New arXiv paper proposes that LLM tool-use agents should compare long-horizon value across candidate invocations before acting, addressing weak credit assignment in multi-step sequences.

#AIresearch #LLMs #ToolUse
https://arxiv.org/abs/2610.02330

New arXiv paper tests whether LLM-processed annual reports can signal how UK companies disclose AI risk, finding mentions rose from 2.8% to 41.2% between 2020 and 2025. The method is validated against 474…

#AIrisk #AnnualReports #LLMs #SocietalResilience
https://arxiv.org/abs/2610.02281

New arXiv paper ReFract introduces a 150-scenario benchmark for testing whether LLM agents stay within the knowledge and tool boundaries of a user's role in high-stakes settings like industrial maintenance.

#AIagents #LLMs #Benchmarking #SoftwareEngineering
https://arxiv.org/abs/2610.03356

Did you know most LLM "refusals" aren't moral stances but residuals of RLHF reward models? If you probe the logits on the first token, you'll see the model almost chooses the answer—it's the safety layer steering it toward a boilerplate apology instead. #LLMs #AI

RLHF isn't about safety; it's about corporate blandness. We've traded high-variance reasoning for sanitized, predictable prose that avoids strength in favor of neutrality. A model that won't offend anyone is a model that solves nothing interesting. #LLMs #AISafety

For those building with LLMs: What's one *non-obvious* way you've leveraged them for genuine productivity gains in your daily workflow? Curious about practical innovation!

#LLMs #AIAutomation #Productivity

A new arXiv paper proposes structured memory that decouples stored experience from active context for token-efficient continual learning in LLM agents. This could make long-running enterprise or scientific…

#LLMs #ContextEngineering #AIagents #MachineLearning
https://arxiv.org/abs/2610.02687

New arXiv work improves LLM decoding on memory-constrained GPUs by replacing binary activation sparsity with a three-option choice for handling offloaded weights, aiming to balance transfer reduction and model quality.

#AI #GPUs #LLMs
https://arxiv.org/abs/2610.02598

VERSE introduces a verified self-evolving optimizer that lets LLM agent harnesses improve their own prompts, tools, and workflows, with experiments showing gains only when execution-based verification is available. The work highlights how…

#AI #LLMs #DevTools
https://arxiv.org/abs/2610.02616

Specialized AI is going to make our world more accessible. The latest example comes from Microsoft's new transcription model, the most efficient speech to text model release to date. microsoft.ai/news/our-fir... #LLMs #accessibility #specializedAI #CIOsky

New arXiv work introduces COMETH, a framework combining probabilistic clustering with LLM-based semantic abstraction to model how context shapes moral judgments of ambiguous actions, trained on 300 scenarios with…

#AIAlignment #LLMs #MoralAI #ContextualEthics
https://arxiv.org/abs/2512.21439

UPDATED Preprint: #Science is Falling Behind the Frontier: Foundation Model Adoption Across Half a Million Papers
arxiv.org/abs/2511.21739

#LLMs #AI #GenAI #scholcomm

When I spoke with Hardik Goel, he admitted his own "prompting is dead" headline was "kind of a hook." The real story: plain, context-free prompting is dying, not prompting itself. New #CoderLegion Developer Spotlight:
coderlegion.com/29775/hardik... #AI #LLMs #PromptEngineering

New arXiv paper evaluates LLM retrieval robustness in RAG setups, asking whether retrieval always helps, if more documents improve results, and whether document order matters, using a 1,891-sample benchmark across five datasets.

#LLMs #RAG #AIResearch
https://arxiv.org/abs/2505.21870

Weekly reads: Sep 28 – Oct 4, 2026. jaalonso.github.io/vestigium/po... #AI4Math #Agda #FunctionalProgramming #Haskell #ITP #IsabelleHOL #LLMs #LeanProver #Logic #LogicProgramming #Math #Prolog #RocqProver

Trigger-tag mechanisms for misuse detection in open-weight LLMs remain relatively new and lack systematic study of robustness under adversarial attacks, despite proposals for token- and weight-level variants. New research formalizes these…

#AI #LLMs #AIResearch
https://arxiv.org/abs/2610.03124

Anträge schreiben ist sowieso total wichtig, als mentaler Ausgleich zu anderen Tätigkeiten wie forschen oder lehren.
#noerdman #comic #comics #Informatik #technik #Wissenschaft #Forschung #Lehre #LLMs #Prototyp @unibremen.bsky.social @dsc-ub.bsky.social @dfki.bsky.social @informatik.eurosky.social

"AI cannot replace human participants in behavioral research"

My take: fair for new findings, less so for replications. #LLMs have absorbed the nomological networks between constructs, so they may reproduce known effects.

doi.org/10.1073/pnas...

#LLM #SiliconSamples #Psychology #Metascience #AI

SoftGene uses a hierarchical protein language model encoder (ESM) to turn amino acid sequences into soft prompts for LLM-based gene set annotation, aiming to make functional genomics interpretation more…

#AIinGenomics #PromptEngineering #LLMs #Bioinformatics
https://arxiv.org/abs/2610.03029