The paper introduces Concise-9B, a version of the Qwen3.5-9B reasoning model that reduces unnecessary thinking tokens through concise self‑training. By applying two rounds of on‑policy rejection‑sampling fine‑tuning and a small share of answer‑synthesis targets, Concise-9B cuts token usage to roughly two‑thirds of the original model while maintaining comparable accuracy on a wide range of hard mathematics and reasoning benchmarks. The study also shows that most of the token savings occur after the answer has already formed, and that the approach’s effectiveness depends on the verbosity of the base model.
By Alexander Boesgaard Lorup
arXiv:2609.01274v1 Announce Type: new
Abstract: Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but how these gains relate to inference-time decoding and sear...
By Wenhe Sun, Cunxiang Wang, Zijun Yao, Yixin Cao
The paper investigates whether the reported out‑of‑domain performance gains of Reinforcement Learning with Verifiable Rewards (RLVR) are due to improved answer quality or simply to changes in answer formatting. By comparing scores from standard extractors with a custom extractor that distinguishes answers inside a oxed{} notation, the authors find that a large portion (about 82%) of the apparent gains on 7B SimpleRL models vanish, indicating that the gains are largely due to formatting rather than better answers. In contrast, a 32B model trained with a different recipe shows minimal format‑related gains, suggesting that the effect depends on the training setup.
By Zongye Lyu
The paper evaluates how three large mixture‑of‑experts models (Alibaba, OpenAI, NVIDIA) can be fine‑tuned to reason in a low‑resource language, specifically Greek. Accuracy metrics show little change, but the authors uncover significant qualitative improvements: after supervised fine‑tuning, models reason in Greek on ~98% of items, with better grammaticality and retained general ability. Reinforcement learning with pre‑registered rewards further eliminates reasoning‑channel leaks and format skips, while the Greek‑reasoning habit remains robust to an accuracy‑only gradient.
By Ayoub Kirouane, Christos Petrocheilos
Reinforcement learning with verifiable rewards (RLVR) makes a language model solve more problems on the first try and fewer when given many tries: it raises pass@1 while the base model overtakes it at...
The paper investigates whether large language models’ reasoning traces truly contain early, informative signals or merely reflect budget and difficulty confounds. Using a restart‑controlled truncation probe, the authors compare continuation success rates against from‑scratch restart curves across 178 problem‑model pairs, finding that only one case shows prefix‑limited success and that continuing a model’s own prefix generally outperforms restarting. A difficulty‑controlled test and two generation‑free analyses reveal that early internal signals do not carry outcome information beyond a problem‑difficulty baseline, underscoring the need for proper counterfactual controls.
By Yigit Utku Bulut
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
The paper argues that the observed crossover in pass@$k$ performance between reinforcement learning with verifiable rewards (RLVR) and its base model is not always statistically confirmed. By constructing confidence bands across sampling budgets and performing power analyses, the authors show that many reported crossovers lack statistical support and that more prompts can improve detection more than more answers per prompt. They further demonstrate that RLVR’s effect on a prompt is better described as a conditional distribution (a Markov kernel) rather than a single curve, allowing predictions of crossovers in new data and clarifying how losses on difficult prompts can overturn early gains.
By Chen Yang, Xianyang Zhang, Jun Chen
The paper proves that in reinforcement learning with verifiable rewards (RLVR), any improvement in pass@1 necessarily incurs a cost in pass@k, establishing a precise price for such gains. By modeling RLVR as an exponential tilt of a base policy, the authors show that every update that raises pass@1 strictly lowers success on at least one prompt, and that the number of pass@k crossovers is bounded by the sign variation of per‑prompt success shifts. The work also derives an entropy‑performance law and provides a diagnostic to measure this price from evaluation logs.
By Yinuo Liu
The paper reports on training and calibration experiments for three open models on the ARC‑AGI‑2 benchmark, using offline runs on four L4 GPUs for up to 12 hours. For the NVARC Qwen3‑4B grid model, a curated corpus of 12,000 executable rules and 95,993 tasks was built, and fine‑tuning on 1,920 real task records raised the semi‑private score from 28.96 to 33.06, with no synthetic‑task checkpoint outperforming the real‑task model. For the Qwen3.8‑27B reasoning model, fine‑tuning on 1,036 of its own reasoning traces and a specialized inference procedure increased the score from 26.81 to 44.03, while the Qwen3.8‑Flash‑Next mixture‑of‑experts model achieved a best score of 56.67 after expert pruning, routing statistics adaptation, and an FP8 indexer cache.
By Nguyen Minh Thang, Trinh Thi Phuong
The paper examines on‑policy self‑distillation, where a teacher scores a student’s rollout to provide dense token‑level credit. Experiments show that the raw scoring signals neither localize errors nor effectively penalize incorrect tokens, because locating errors is a search problem that the teacher is not explicitly asked to perform. Introducing a critique‑conditioned approach—first generating a critique of the rollout and then conditioning the teacher on it—produces signals that accurately localize errors, correct erroneous spans, and guide the policy toward backtracking tokens, improving performance on math problems and in cold‑start GRPO settings.
By Mayug Maniparambil, Ayush Rai, Ekansh Verma, Noel E. O'Connor, Fergal Reid
arXiv:2606. 15455v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a key approach for enhancing the reasoning abilities of large language models.
By Suqin Yuan, Jinkun Chen, Jiyang Zheng, Muyang Li, Lei Feng, Dadong Wang, Tao Xiang, Tongliang Liu, Bo An