Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies
Paper • 2609.38537 • Published • 28
Dipankar Sarkar PRO
dipankarsarkar
AI & ML interests
Building the AI-native stack. Agents as infrastructure, safety as architecture, performance as plumbing. I publish the receipts: papers, datasets, demos.
Recent Activity
upvoted a paper about 2 hours ago
Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies upvoted a paper about 2 hours ago
EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making upvoted a paper about 2 hours ago
WorldAuditBench: Interactive 3D World Auditing with Multimodal AgentsOrganizations
upvoted 6
papers about 2 hours ago
EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making
Paper • 2609.38334 • Published • 64
WorldAuditBench: Interactive 3D World Auditing with Multimodal Agents
Paper • 2609.40325 • Published • 90
DAGent: Evaluate-then-Grow Planning for Deep Research Agents
Paper • 2609.39154 • Published • 3
CUA-SWE: When Computer-Use Agents Meet Visual Software Engineering
Paper • 2609.32600 • Published • 12
MILO: Automated Harness Discovery via Orchestrated Multi-Agent Evolution
Paper • 2609.38349 • Published • 15
Sure on 100%, right on 66.1% is the number worth keeping. The threshold learned the templates, not the task.
Cheap fix before the card: pick each head's threshold on a holdout split by template, not at random. The same leave-one-template-out you just ran, used for calibration and not only for the eval. Then "sure" has to mean sure on a template it never saw.
urgency held up and category didn't, so the per-head thresholds would likely move by very different amounts. That gap is itself worth a line on the card.
On real traffic there is no template id, but distance to the nearest training ticket plays the same role. It flags a ticket before the head answers, where agreement drift only notices after a batch has gone wrong.
Would you route the far ones straight to the teacher, or keep them inside the sampled share?
liked 2
datasets about 5 hours ago
upvoted 6
papers about 5 hours ago
Rubric Rewards from Item Response Theory
Paper • 2609.35646 • Published • 10
It's Not What the Image Shows: Irrelevant Context Destabilises VLM Judges Without Informing Them
Paper • 2609.37863 • Published • 35
PatchHolmes: Agentic Patch Retrieval via Listwise Selection
Paper • 2609.38807 • Published • 7
CheatBench: Measuring Reward Gaming in AI Agents
Paper • 2609.36308 • Published • 4
See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs
Paper • 2609.34970 • Published • 4
SkillSeek: Revisiting Agent Skill Retrieval at Marketplace Scale
Paper • 2609.38822 • Published • 4
upvoted 5
papers about 6 hours ago
Dating the Model: Hidden Dates in System Prompts Affect LLM Evaluation
Paper • 2609.36931 • Published • 3
WUSH-KV: KV Cache Quantization with Data-Adaptive Transforms
Paper • 2609.38121 • Published • 14
Thinking Outside the Box: Can Language Models Rely on External Guidance Selectively?
Paper • 2609.39578 • Published • 60
PivotOPD: Learning to Recover from Pivotal Mistakes in Multi-Turn Agents
Paper • 2609.40285 • Published • 20
A2Z GameSpec-Bench: How Faithfully Can Coding Agents Generate Games from Game Design Specifications?
Paper • 2609.39564 • Published • 9