The paper evaluates Laya, a small model that processes a multiple‑choice question, its options, and a text in a single 512‑token sequence. Experiments show that when the relevant evidence lies beyond the window, accuracy drops to about 26–27%, and confidence metrics fail to warn of truncation. Fine‑tuning on longer documents improves performance on visible evidence but not on cut evidence, and a larger window or a 312‑token scan can yield modest gains.
By Mohammed Nayeem Khan, Allan Kevin T
The paper introduces a selection‑safe conformal early‑exit method that calibrates a single split‑conformal quantile using a terminal‑path score derived from prefix maxima. This approach guarantees coverage for any data‑dependent stopping policy that preserves terminal‑set containment, without requiring an α/L correction. Empirical results on eight encoder backbones and CIFAR‑100 show that pathwise selection achieves higher macro coverage (0.901 vs. 0.814) with smaller set sizes and lower latency fractions compared to naive layer‑wise selection, and remains robust to deployment‑time budget changes.
By Adria Binte Habib, Ibne Farabi Shihab
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
arXiv:2610.15218v1 Announce Type: new
Abstract: A recurrent-depth language model reuses one block of weights several times in a row. The number of repetitions, which we call loops, is chosen when the...
By Seyedarmin Azizi, Erfan Baghaei Potraghloo, Arya Fayyazi, Saeid Shokoufa, Massoud Pedram
arXiv:2608.20873v1 Announce Type: new
Abstract: Every way of teaching a deployed language model something new -- full fine-tuning, adapter merging, model editing -- replaces the released checkpoint,...
By Zifeng Liu, Zhiyong Du, Yaxin Lu, Yiming Mao, Zhenhe Wang, Wenqi Shi, Zhengkun Jing
The paper shows that AI‑text detectors, rather than learning a clear AI‑versus‑human boundary, amplify an inherited predictability axis that already exists in language models. This amplification causes detectors to over‑flag fluent, formal human writing while missing high‑temperature AI outputs, and the bias persists across languages, code, and detector architectures. A training‑free operator can relocate the bias but cannot erase it, underscoring that the unfairness is a structural cost of out‑of‑distribution generalization.
By Alexander Smirnov
The paper reports that in English all‑words word sense disambiguation (WSD), the scarcity of high‑quality labels—not the models—has become the limiting factor. The authors introduce lexEN, a human‑adjudicated correction layer over the Maru2022 ALL_NEW benchmark, and SenseBench, a living leaderboard for LLM WSD evaluation. They show that frontier large language models reach about 95 % accuracy on lexEN‑v1, that relabeling corpora with these models improves downstream systems, and that fine‑grained WordNet senses are often ill‑posed, with coarsening improving both annotator agreement and model performance.
"whyItMatters":"The study highlights that improving label quality and managing annotation costs are now the critical challenges for advancing WSD performance, as model accuracy is already near its theoretical ceiling."
By Vassili Philippov, Amro Salman, Dmitrii Andreev, Penny Hands, Emil Kaiumov, Pavel Katunin, Anton Nikolaev
Depth pruning removes entire Transformer blocks to shrink and speed up language models, but the resulting damage is uneven across capabilities. In a study of four 1.2–1.7 B instruction‑tuned models, function calling and mathematical reasoning collapsed in every case, while other skills varied in order of degradation. Fine‑tuning a short ‘healing’ run can restore some abilities, with the choice of which blocks to retrain and the allocation of a limited token budget influencing recovery outcomes.
By Piyush Dhoka, Vedant Pandhare, Varun Inamdar, Aadarsh Pathre, Pavitha Nooji
Prefill‑only decision models evaluate every candidate in a menu in a single forward pass, avoiding decoding and reducing cost by one to two orders of magnitude compared to generative language models. The paper demonstrates that when only the candidate menu changes, the model’s post‑intervention accuracy can be predicted solely from the cached first‑pass distribution using a simple estimator that renormalizes and selects the argmax, without any labels or second pass. Across seven model families, ten datasets, and two task types, this menu‑only intervention prediction is within 4.2 points of actual accuracy, and in one family it is exact, whereas a probability‑level variant fails by 21 points, indicating the property resides in ranking rather than calibrated probabilities.
By Ran Li, Lei Chen
arXiv:2607. 20436v1 Announce Type: cross Abstract: Safety evaluations often assume that behavior observed during testing reflects behavior in ordinary use, but fine-tuning can break this assumption.
By Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz
The paper investigates how synthetic tasks influence language‑model pretraining by distinguishing diagnostic, teachable, and transferable signals. In a controlled setting where 70% of the data is general Python and 30% is a mixture of curated synthetic tasks, literature‑derived probes, and OpenCodeInstruct, 14 of 27 tasks show teachability, with curated tasks outperforming literature probes. Curated synthetic data provides conditional value: it boosts downstream performance when mixed with OpenCodeInstruct but is less effective when used alone, and a loss‑based scheduler can misalign short‑term loss reduction with long‑term transfer.
By Ohad Rubin
Model Annealing is a data‑efficient learning method that, after an initial memorization phase, spends the remaining training budget on review and consolidation. The training is split into episodes with cyclic learning rates; each episode learns from true labels, tempered predictions stored in a memory of past models, and adversarial ‘dreams’ where the student and memory disagree most. Across several datasets and architectures, Model Annealing improves top‑1 accuracy, likelihood, calibration, and robustness compared to standard cross‑entropy training and matches snapshot distillation while adding no inference overhead.
By Ali Mohammadi Ruzbahani, Rezvan Abouzadeh