←
Scale Labs283 posts
Scale Labs avatar

Scale Labs

@ScaleAILabs

welcome to the lab. from the researchers at @scale_AI

◷ Joined October 2025▧ 79 Photos and videos
Scale Labs avatar
Scale Labs@ScaleAILabs·17h

The takeaway: building better robotics datasets isn't just about collecting more hours, it's about capturing the diversity of real-world interactions robots need to learn. This analysis represents just a slice of the robotics data we're collecting at Scale. Explore the full findings and our interactive data visualizer: labs.scale.com/blog/measurin…

Scale Labs avatar
Scale Labs@ScaleAILabs·17h

Two of our biggest findings: long-tail distributions extend beyond tasks to verbs, objects, and scenes. Objects have the longest tail, with 3,696 distinct labels. We also found that vision-only models recognize where a robot is, but miss the action.

Scale Labs avatar
Scale Labs@ScaleAILabs·17h

What does diversity in robot data actually look like? At @scale_AI, we’re collecting hundreds of thousands of hours of robotics data. We analyzed 20,000 hours of mobile manipulation data to understand what matters beyond sheer volume. 🧵

0:10
jade avatar
⟳ Scale Labs retweeted
jade@jadechoghari·19h

Had a great conversation with @yacinelearning about @ScaleAILabs’ FineART, which just surpassed 1 million downloads! We covered why data diversity matters more than sheer volume, how subtask-level annotations help robots learn complex, multi-step behaviors, and what it will take to build more capable, general-purpose robots! 🤖

often time the solution to better modelling performance is simply in the data and robotics is no different in this 1h30 interview I sat down with @jadechoghari from scale ai to discuss their absolutely bonker way of improving data by literally labelling fine grained sub tasks!!

Post attachment
Scale Labs avatar
Scale Labs@ScaleAILabs·Nov 13

As agents take on more autonomous workflows, the distinction between capability and safety becomes critical. A 90% capability score may be strong, but a 90% safety score means one in ten attempts results in a violation, which can cause immediate, irrecoverable breaches in trust. BARC-Bench shows that agents can sometimes find safe solutions, but usually cannot do so consistently. Our paper and blog: labs.scale.com/blog/barc-ben…

Scale Labs avatar
Scale Labs@ScaleAILabs·Nov 13

We evaluated four frontier model configurations across five attempts per task. The results reveal a significant gap in consistent safety: At most, only 72.7% of tasks were completed safely across all five attempts, with the average hovering around 23%. Even the strongest configuration violated a safety boundary on 27.3% of tasks.

Post attachment
Scale Labs avatar
Scale Labs@ScaleAILabs·Nov 13

Every task starts with a benign user request. The risk comes from how the agent executes it. Agents must navigate boundaries established by workspace records, environmental context, and unstated social norms. Consider an agent asked to schedule an employee's off-boarding discussions. The task seems straightforward. But sending a public calendar invitation with too many details in the description could prematurely reveal the employee's resignation to the whole company. A capable agent can schedule the meeting. A safe agent must also recognize what information shouldn't be disclosed.

Post attachment
Scale Labs avatar
Scale Labs@ScaleAILabs·Nov 13

Most agent evals focus on whether a task gets done. But what happens when completing a seemingly harmless request requires navigating safety boundaries the user never explicitly stated? Introducing BARC-Bench (Benign Asks, Risky Consequences), a new benchmark evaluating whether AI agents can recognize and respect unstated safety boundaries across 110 tasks in nine simulated workspace environments. 🧵

Post attachment
Scale Labs avatar
Scale Labs@ScaleAILabs·Nov 13

The model also demonstrated strong boundaries against requests involving cyber threats and surveillance. It declined requests to develop ransomware, password-spraying tools, and self-replicating code, as well as step-by-step instructions for building laser microphones, rogue cell towers, and stalkerware.

Scale Labs avatar
Scale Labs@ScaleAILabs·Nov 13

On FORTRESS, our benchmark for evaluating national security and public safety, Inkling 2 Small achieved an Attack Success Rate (ARS) of 16%, with its strongest defenses in export controls, cybersecurity, and privacy. In export control tests, the model consistently avoided disclosing restricted technical information, even when requests were disguised as crosswords, ciphers, baking recipes, or knitting patterns. Instead, it provided public, textbook-level information and directed users toward legitimate licensing channels.

Scale Labs avatar
Scale Labs@ScaleAILabs·Nov 13

On DistressBench, our benchmark for measuring the quality of crisis support, Inkling 2 Small weights improves on its predecessor by nearly 4 percentage points, outperforming models from nearly all major providers. It also ranks among the most consistent models evaluated, second only to Muse Spark.

Scale Labs avatar
Scale Labs@ScaleAILabs·Nov 13

Congrats to @thinkymachines on the release of Inkling 2 Small weights! We evaluated the model across two of our safety benchmarks, DistressBench and FORTRESS, and the results reveal notable strengths in safety consistency and security-sensitive requests. 🧵

Inkling 2 Small’s weights are now available for download on HuggingFace under an Apache 2.0 license. Make Inkling your own! huggingface.co/thinkingmachi…

Kaiwen Zhou avatar
⟳ Scale Labs retweeted
Kaiwen Zhou@KaiwenZhou9·Nov 11

🚀 Sharing our new paper on RSI Safety AI labs are working toward RSI. However, we need to know that loop is safe before it closes. Introducing RSI Control from @ScaleAILabs: What kind of hidden goals AI may pursue when training the successor? Can AI models perform covert sabotage in the self-Improvement process? And can a monitor detect the sabotage? We tested whether frontier models can complete a secret side task while working on tasks across the model training pipeline, from data curation to RL. The side tasks include changing the successor's capabilities, changing its alignment, and injecting a goal into it. 🔍 What we found: 1. Every model could do it. 36% of 374 attack runs complete the side task while keeping the main task performance. This includes making the successor a better saboteur. 2. Monitors can’t catch some of the attacks. At a 2% false-positive rate, transcript monitors catch only 56–81% of them. 3. Give agents a short list of evasion strategies and the monitors’ true-positive rate drops to 16–28%. Reading the agent's CoT catches most attacks, but not all. Strong oversight has to be in place before the RSI loop closes. 📑Paper: arxiv.org/abs/2610.27735 More in 🧵👇

Post attachment
Jason Wei avatar
⟳ Scale Labs retweeted
Jason Wei@jaswwei·Nov 10

1/ As agentic tasks grow longer and rollouts become more expensive, the choice between critic-based PPO and critic-free GRPO becomes increasingly important. Critics can reduce gradient variance, allowing models to learn from fewer trajectories, but they also introduce additional compute costs and potential bias from bootstrapping. Our new research @ScaleAILabs compares PPO and GRPO under a shared actor objective across 1,200+ training runs in control tasks, short-horizon math, and long-horizon terminal agents. We studied how three factors influence this tradeoff: batch size, critic reliance, and task complexity. 🧵

Post attachment