The paper reports on a verifiable‑reward tuning approach applied to an Ornith‑style vision‑language model for multi‑view driving question answering. Using a 256‑token limit, the GRPO method raises answer extraction from 78.25 % to 90.25 % and accuracy from 55.50 % to 63.75 %. On a larger 2,511‑question set, the tuned pipeline cuts average token usage from 108.6 to 47.1, though the accuracy improvement is not statistically significant.
arXiv:2606. 01080v1 Announce Type: cross Abstract: Large language models often improve on difficult tasks by spending inference-time compute on a reasoning trace before producing the final answer.
arXiv:2607. 14099v1 Announce Type: cross Abstract: Deploying Vision-Language Models (VLMs) in real-world settings requires not only strong visual reasoning but also stability under sustained conversational pressure.
The paper introduces MOTIVE, a multi-view self-verification framework for vision-language models that evaluates candidate answers from complementary perspectives and learns a reliability score aligned with correctness. By using this score to decide whether to accept an answer or trigger history-guided rethinking, MOTIVE improves reliability and efficiency compared to existing self-verification and self-correction methods. Extensive experiments on diverse multimodal benchmarks and VLM backbones show consistent performance gains and reduced unnecessary reasoning turns.
arXiv:2606. 11209v1 Announce Type: cross Abstract: Visual question answering increasingly requires multi-step reasoning.
arXiv:2512. 11995v2 Announce Type: replace-cross Abstract: While many vision-language models (VLMs) are developed to answer well-defined, straightforward questions with highly specified targets, as in most benchmarks, they often struggle in practice with complex open-ended tasks, which usually require multiple rounds of exploration and reasoning in the visual space.
arXiv:2607. 15565v1 Announce Type: cross Abstract: Where should the question go in a vision-language model (VLM) prompt: before the image or after it?
VBVR-Pro is a closed‑loop testbed that enables native visual reasoning through generation, offering 300 procedurally generated tasks that scale training and allow strong transfer to external benchmarks. It supplies verifiable reward scorers based on deterministic, task‑specific rules, outperforming VLM‑as‑a‑judge approaches and providing reliable signals for reinforcement learning. The suite also facilitates controlled modality studies, revealing that video generation excels at persistent spatiotemporal tracking while interleaved generation offers a compute‑efficient alternative, and highlights the importance of vision‑native trajectories for reasoning.
VAMR is a video agent that handles multiple questions about a single long‑form video using a single shared tool‑use trajectory, rather than treating each question separately. It employs question‑conditioned visual perception, layered multi‑question memory, and a question‑horizon policy optimization to jointly acquire evidence and progressively build a shared understanding. In benchmark tests (LVBench, Video‑Holmes, LongVideoBench), VAMR outperforms prior methods, achieving higher accuracy and fewer reasoning rounds and processed frames.
PRISM-BENCH is a new benchmark comprising 1,000 expert‑curated vision‑language questions across five under‑represented reasoning categories, each paired with an idealized reasoning trace validated by a second annotator. It introduces diagnostics beyond final‑answer accuracy: a perception probe that attributes about 85% of frontier‑model failures to misperception, and trace alignment that predicts answer correctness with a macro AUROC of 0.66. The benchmark also presents CORE, an efficiency‑aware metric that reorders 24.2% of model pairs relative to accuracy while remaining uncorrelated with response length, revealing that recent gains are driven by recognition rather than spatial grounding.
arXiv:2607. 05438v1 Announce Type: cross Abstract: Multimodal retrieval-augmented generation (RAG) grounds a generator in evidence drawn from heterogeneous modalities -- text, tables, and images.
The paper shows that chain‑of‑thought (CoT) instructions can distort evaluation of vision‑language models (VLMs) when a scorer reads answer‑label logits before the model generates a rationale. On ScienceQA, Qwen2.5‑VL‑7B’s accuracy falls from 80.76% to 45.48% under this CoT‑prefix scoring, and most predictions incorrectly pick the first option. Linear probes and free generation recover most of the lost accuracy, indicating that the answer information remains in the hidden states but is missed by the early readout. The authors explain the mismatch with vocabulary and layer diagnostics, noting that probability mass shifts toward continuation tokens while answer information stays linearly accessible in later layers. The effect varies across datasets and models, but the study demonstrates that CoT‑prefix scoring can misrepresent model knowledge unless the requested and scored outputs are aligned.
arXiv:2607. 28374v1 Announce Type: new Abstract: Multimodal agents for visual question answering increasingly operate as multi-step trajectories that interleave perception, retrieval, and reasoning, yet evaluation still largely reduces to final-answer accuracy.