Research Papers

Nemotron-Terminal (arXiv:2602.21193, 66 HF upvotes) — First systematic study of data engineering for terminal/CLI agents. Terminal-Task-Gen pipeline with Dockerized environment interaction. Qwen3-initialized 8B model goes from 2.5% to 13.0% on Terminal-Bench 2.0. All checkpoints and synthetic datasets open-sourced at nvidia/nemotron-terminal. The open-source recipe for building your own Claude Code-like terminal agent. (arXiv)

PyVision-RL (arXiv:2602.20739, 43 HF upvotes) — Addresses "interaction collapse" where RL-trained agents learn to reduce tool usage and multi-turn reasoning. Oversampling-filtering-ranking rollout strategy sustains interaction. If you're training agents via RL, this solves a real problem. (arXiv)

TAPE: Tool-Guided Adaptive Planning (arXiv:2602.19633) — Graph-based multi-plan aggregation + constrained decoding + adaptive re-planning on environmental deviation. For agents in costly real-world environments (deployments, DB operations, financial transactions). (arXiv)

TTT with KV Binding = Linear Attention (arXiv:2602.21204) — NVIDIA/UofT/Vector Institute reveal test-time training with KV binding is mathematically equivalent to learned linear attention. Enables principled simplifications and fully parallel formulations. Replace complex TTT with simpler linear attention for same results. (arXiv)

DREAM: Deep Research Evaluation (arXiv:2602.18940) — AWS framework for evaluating deep research agents with metrics for multi-turn reasoning, tool use patterns, and research quality. If building Perplexity-style research agents, this provides standardized measurement. (arXiv)