@MaxForAI: Interesting, lately I've been feeling more and more that the RSI direction is being gradually advanced bit by bit by an…

X AI KOLs Timeline Papers
agentic-rl long-horizon-tasks reinforcement-learning verifiable-rewards training-environments rsi

Summary

评论介绍了 UCLA 与 Adobe Research 的 Tev-Gym 工作：通过构建可验证的长任务环境与稠密中间进度奖励来训练长时程 Agent，Tev-27B 在 ALE-CM 计算与数学子集上以 Pass@5 比肩 GPT-6 Astra 与 Claude Opus 5，Tev-320B 则将 GLM-5.3-Flash 在 Terminal-Bench 4.0-CM 上的 Pass@1 从 29% 提升至 37%。

Interesting, lately I've been feeling more and more that the RSI direction is being gradually advanced bit by bit by an increasing number of practical studies. Just saw a really intriguing piece of work called Tev from UCLA and Adobe Research, specifically focused on how to scale up training for long-task Agents. They directly trained a 27B model that achieved a Pass@5 of 28% on the computational and mathematical science tasks in Agents'LastExam, matching GPT-6Astra and Claude Opus5??? Of course, this is just the result on a specific subset, but considering the model is only 27B, it's still pretty noteworthy. And I think their approach is even more worth paying attention to than the results themselves. Specifically: Right now, many Agents can already work continuously for a few hours, or even longer, but figuring out how to train these kinds of Agents through reinforcement learning is actually a really thorny problem. First off, it's hard to construct these complex tasks at scale. Second, even if you do manage to create them, the model still struggles to learn from them. Because a single task might require dozens of consecutive steps, and if it fails at the end, traditional outcome-based rewards often just give it a zero. Even if the model got the first 90% right, a problem on the very last step means the whole process might not yield any effective learning signal. Tev's idea is to first construct a perfectly correct task world, complete with full workflows, right answers, and the expected states at each stage. Then, they reverse-engineer a flawed environment from it, letting the Agent fix it on its own (this part is absolutely crucial). This way, the training system knows exactly how much the Agent got right at each step, and even if the task isn't ultimately completed, it can still provide rewards based on intermediate progress. Using this method, they built 1170 verifiable long-task environments, covering 7 domains like software engineering, data engineering, AI research, mathematics, security, and infrastructure. Even more interesting is that this whole approach can transfer to models of different scales. They took GLM-5.3-Flash and continued training it to get Tev-320B, boosting Pass@1 from 29% to 37% on the computational and math subset of Terminal-Bench4.0. I think this really highlights some key points. We've been talking nonstop about how to make Agents keep getting stronger, how to make capability improvements rely less and less on human involvement. But behind that, there are a ton of very specific engineering and research problems that need solving. Where does the training data come from, how to automatically generate complex tasks, how to verify intermediate states, how to turn failure trajectories into effective learning signals. Tev is hitting on several of those really critical ones. And it's already shown through actual training that it can boost long-task capabilities across models of different sizes. A lot of RSI's foundational capabilities are built up just like this, step by step.
Original Article
View Cached Full Text

Cached at: 10/15/26, 02:14 PM

Interesting, lately I’ve been feeling more and more that the RSI direction is being gradually advanced bit by bit by an increasing number of practical studies.

Just saw a really intriguing piece of work called Tev from UCLA and Adobe Research, specifically focused on how to scale up training for long-task Agents.

They directly trained a 27B model that achieved a Pass@5 of 28% on the computational and mathematical science tasks in Agents’LastExam, matching GPT-6Astra and Claude Opus5???

Of course, this is just the result on a specific subset, but considering the model is only 27B, it’s still pretty noteworthy.

And I think their approach is even more worth paying attention to than the results themselves.

Specifically: Right now, many Agents can already work continuously for a few hours, or even longer, but figuring out how to train these kinds of Agents through reinforcement learning is actually a really thorny problem.

First off, it’s hard to construct these complex tasks at scale. Second, even if you do manage to create them, the model still struggles to learn from them. Because a single task might require dozens of consecutive steps, and if it fails at the end, traditional outcome-based rewards often just give it a zero. Even if the model got the first 90% right, a problem on the very last step means the whole process might not yield any effective learning signal.

Tev’s idea is to first construct a perfectly correct task world, complete with full workflows, right answers, and the expected states at each stage.

Then, they reverse-engineer a flawed environment from it, letting the Agent fix it on its own (this part is absolutely crucial).

This way, the training system knows exactly how much the Agent got right at each step, and even if the task isn’t ultimately completed, it can still provide rewards based on intermediate progress.

Using this method, they built 1170 verifiable long-task environments, covering 7 domains like software engineering, data engineering, AI research, mathematics, security, and infrastructure.

Even more interesting is that this whole approach can transfer to models of different scales.

They took GLM-5.3-Flash and continued training it to get Tev-320B, boosting Pass@1 from 29% to 37% on the computational and math subset of Terminal-Bench4.0.

I think this really highlights some key points.

We’ve been talking nonstop about how to make Agents keep getting stronger, how to make capability improvements rely less and less on human involvement.

But behind that, there are a ton of very specific engineering and research problems that need solving.

Where does the training data come from, how to automatically generate complex tasks, how to verify intermediate states, how to turn failure trajectories into effective learning signals.

Tev is hitting on several of those really critical ones.

And it’s already shown through actual training that it can boost long-task capabilities across models of different sizes.

A lot of RSI’s foundational capabilities are built up just like this, step by step.

Wenbo Hu (@gordonhu608): How to scale long-horizon agentic RL Envs? 🤔

We introduce Tev-Gym: 1,170 envs with dense progress rewards. Trained on it:

🔥 Tev-27B matches GPT-6 Astra & Claude Opus 5 on ALE-CM (computing and math domains) at Pass@5. 🚀 Tev-320B tops open models on Terminal-Bench 4.0-CM at

Similar Articles

Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards

Hugging Face Daily Papers

本文提出Tev-Gym——首个具有可验证、密集渐进奖励的长周期智能体训练数据集（7个领域1170个任务），并基于进度-势能奖励塑形（PPRS）训练出Tev-27B，在Agents' Last Exam计算与数学科学领域匹配GPT-6 Astra和Claude Opus 5的Pass@5表现，并可泛化至Terminal-Bench 3.0/4.0。

@cuisitekp: A 9B model outperforms models several times larger. The team behind OLMo/Tülu from Ai2 and the University of Washington released a new paper called Tmax, claiming it's the strongest open-source RL training recipe for 'terminal agents'. Result: A 9B model on Terminal-Be…

X AI KOLs Timeline

Ai2 and the University of Washington released a paper titled Tmax, proposing the strongest open-source terminal agent RL training recipe to date. A 9B parameter model outperforms larger models on Terminal-Bench 2.0, with the key being low-cost generation of vast amounts of verifiable training data, not model size or algorithm.