REBEL: Hidden Knowledge Recovery via Evolutionary-Based Evaluation Loop
Does unlearning really work? REBEL evolves adversarial prompts that recover "forgotten" facts in up to 60% of cases (TOFU) and 93% (WMDP).
https://t.co/08w9lABmBH
Three papers from Sarmata group at HHU Düsseldorf accepted to EACL 2027 (Findings), joint work with @PawelBatorski and @przem_spurek. Infos on papers below.
GPS: General Per-Sample Prompter
GPS is an automatic prompt optimizer giving sample-adaptive prompts without access to task-specific training sets. The prompt optimizer is pertained with RL on large corpora of problems and generalizes to new problems.
https://t.co/KAvgK5GgWP
Our paper got into #NeurIPS2026! 🎉
"Transcoda: End-to-End Zero-Shot Optical Music Recognition via Data-Centric Synthetic Training"
Transcoda reads sheet music and writes the notes as text. It has 59M parameters, trains in 6 hours on one GPU, and beats a model 16x its size. 🧵
Can prompts unrelated to a task still steer an LLM?
With Abtin Pourhadi, Jerzy Sarosiek, Przemysław Spurek, and @pswoboda3, we present Spurious Prompts.
They can improve accuracy, and steer models toward unintended output patterns.
Full paper:
https://t.co/7moApCxIzY
1/3
🧠 Can we unlearn a model without iterative optimization?
🚀 Introducing GROM: Gradient-Free Rapid One-Shot Machine Unlearning, a new approach to efficient machine unlearning without costly iterative procedures.
📄 Paper: https://t.co/cTTUzcgq06
@GMUMJU
Can prompts unrelated to a task still steer an LLM?
With Abtin Pourhadi, Jerzy Sarosiek, Przemysław Spurek, and @pswoboda3, we present Spurious Prompts.
They can improve accuracy, and steer models toward unintended output patterns.
Full paper:
https://t.co/7moApCxIzY
1/3
Excited to share that our paper “PIAST: Rapid Prompting with In-context Augmentation for Scarce Training Data” has been accepted to the ACL Main Conference 2026 (CORE A*) . 1/3
We study how to automatically construct strong prompts in low-data settings by augmenting human-written instructions with automatically synthesized few-shot examples. Our results show that such examples can be a much stronger lever than exhaustive instruction search. 2/3
@Slawomir_Ozdyk W sprawie różnic dusz: Też kościół uczy, że poprzez sakramenty dusza katolika jest przemieniana i otrzymuje signum indelebile. N.p. tylko ksiądz otrzymuje "configuratio ad Christum Sacerdotem". Jednak żydzi nie narzekają, że kościół twierdzi, że oni mają inne dusze.
@mischrev I recommend "Jesus of Nazareth, From the Baptism in the Jordan to the Transfiguration" and "Jesus of Nazareth, From the Entrance into Jerusalem to the Resurrection" by Joseph Ratzinger (Benedict XVI) how the Mosaic law relates to the person of Jesus.
How important is prompt quality-and how can we improve it?
Together with @PaweB96220 and @pswoboda3, we present PRL (Prompts from Reinforcement Learning): a novel RL-based method for automatic prompt generation that achieves SOTA performance across a range of benchmarks.
Excited to share our #arxiv paper “Normalized Matching Transformer” (w/ @pswoboda3): an end-to-end model that learns keypoint correspondences directly, replaces external solvers, and sets new SOTA results.
paper: https://t.co/BU7ODermgs
code: https://t.co/jVtPqM3YnT
Απ᾽ όπου γεννιούνται τα όντα εκεί και καταλήγουν πεθαίνοντας, όπως επιβάλλει η αναγκαιότητα· γιατί λογοδοτούν και επανορθώνουν το ένα στο άλλο για την αδικία που διέπραξαν, σύμφωνα με την τάξη του χρόνου.