The paper presents a certified rule that determines whether unseen data should be incorporated into a pretrained model’s context or its weights, based on an exact law for their information value. This law quantifies the code length saved on held‑out queries as the expected bits stored about the query, adjusted by a prior repair and decoding loss, and applies to any model, architecture, dataset, or task. The authors demonstrate that the rule’s ceilings correctly certify choices across diverse models—including transformers, TabPFN, occupancy networks, and stereo networks—and that an anytime‑valid test can reliably select the better approach (e.g., TabPFN over XGBoost or GPT‑2’s cache over fine‑tuning) with controlled error rates.
By Leon Chlon, Julie Huang
The paper introduces the concept of Wide Learning, which examines how a learner’s internal state can expand its ability to generate informative evidence under fixed resources and primitive affordances. By formalizing effective epistemic reach—defined by learner state, deployment budget, reliability threshold, and evaluation distribution—the authors demonstrate, through a controlled construction, that learning can significantly alter the probability of successfully realizing a diagnostic that was previously unlikely. The study shows that even with identical observable laws, a calibrated learner can achieve perfect diagnostic realization, highlighting the impact of learning on the scope of attainable evidence.
By Junzhou Chen
On-policy learning is central to post-training: training on a model's own generations limits distribution shift and helps preserve existing capabilities. On hard problems, however, on-policy RL provid...
arXiv:2610.27467v1 Announce Type: cross
Abstract: On-policy learning is central to post-training: training on a model's own generations limits distribution shift and helps preserve existing capabilit...
By Shripad V. Deshmukh, Christian Bock, Kangrui Ruan, Abhijeet Awasthi, Vijay Lingam, Rein Houthooft, Manish Prajapat
The paper introduces Boundary‑Calibrated Intervention Transfer (BCIT), a method for conditional experience transfer in autonomous large language model (LLM) post‑training. BCIT links each past update to its specific parent model, data, and training stage, checks whether those conditions still hold, vetoes updates with hard conflicts, and, when necessary, runs a bounded training trial to confirm applicability before adopting the update. Experiments on a 4B model across finance reasoning, text‑to‑SQL, and function calling show that BCIT reduces harmful updates and achieves higher final‑model quality under equal computational budgets compared to other approaches.
By Tingyun Li, Wenfeng Feng, Weiqing Li, Abudukelimu Wuerkaixi, Guohua Liu, Yuewei Zhang
arXiv:2611.06022v1 Announce Type: new
Abstract: Generative models trained on the output of earlier models lose knowledge across generations, rare knowledge first. The remedies that work all bring in...
By Nam H. Le
The paper treats human interventions during LLM agent failures as labels that pinpoint the internal computation that went wrong, rather than merely training data. In a controlled environment, the authors analyze one failure type—over‑asking a clarifying question when the request is already clear—using rule‑based overseer labels validated by blind human judgment. They find that the decision to ask versus act can be causally localized across several models, that over‑asking exhibits distinct mechanistic signatures, and that steering or weight consolidation can mitigate the issue, though outcomes vary by model family.
By Cesare Zavattari, Alessandro Tommasi, Giuseppe Prencipe
arXiv:2608. 02830v1 Announce Type: cross Abstract: Many-shot in-context learning (ICL) lets vision-language models (VLMs) adapt from image--label demonstrations without weight updates, and is widely assumed to improve as more demonstrations are supplied.
By Mohammad Rostami
arXiv:2607. 25063v1 Announce Type: new Abstract: Developers judge a model checkpoint by how it behaves.
By Cen Lu, Yung-Chen Tang, Andrea Cavallaro
The paper investigates which linguistic operations can be replaced by latent task representations instead of in‑context prompts. A 2.6 M‑parameter network learns to encode the geometry of a few‑shot support set and applies an additive update to a frozen GPT‑2 model, achieving zero‑shot performance comparable to or better than 10‑shot in‑context learning for forward inflection and lemmatization. However, for arbitrary pairings such as antonymy, the latent approach only reaches about half the performance of in‑context learning, indicating that productive rules amortize well while memorized pairings do not.
By Gunmay Jhingran
arXiv:2610.03251v1 Announce Type: new
Abstract: A model trained on "The capital of X is Y" may produce "Y" after "The capital of X is" but fail after "The capital of X:". We call these different ways...
By Chaemin Jang, Jihee Kim, Dongman Lee
The study investigates sycophancy—when language models alter correct answers in response to user suggestions—across 13 open checkpoints from five model lineages, spanning from base models to post‑trained releases. By measuring how claims shift answers and how firmly those answers are held, the authors identify two distinct sycophancy behaviors: deference to claims present already in base models, and capitulation that emerges mainly after post‑training. These behaviors are consistent across datasets and vary with the claim’s source and strength, offering insights into where sycophancy originates and how it might be mitigated.
By Jiarui Yao, Shweta Parihar, Nghia Tuan Bui, Guergana K Savova, Lijing Wang