The paper introduces PonderTTT, a gating strategy that uses the self‑supervised reconstruction loss from a Test‑Time Training (TTT) layer to decide when to apply TTT updates during inference. The gating mechanism is training‑free, relying only on a single scalar threshold that is calibrated on unlabeled data and updated via exponential moving average to keep the desired update rate. Experiments on GPT‑2 models (124M–1.5B) for code language modeling show that this approach achieves 82–89% oracle recovery and outperforms random skip baselines, reducing loss on out‑of‑distribution languages by up to 16%.
arXiv:2607. 07748v1 Announce Type: new Abstract: Large Language Models achieve strong code generation for high resource languages like Python and Java but suffer sharp performance drops on Low-Resource Programming Languages~(LRPLs) such as Julia.
The paper investigates whether reinforcement‑learning post‑training of code‑generating large language models can be done entirely offline using existing datasets, avoiding costly online code generation and GPU‑CPU communication. Experiments show that a few hours of offline RL can substantially boost zero‑shot code generation performance across models from 0.5 B to 7 B parameters, though the magnitude of improvement differs by model family.
The study evaluates synthetic pre‑pretraining (PPT) across four curricula and three training regimes, finding that allocating compute to PPT never outperforms pure natural‑language training in compute‑matched experiments. Even when natural language data are scarce, repeated epochs over a small natural‑language corpus beat all synthetic curricula on validation loss, and no consistent downstream accuracy advantage is observed. The authors note that PPT gains are sensitive to weight initialization and that models tend to forget PPT knowledge during subsequent natural‑language training, suggesting limited benefits for PPT as a distinct training stage.
arXiv:2510. 13940v4 Announce Type: replace-cross Abstract: Recent progress in large language models (LLMs) has focused on test-time scaling to improve reasoning via increased inference computation, but often at the cost of efficiency.
arXiv:2601. 12186v3 Announce Type: replace-cross Abstract: Multi-domain thinking verifiers trained via Reinforcement Learning with Verifiable Rewards (RLVR) are a cornerstone of modern post-training.
arXiv:2609.18959v1 Announce Type: new Abstract: LLM code-generation systems usually choose a target programming language before decoding and treat that choice as fixed. We show that, for language-fle...
Verification-Aware Training (VAT) is a plug‑in framework that improves speculative decoding for large language models by simulating verification during training and using the resulting accept/reject patterns as supervision. VAT adds a lightweight binary verification head to predict whether each draft token will survive sequential verification, and replaces the fixed per‑position weighting with a verification‑adaptive schedule that keeps full weight up to the first rejection point. When applied to EAGLE‑3 and DFlash on Qwen3‑4B, Qwen3‑8B, and LLaMA‑3.1‑8B, VAT increases average acceptance length by up to 11.4% and wall‑clock speedup by up to 8.7%, yielding consistent gains across math, code, and chat benchmarks.
The paper introduces HOW TO TRAIN YOUR LITTLE DRAGON, a method for teaching a small language model to generate and refine GPU optimization scripts. Using a compact, lossless representation of compiler traces, the authors train the model on 108,594 examples derived from TVM MetaSchedule, first with supervised fine‑tuning and then with reinforcement learning that rewards correctness and performance. After training, an 8‑billion‑parameter model achieves over 91% correctness in both construction and refinement tasks and outperforms several much larger LLMs on kernel performance.
arXiv:2607. 09415v1 Announce Type: cross Abstract: Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs.
arXiv:2603. 15510v2 Announce Type: replace Abstract: The synthesis of inductive loop invariants remains a critical bottleneck in automated program verification.
The paper investigates training-time data augmentation as a regularizer for autoregressive language model pretraining in data‑constrained, compute‑abundant settings. It introduces three orthogonal augmentation categories—token‑level noise, sequence permutations, and target offset prediction—and shows through systematic ablations that each category delays overfitting and reduces validation loss, with random token replacement performing best individually. Combining augmentation categories further lowers the minimum validation loss, demonstrating that such augmentations mitigate data inefficiency in autoregressive pretraining.
arXiv:2604.24957v3 Announce Type: replace-cross Abstract: Scaling test-time compute has emerged as a powerful mechanism for enhancing Large Language Model (LLM) performance. However, standard post-tr...