@MSFTResearch: LLMs that get lost in conversation. Robots that learn zero-shot. Diffusion models built for imbalanced data. This is th…
Summary
Microsoft Research highlights ICLR papers on LLMs losing conversational coherence, zero-shot robot learning, and diffusion models for imbalanced data.
Similar Articles
@HowToAI_: Microsoft Research + Salesforce has published a paper that should scare every single AI builder right now. It’s called …
A new paper by Microsoft Research and Salesforce reveals that LLM performance drops significantly in multi-turn conversations due to a 'Lost in Conversation' phenomenon, challenging the reliability of current single-turn benchmarks.
@MSFTResearch: Small language models learn to negotiate with SocialRL, PazaBench V2 expands speech AI evaluation across African langua…
Microsoft Research highlights new research on SocialRL for small language model negotiation, PazaBench V2 for African language speech evaluation, EvoLib for agent experience learning, improved A/B testing methods, and AI-driven precision oncology.
@MSFTResearch: New tools, models, repos, and papers out of Microsoft Research are here. Use AI and agents? It's worth watching: • Mage…
Microsoft Research announced new tools, models, repositories, and papers, including MagenticLite, agentic GitHub workflows, verification-first agents, and meaning-matching fine-tuning, during the Microsoft Research Forum virtual series.
@MSFTResearch: Among Microsoft’s 100+ accepted papers, 3 oral presentations, and 1 expo demo at ICML in Seoul, highlights include Fara…
Microsoft Research highlights from ICML 2026 include the Fara 1.5 computer-use agent family, critique-resilient benchmarking, expanded protein ML benchmarks with FLIP2, and improved LLM reasoning stability, with over 100 accepted papers.
@MSFTResearch: LLMs do not get smarter just by remembering more. EvoLib turns experience into evolving knowledge, taking reusable skil…
Microsoft Research introduces EvoLib, a framework that enables LLMs to continually learn from their own experience during inference by extracting reusable skills and insights, without model updates or external labels.