arXiv:2609.37169v1 Announce Type: cross Abstract: Mid-training equips pretrained large language models with specialized and reasoning capabilities, but the returns of this stage are bounded since add...
UpgradeBench is a decision‑centric longitudinal benchmark that evaluates how fine‑tuned language‑model specialists should be handled when new base‑model releases occur. It covers four consecutive Qwen releases, a continuation checkpoint, six tasks, two model sizes, and OLMo checkpoints with known training lineage, and examines whether retraining, adapter transfer, or other recovery strategies improve specialist performance. The benchmark reveals that upgrade gains vary by task and release interval, that direct adapter copying is sensitive to pretraining distance, and that teacher relabeling can recover specialists without new annotations. "whyItMatters":"The study provides actionable insights into the cost‑effective management of specialist models across model releases, showing how to balance retraining effort with performance gains."
Osprey is a target‑agnostic pre‑training method that bootstraps draft models for speculative decoding from existing small language models. By pruning to a shallow backbone, restoring language‑modeling capability with next‑token pretraining, and adapting via vocabulary alignment and distillation, Osprey reduces per‑target work to a lightweight adaptation step. Experiments show that a single Osprey backbone improves mean acceptance length by up to 22.7% and increases tokens per second by 17.5% across several large target models, especially on out‑of‑domain and multilingual data.
PROOF-Gen is a method that improves distillation of tool‑calling models by recovering successful trajectories from teacher failures. It uses per‑scenario prompt optimization to generate corrective guidance that steers the teacher to a passing trajectory, then removes this guidance before training so the student learns from clean demonstrations. On τ2‑bench, PROOF-Gen recovers 93% of failed scenarios, boosting Qwen3‑4B‑Instruct‑2507’s Pass^1 from 0.132 to 0.529 and improving Gemma 4 E4B‑it by 7.2pp on BFCL v4 multi‑turn, while also raising deployed on‑device model performance by up to 5.0pp across response‑quality metrics.
arXiv:2606. 26671v1 Announce Type: new Abstract: Post-training alignment determines the reasoning and human preference following capabilities of large language models, yet most existing works withhold detailed data construction, filtering rules and training recipes, which hinders community reproducibility and lightweight model optimization.
arXiv:2609.35793v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly trained with reinforcement learning from verifiable rewards (RLVR). An exact verifier can also support te...
arXiv:2606. 06902v1 Announce Type: new Abstract: Targeted post-training aims to improve reasoning, math, and code without degrading strengths.
arXiv:2608. 09351v1 Announce Type: cross Abstract: Test-time scaling improves LLM accuracy but multiplies inference cost, making the accuracy gained per unit of compute the metric that matters in deployment.
arXiv:2606. 03938v1 Announce Type: cross Abstract: Multi-epoch training is becoming the standard now that compute is growing faster than the supply of high-quality text.
The paper introduces READ, a method for composing low‑rank adapters (LoRA) in large language models. By rewriting each adapter into a balanced canonical form and enforcing a one‑directional coupling, READ allows new skills to read but never write into the output subspaces of existing skills, eliminating interference. Experiments on four benchmark suites and two model families show that READ consistently outperforms existing baselines, improving SuperGLUE scores by over twenty points and domain suite scores by more than seven points.
The study evaluates whether a single large language model (LLM) can handle multiple customer‑support tasks or if separate specialist models are preferable. Using 13 models from five families and 200+ checkpoints across eight datasets, the authors find that multi‑task full fine‑tuning consistently outperforms other strategies. They also show that sequential LoRA and model merging can preserve earlier skills and improve off‑task robustness, offering practical guidelines for real‑world deployment.
FCPRAG introduces a fusion-controlled parametric retrieval‑augmented generation framework that uses a lightweight controller to predict per‑passage fusion scores and sample‑level calibration signals, such as a mixing gate and adaptive temperature. This approach mitigates the bottleneck of evidence‑level fusion when multiple passages are retrieved, enabling selective fusion under informative signals and conservative fusion under uncertainty. Experiments on HotpotQA, 2WikiMultiHopQA, PopQA, and ComplexWebQuestions demonstrate consistent F1 improvements over standard RAG and parametric RAG baselines, with gains up to 4.65% on 2WikiMultiHopQA and 7.55% on CWQ, while also reducing tuning cost and enhancing robustness to retrieval perturbations.