Research on LLM agents at IBM and the Weizmann Institute of Science:
what agents pursue without being asked, and how safely they act.
Ido Levy, Segev Shlomov, Leshem Choshen
RAG techniques are usually tested one at a time. Do their gains survive when they are combined? Across 450 configurations, eight benchmarks and corpora from 5K to 200K documents, the three strongest single retrieval techniques predict a +47% gain in NDCG@10 when combined and deliver +17%, and the pipeline that leads on small corpora trails by 13.7% NDCG@10 at 200K documents. Lasagna, the open-source framework behind the study, swaps one stage of a RAG pipeline while holding the rest fixed and tests whether the difference is real.
Paper · Project page · Code · PyPI · Pre-registration
Ido Levy, Segev Shlomov, Avi Yaeli, Nir Mashkif, Asaf Adi, Yinon Goldstein, Eilam Shapira, Leshem Choshen
Agents call a frontier LLM again and again for closed-set decisions: which tools to shortlist, which skill or application to use, where to route. TabAgent replaces that call with a compact classifier trained on the agent's own successful runs, leaving the rest of the agent intact. On tool shortlisting it recovers essentially all of the tools the GPT-5.4 shortlister used on its successful runs, at about 10³× lower latency and 10⁵× lower cost per call.
Paper · Project page · Code · Models 🤗
Ido Levy, Asaf Yehudai, Segev Shlomov, Asaf Adi, Leshem Choshen
What should an LLM agent pursue that the user never asked for? Need graphs measure it without a model judge, and Q&D trains it from the consequences of its own questions. The trained 8B questioner recovers 90% of the required evidence where the same model, prompted, recovers 78%, and it more than doubles retail task success in a τ²-bench customer-service agent it was never trained on.
Paper · Project page · Code · Model 🤗
Ido Levy, Ben Wiesel, Sami Marreed, Alon Oved, Avi Yaeli, Segev Shlomov
A benchmark for evaluating the safety and trustworthiness of web agents in enterprise scenarios.
Paper · Website · Code · Leaderboard · Dataset
An agent harness for complex tasks on the web and APIs, with OpenAPI and MCP integrations, a composable architecture, reasoning modes and policy-aware features.