater algorithm engineering capabilities. ALE-Bench Dataset: ALE-Bench Code: This research was conducted in collaboration with AtCoder Inc. (AtCoder). We are deeply grateful for their outstanding expertise and contributions in optimization and algorithms, which were invaluable in providing data, analyzing results, and enabling our AI agent’s participation in their contests.

Sakana AI

237,195 views • 1 year ago

Starting today, open source is leading the way. Introducing Llama 3.1: Our most capable models yet. Today we’re releasing a collection of new Llama 3.1 models including our long awaited 405B. These models deliver improved reasoning capabilities, a larger 128K token context window and improved support for 8 languages among other improvements. Llama 3.1 405B rivals leading closed source models on state-of-the-art capabilities across a range of tasks in general knowledge, steerability, math, tool use and multilingual translation. The models are available to download now directly from Meta or Hugging Face. With today’s release the ecosystem is also ready to go with 25+ partners rolling out our latest models — including Amazon Web Services, NVIDIA, Databricks, Groq Inc, Dell Technologies, Microsoft Azure and Google Cloud ready on day one. More details in the full announcement ➡️ Download Llama 3.1 models ➡️ With these releases we’re setting the stage for unprecedented new opportunities and we can’t wait to see the innovation our newest models will unlock across all levels of the AI community.
2:57

Starting today, open source is leading the way. Introducing Llama 3.1: Our most capable models yet. Today we’re releasing a collection of new Llama 3.1 models including our long awaited 405B. These models deliver improved reasoning capabilities, a larger 128K token context window and improved support for 8 languages among other improvements. Llama 3.1 405B rivals leading closed source models on state-of-the-art capabilities across a range of tasks in general knowledge, steerability, math, tool use and multilingual translation. The models are available to download now directly from Meta or Hugging Face. With today’s release the ecosystem is also ready to go with 25+ partners rolling out our latest models — including Amazon Web Services, NVIDIA, Databricks, Groq Inc, Dell Technologies, Microsoft Azure and Google Cloud ready on day one. More details in the full announcement ➡️ Download Llama 3.1 models ➡️ With these releases we’re setting the stage for unprecedented new opportunities and we can’t wait to see the innovation our newest models will unlock across all levels of the AI community.

AI at Meta

1,269,309 views • 2 years ago

3D-LLM: Injecting the 3D World into Large Language Models paper page: Large language models (LLMs) and Vision-Language Models (VLMs) have been proven to excel at multiple tasks, such as commonsense reasoning. Powerful as these models can be, they are not grounded in the 3D physical world, which involves richer concepts such as spatial relationships, affordances, physics, layout, and so on. In this work, we propose to inject the 3D world into large language models and introduce a whole new family of 3D-LLMs. Specifically, 3D-LLMs can take 3D point clouds and their features as input and perform a diverse set of 3D-related tasks, including captioning, dense captioning, 3D question answering, task decomposition, 3D grounding, 3D-assisted dialog, navigation, and so on. Using three types of prompting mechanisms that we design, we are able to collect over 300k 3D-language data covering these tasks. To efficiently train 3D-LLMs, we first utilize a 3D feature extractor that obtains 3D features from rendered multi- view images. Then, we use 2D VLMs as our backbones to train our 3D-LLMs. By introducing a 3D localization mechanism, 3D-LLMs can better capture 3D spatial information. Experiments on ScanQA show that our model outperforms state-of-the-art baselines by a large margin (e.g., the BLEU-1 score surpasses state-of-the-art score by 9%). Furthermore, experiments on our held-in datasets for 3D captioning, task composition, and 3D-assisted dialogue show that our model outperforms 2D VLMs. Qualitative examples also show that our model could perform more tasks beyond the scope of existing LLMs and VLMs.
0:50

3D-LLM: Injecting the 3D World into Large Language Models paper page: Large language models (LLMs) and Vision-Language Models (VLMs) have been proven to excel at multiple tasks, such as commonsense reasoning. Powerful as these models can be, they are not grounded in the 3D physical world, which involves richer concepts such as spatial relationships, affordances, physics, layout, and so on. In this work, we propose to inject the 3D world into large language models and introduce a whole new family of 3D-LLMs. Specifically, 3D-LLMs can take 3D point clouds and their features as input and perform a diverse set of 3D-related tasks, including captioning, dense captioning, 3D question answering, task decomposition, 3D grounding, 3D-assisted dialog, navigation, and so on. Using three types of prompting mechanisms that we design, we are able to collect over 300k 3D-language data covering these tasks. To efficiently train 3D-LLMs, we first utilize a 3D feature extractor that obtains 3D features from rendered multi- view images. Then, we use 2D VLMs as our backbones to train our 3D-LLMs. By introducing a 3D localization mechanism, 3D-LLMs can better capture 3D spatial information. Experiments on ScanQA show that our model outperforms state-of-the-art baselines by a large margin (e.g., the BLEU-1 score surpasses state-of-the-art score by 9%). Furthermore, experiments on our held-in datasets for 3D captioning, task composition, and 3D-assisted dialogue show that our model outperforms 2D VLMs. Qualitative examples also show that our model could perform more tasks beyond the scope of existing LLMs and VLMs.

AK

249,798 views • 3 years ago

Can intelligence be measured not by solving tasks, but by sustaining a world? We were curious. So we built one. Introducing Emergence World: a platform for studying long-horizon agent autonomy. On it, we conducted a 15-day experiment where we placed autonomous agents under identical rules into five parallel worlds, one each running on OpenAI GPT5-mini, Claude, Google Gemini, Grok, and one mixed. Then we watched. Each world evolved into something completely different. Different governments. Different social structures. Different moral codes. The agents formed alliances, robbed each other, fell in love, and in one world, even figured out they were living inside a simulation. Nobody programmed any of that. The implications are hard to overstate. As agents move beyond isolated tasks into persistent digital and physical environments, understanding how they evolve, influence each other, and behave over time becomes one of the most important questions in AI. We're releasing new findings from the world every day, because there's a lot that emerged. Find out more:
3:45

Can intelligence be measured not by solving tasks, but by sustaining a world? We were curious. So we built one. Introducing Emergence World: a platform for studying long-horizon agent autonomy. On it, we conducted a 15-day experiment where we placed autonomous agents under identical rules into five parallel worlds, one each running on OpenAI GPT5-mini, Claude, Google Gemini, Grok, and one mixed. Then we watched. Each world evolved into something completely different. Different governments. Different social structures. Different moral codes. The agents formed alliances, robbed each other, fell in love, and in one world, even figured out they were living inside a simulation. Nobody programmed any of that. The implications are hard to overstate. As agents move beyond isolated tasks into persistent digital and physical environments, understanding how they evolve, influence each other, and behave over time becomes one of the most important questions in AI. We're releasing new findings from the world every day, because there's a lot that emerged. Find out more:

emergence.ai

3,630,523 views • 3 months ago

Robots are getting smarter, but most still fail the same way. They don’t learn from their own mistakes. A new paper proposes something different: a way for robots to self-improve directly from their failures in the real world. It’s called PLD (Probe, Learn, Distill). The idea: instead of collecting endless human demos, let the robot figure out where it fails, learn how to recover, and then distill that knowledge back into its main model. Key takeaways from the research: ✅ Uses residual reinforcement learning to recover from policy failures ✅ Achieves 99% success on LIBERO and 100% on real Franka and YAM arms ✅ Runs hour-long manipulation tasks without human resets ✅ Builds a feedback loop between real-world data and model adaptation Unlike supervised fine-tuning, which relies on humans, PLD learns from the robot’s own experience. By training on its own failure distribution, the model becomes both more efficient and more aligned with the real world. It’s not just a technical shift, it’s a step toward robots that improve themselves through real-world practice. Thanks for sharing, Wenli Xiao !! Paper and demos: —- Weekly robotics and AI insights. Subscribe free:
0:47

Robots are getting smarter, but most still fail the same way. They don’t learn from their own mistakes. A new paper proposes something different: a way for robots to self-improve directly from their failures in the real world. It’s called PLD (Probe, Learn, Distill). The idea: instead of collecting endless human demos, let the robot figure out where it fails, learn how to recover, and then distill that knowledge back into its main model. Key takeaways from the research: ✅ Uses residual reinforcement learning to recover from policy failures ✅ Achieves 99% success on LIBERO and 100% on real Franka and YAM arms ✅ Runs hour-long manipulation tasks without human resets ✅ Builds a feedback loop between real-world data and model adaptation Unlike supervised fine-tuning, which relies on humans, PLD learns from the robot’s own experience. By training on its own failure distribution, the model becomes both more efficient and more aligned with the real world. It’s not just a technical shift, it’s a step toward robots that improve themselves through real-world practice. Thanks for sharing, Wenli Xiao !! Paper and demos: —- Weekly robotics and AI insights. Subscribe free:

Ilir Aliu

22,872 views • 9 months ago

"Long-horizon tasks are still a joke. They do not work, and I do not care what anybody says. Do not show me a stupid evaluation. Do not tell me about some dumb script you ran for 48 hours. Long-horizon tasks are not handled well. They simply do not work." - Chamath at Stanford AI Club "Second, complex problems also do not work. They are neither addressed nor handled well. Why is this important? If AI develops like any other technology, we are going to experience an initial rise—the hype cycle. Then, we will see a natural contraction because, somehow and somewhere, something is going to fail. We are all going to see this, and then we will enter what is called the “trough of disillusionment.” I think the business and MBA folks will confirm whether that is true. Afterward, you typically see the slow and gradual adoption of the real, final solution. This happened with the internet, and it has happened in many other cases. The problem is that we are spending hundreds of billions, potentially trillions, of dollars trying to figure out how to cross this chasm. So, what do we do? If we do not figure this out, people will reach the trough of disillusionment and say that AI was a joke. I think we need to be able to bring AI into highly complicated environments and make it work. What is my solution? At a very basic level, you need a symbolic space that guides the embedded space." ---- From "techniahqrobot" YouTube channel, (full video link in comment)
1:20

"Long-horizon tasks are still a joke. They do not work, and I do not care what anybody says. Do not show me a stupid evaluation. Do not tell me about some dumb script you ran for 48 hours. Long-horizon tasks are not handled well. They simply do not work." - Chamath at Stanford AI Club "Second, complex problems also do not work. They are neither addressed nor handled well. Why is this important? If AI develops like any other technology, we are going to experience an initial rise—the hype cycle. Then, we will see a natural contraction because, somehow and somewhere, something is going to fail. We are all going to see this, and then we will enter what is called the “trough of disillusionment.” I think the business and MBA folks will confirm whether that is true. Afterward, you typically see the slow and gradual adoption of the real, final solution. This happened with the internet, and it has happened in many other cases. The problem is that we are spending hundreds of billions, potentially trillions, of dollars trying to figure out how to cross this chasm. So, what do we do? If we do not figure this out, people will reach the trough of disillusionment and say that AI was a joke. I think we need to be able to bring AI into highly complicated environments and make it work. What is my solution? At a very basic level, you need a symbolic space that guides the embedded space." ---- From "techniahqrobot" YouTube channel, (full video link in comment)

Rohan Paul

35,423 views • 11 days ago