🚀 Today we’re introducing 𝗔𝘂𝘁𝗼𝗜𝗻𝗱𝗲𝘅: Learning Representation Programs for Retrieval. 🚀 AutoIndex learns executable programs that determine how a corpus is represented to a search engine. We think this points to a broader category of research: 𝗔𝗜 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝘁𝗵𝗮𝘁 𝗼𝗽𝘁𝗶𝗺𝗶𝘇𝗲 𝘁𝗵𝗲 𝗲𝘅𝗲𝗰𝘂𝘁𝗮𝗯𝗹𝗲 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝘀 𝗮𝗿𝗼𝘂𝗻𝗱 𝘁𝗵𝗲𝗺.

I was really surprised how well this worked out of the box, and we're only getting started. So much untapped potential! All of the impressive results and architecture designs should be credited to the UMass MS (!) students here.

Introducing Harbor-Index, a compact, diverse, and high-quality benchmark built to challenge frontier agents. We carefully select, audit and fix 82 high-signal tasks out of 6,627 candidates spanning 54 benchmarks. No agent gets above 30%. (1/5)

Thanks for the heroic work! @zhuofengli96475 @GhxIsaac @chuanli11 @yuz9yuz @ShuiwangJi @DongfuJiang @YuyangBai02 If you're attending ACL, we'd love to see you at our poster session! 🗓 Tuesday, July 7 🕚 11:00 AM–12:30 PM 📍 Grand Hall #acl# #llm# #agenticAI# #AIforResearch# #nlp# @aclmeeting

Came across DR-DCI via @jobergum's talk at @aiDotEngineer. It's tackling a problem that comes up constantly once you're building agents: how does an agent actually search over a large and distributed set of documents. Traditional RAG has known limits. It chunks documents, embeds them, and hands the agent a handful of disconnected snippets. If the answer got split across a chunk boundary, or needs cross-referencing two documents, the agent has no way to go look further, it only gets the one retrieval pass. The other extreme, the current "grep is all you need" trend, just gives the agent shell tools and lets it search the raw corpus directly. That's more flexible, but it doesn't hold up at scale, full-corpus grep gets slow and starts timing out as the collection grows across distributed physical nodes. DR-DCI (paper by @zhuofengli96475 and team) is a hybrid of the two. BM25 runs first and narrows a huge corpus down to a manageable set of candidate documents. Those relevant docs get pulled into a bounded, sandboxed workspace, and only then does the agent run real grep/cat/find against that smaller set instead of the whole corpus. 71% accuracy vs 63% for raw grep-everything, and about 20x faster. I made some observations of my own on top of the paper, mainly around productionizing this for distributed systems. These include shard locality, tiered caching, caching differences for immutable vs mutable content, fast sandbox setup/teardown for the VFS workspace, and precomputing common queries. This area (RAG + grep hybrids) feels like a real harness engineering lever for inference providers to differentiate on, both on tool-calling benchmarks and in how agents actually perform on real work.

Two years ago, we built OSWorld 1.0 — the benchmark that became the standard for computer-use agents. Agents now score 83.5% on it. Problem solved? Not even close. 🚀Today we introduce OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks. What's new: 🎯 108 real-world workflows, each ~1.6 hours ⏱️ for a skilled human ⚙️ ~318 tool calls/task vs. ~30 in OSWorld 1.0 🌍 Grounded in authentic artifacts & stateful user profiles ⚡ Captures real phenomena: dynamic environments, streaming interaction, cross-source reasoning, implicit-state inference & more 📊 Best results: Claude Opus 4.8 reaches the highest accuracy at 20.6%, while GPT-5.5 is far more token-efficient but plateaus near 13%. No one is close to solving real computer use. 🏠 Homepage: 📄 Paper: 💻 Code: 🤗 Dataset: 🧵 [1/8]

Our work on multi-agent teamwork is accepted to #ICML2026#! When agents operate beyond fixed workflows, teamwork is critical. We show that frontier LLMs often make poor teammates—not because they're unhelpful, but because they're too accommodating. Great job by @aneeshpappu w/ awesome collaborators!

🚀 Impressive work. Training a strong Deep Research agent with just 32 H100s, 8K synthetic samples, and a fully open recipe highlights how fast open ecosystems are advancing. Great to see OpenResearcher ( included as a baseline. Exciting progress for open Deep Research. 👏

🚀Excited to announce QUEST today! Benchmark numbers will fade, but the know-how behind them endures. With QUEST: See how we synthesize diverse Deep Research tasks through a unified framework at scale. See how we train Deep Research agents through three stages: Mid-Training → SFT → RL. See how our caching system dramatically reduces API costs during training. See how our context management enables unbounded deep research. We’ve released everything we can!

🚀 Impressive work. Training a strong Deep Research agent with just 32 H100s, 8K synthetic samples, and a fully open recipe highlights how fast open ecosystems are advancing. Great to see OpenResearcher ( included as a baseline. Exciting progress for open Deep Research. 👏

🚀 DR-DCI improves raw DCI from 62.9% → 71.2% on BrowseComp-Plus while also reducing 💰 cost and ⏱️ runtime. 🔍 Retrieval-for-discovery + 🛠️ DCI-for-investigation is a practical decomposition for scaling DCI agents to large-scale corpora (10M). Congrats to the authors! 👏

Excited to share the 𝗗𝗿-𝗗𝗖𝗜 project led by Yi Lu, @zhuofengli96475, and @DongfuJiang! Dr-DCI scales DCI (our prev work): retrieval 𝗽𝘂𝗹𝗹𝘀 𝗿𝗲𝗹𝗲𝘃𝗮𝗻𝘁 𝗱𝗼𝗰𝘀 𝗶𝗻𝘁𝗼 𝗮 𝘄𝗼𝗿𝗸𝘀𝗽𝗮𝗰𝗲; agents then use terminal tools there for evidence search and verification.