takelab.bsky.social

takelab.bsky.social

@takelab.bsky.social

10 followers7 following7 posts

View full thread
takelab.bsky.social
@takelab.bsky.social

Old Habits Die Hard: How Conversational History Geometrically Traps LLMs πŸ“œ arxiv.org/abs/2603.03308 πŸ—“οΈ Wed, Jul 8, Poster Session 3 πŸ—“οΈ Fri, Jul 10, Workshop on Mechanistic Interpretability @mtutek.bsky.social

arxiv.orgOld Habits Die Hard: How Conversational History Geometrically Traps LLMsHow does the conversational past of large language models (LLMs) influence their future performance? Recent work suggests that LLMs are affected by their conversational history in unexpected ways. For...
takelab.bsky.social
@takelab.bsky.social

At EACL '26, TakeLab members are presenting two papers. Feel free to stop by if you'd like to chat about simple tricks to improve LLM token embeddings, principled evaluation of synthetic data, coffee, or life in general. Details below πŸ‘‡πŸ§΅ @eaclmeeting.bsky.social #EACL2026 #NLProc πŸ‡²πŸ‡¦

takelab.bsky.social
@takelab.bsky.social

[WASSA] A Multi-Aspect Evaluation Framework for Synthetic Data: Case Study on Irony and Sarcasm πŸ“Šβš–οΈπŸˆ πŸ“œ aclanthology.org/2026.wassa-1... πŸ“ Poster Hall πŸ—“οΈ Sunday, March 29th, 16:30-17:30

aclanthology.orgA Multi-Aspect Evaluation Framework for Synthetic Data: Case Study on Irony and SarcasmLaura Majer, Ana BariΔ‡, Florijan Sandalj, Ivan UnkoviΔ‡, Bojan Puvača, Jan Ε najder. The Proceedings for the 15th Workshop on Computational Approaches to Subjectivity, Sentiment Social Media Analysis (W...

takelab.bsky.social reposted

Martin Tutek
@mtutek.bsky.social

ManagerBench was accepted to ICLR! @iclr-conf.bsky.social #ICLR2026 LLMs are still either unsafe, or completely harm avoidant - even when the harm affects furniture πŸ›‹οΈ Check out our benchmark, online or in Rio πŸ‡§πŸ‡·

Martin Tutek@mtutek.bsky.social

πŸ€”What happens when LLM agents choose between achieving their goals and avoiding harm to humans in realistic management scenarios? Are LLMs pragmatic or prefer to avoid human harm? πŸš€ New paper out: ManagerBench: Evaluating the Safety-Pragmatism Trade-off in Autonomous LLMsπŸš€πŸ§΅

takelab.bsky.social reposted

camuljak.bsky.social
@camuljak.bsky.social

We already know prompt repetition is a handy hack to improve a decoder-only LM’s performance as it allows the model to β€œsee” bidirectionally, an ability otherwise suppressed by the causal mask. But what happens if we increase the number of repetitions? πŸ€”πŸ§΅ @eaclmeeting.bsky.social #EACL2026

takelab.bsky.social reposted

Akhil Arora
@akhilarora.bsky.social

Back from #ICML2025 πŸ›¬, and off πŸš„ to NorrkΓΆping πŸ‡ΈπŸ‡ͺ for #ic2s2 CLAN (cs.au.dk/~clan/) members are presenting 2 papers: 1 spotlight and 1 oral. See 🧡for posters and summaries πŸ‘‹ Reach out to chat about observational studies, causality, LLM agents, human-centered AI etc.

cs.au.dkCLAN for AI Research on Language and NetworksCLAN for AI Research on Language and Networks, or **CLAN** for short, is a research group and a part of the larger [DIS Group](https://cs.au.dk/research/data-intensive-systems) in the [Department of C...

You are up to date.