RT @EliasEskin: Extremely excited to announce that I will be joining @UTAustin @UTCompSci in August 2025 as an Assistant Professor! 🎉. I’m….

RT @willcfleshman: 🚨 Our latest paper is now on ArXiv! 👻.(w/ @ben_vandurme). SpectR: Dynamically Composing LM Experts with Spectral Routing….

RT @alexdmartin314: Check out our latest work, Video-ColBERT, an efficient late interaction approach for video retrieval! Recently accepted….

RT @Yining__Lu: 📣 New Preprint 📣 Did you realize there is a hidden misalignment between decomposer and verifier in long-form text factualit….

RT @ben_vandurme: Our latest on compressed representations: Key-Value Distillation (KVD). Query-independen transformer compression, with of….

7/ Huge thanks to my advisors @anqi_liu33 and @ben_vandurme!.

6/ The takeaway? LLMs don’t need to hallucinate or stay silent—they can express uncertainty by adjusting specificity, striking a better balance between informativeness and reliability. Full paper here:

5/ More importantly, models can be tuned to adjust their specificity based on confidence levels—and this specificity is adaptive! At the same confidence level, the model provides more detailed responses for common knowledge and hedges more when dealing with obscure topics.

4/ The result? By being less specific when uncertain, an 8B model can achieve even higher accuracy than GPT-4o on SimpleQA. Even when the base model performs acceptably well, our method significantly improves accuracy.

3/ Therefore, we can explicitly control the tradeoff between specificity and accuracy using conformal prediction—ensuring responses remain as factual as possible without unnecessary vagueness.

2/ We propose a unifying view of existing uncertainty-based hallucination mitigating approaches: an LLM is always committing to a set of possible worlds that align with its internal beliefs.

1/ 🚨LLMs can still be factual even when they don’t know the full answer!🚨 Introducing Conformal Linguistic Calibration—a new way for models to express uncertainty by responding less specifically until they are confidently factual.🧵

RT @orionweller: Ever wonder how test-time compute would do in retrieval? 🤔. introducing ✨rank1✨. rank1 is distilled from R1 & designed for….

RT @jeff_cheng_77: ‼️Tired of dealing with long reasoning chains?‼️. Introducing Compressed Chain of Thought (CCoT), a framework to perform….

RT @MASC_Conference: 📢 Want to host MASC 2025? . The 12th Mid-Atlantic Student Colloquium is a one day event bringing together students, fa….

RT @jackjingyuzhang: 🤖 LLMs are powerful, but their "one-size-fits-all" safety alignment limits flexibility. Safety standards vary across c….

RT @kesnet50: 🤖 New Multimodal Task! 🤖. We tackle the problem of piecing complex stories together through short snippets of video and text,….

RT @Dongwei__Jiang: Process supervision for reasoning is 🔥! While previous approaches often relied on human annotation and struggled to gen….

RT @orionweller: Introducing ✨Promptriever ✨ the first retriever that can be prompted like an LM with free-form prompts!. Our secret: quer….