The Flow has not summarised this story yet — read it at arXiv Computation and Language.
The paper introduces V‑Rubrics, a reinforcement‑learning framework that evaluates vision‑language model responses by breaking them into atomic propositions and scoring them on Visual Faithfulness, Reasoning Consistency, and Instruction Following. Using a fine‑tuned Qwen3‑VL‑8B‑Instruct model and a newly created 50K‑example V‑Rubrics dataset, the authors demonstrate that rubric‑based GRPO outperforms both a shared SFT baseline and an answer‑only GRPO, especially on knowledge‑oriented and visually grounded reasoning tasks.
The paper introduces SymCE, a dataset of 4,707 false undergraduate‑algebra and real‑analysis conjectures each paired with a deterministic Python verifier that can be executed to check truth. It studies counterexample generation by training a 4B‑parameter language model (Qwen3‑4B) first with supervised fine‑tuning and then with reinforcement learning (GRPO) using the verifier as a reward. The results show that counterexample‑only fine‑tuning can destroy true‑theorem recognition, while reinforcement learning with a sparse outcome‑only reward restores and surpasses baseline performance, achieving higher success rates than larger open‑weight models and competitive performance on other math benchmarks. "whyItMatters":"The study demonstrates that reinforcement learning guided by a per‑theorem verifier can effectively repair the shortcomings of supervised fine‑tuning in theorem proving, offering a practical approach to improve language models’ mathematical reasoning capabilities."
The paper introduces an audit layer for vision‑language model benchmarks that reuses per‑question artifacts from evaluation runs to assess the benchmark itself. Six detectors analyze saved outputs, and an additional text‑only rerun removes visual inputs to produce per‑question reports and aggregate statistics, including permutation‑null excess and interval estimates. Applying this method to the MMStar benchmark reveals that 20.2% of its questions show no accuracy gain from images, with a small but significant excess in mathematics and science categories, and identifies questions where all models chose the same non‑gold option.
The paper argues that multimodal large language models can answer vision benchmarks using only linguistic cues, revealing non-visual shortcuts. It proposes a Test‑set Stress‑Test (TsT) that trains a text‑only model on the test set to expose exploitable patterns, and introduces Iterative Bias Pruning (IBP) to remove high‑bias questions. Applying these methods to VSI‑Bench, CV‑Bench, and MMMU shows significant gaps between blind and vision‑aware performance and demonstrates that debiased benchmarks can reduce blind scores while preserving vision scores.
arXiv:2607. 05904v1 Announce Type: new Abstract: Training a language model against its own reference-free judgments (the premise of self-rewarding, self-play, and LLM-as-a-judge pipelines) assumes a model's verdict on a shown answer tracks correctness.
arXiv:2608. 15445v1 Announce Type: new Abstract: When a reward is correct on every training example yet consistent with more than one goal, a model can acquire an unintended one, a failure known as goal misgeneralization.