The study investigates how text-to-image models interpret number words beyond ten objects. By keeping the object type and initial noise constant, researchers varied the requested count and observed that models reliably follow the number only up to about ten items; beyond that, the count influence weakens and the image composition is largely governed by the coarse noise. Models such as SD3.5‑Medium and FLUX.1‑dev lose adherence after ten, while FLUX.2 and Qwen‑Image loosely track counts up to twenty‑five, yet no clear directional effect is achieved beyond the models’ limits.
arXiv:2610.27756v1 Announce Type: new Abstract: Vision-language models (VLMs) are exact on small sets and drift into estimation on larger ones, a profile reminiscent of children before they master co...
The study measures the exact probability that language models assign to each integer from 1 to N when prompted for a random number, eliminating sampling noise by summing over all spellings. Across 18 open‑weight models and 29 decoding settings, instruction tuning consistently makes the distribution less uniform, introducing a round‑number avoidance heuristic and sharpening endpoint avoidance. While tuned models rarely output the extreme values 1 or N and largely avoid multiples of 10, they do not converge on a single favorite integer across models.
The paper argues that apparent capability limits in vision‑language benchmarks often stem from the way answers are presented rather than from the models themselves. By comparing performance on COCO images with answer choices given as English names versus pixel coordinates, the authors show that models like Qwen3‑VL‑4B perform far better when answers are in natural language, and that the choice of answer format can swing model rankings by dozens of points. The study also demonstrates that different conventions (e.g., hue angles vs. pixel coordinates) reveal which formats a model can actually interpret, highlighting that a fixed answer vocabulary is not neutral across models.
arXiv:2609.23012v1 Announce Type: new Abstract: Generative vision-language models (VLMs) offer a counting paradigm in which one model produces both a count and a natural-language account of the scene...
The study investigates a bias in text‑to‑image models that consistently generate analog clocks showing 10:10, a common advertising pose. Researchers tested 52 models on images requiring clocks to display 2:35, 6:50, and 11:20, varying the prompt to include no time, digits, hand positions in words, or a drawn reference dial. Results show that adding a reference dial dramatically improves accuracy (up to 75% correct) while other prompt variations only modestly reduce the 10:10 bias.
arXiv:2609.11149v3 Announce Type: replace-cross Abstract: How fast does a language model degrade when trained on its own outputs? Theory traces it to gradually accumulating errors, while experiments...
The paper introduces GLANCE, a one‑pass block drafting method that enables lossless speculative decoding for vision‑language models. By using a block‑diffusion head that reads the fused vision‑language state, GLANCE eliminates the need for the drafter to process the image at every step, allowing it to fill an entire block in a single forward pass. Experiments show that GLANCE can decode up to 2.93× faster than autoregressive decoding while maintaining exact greedy decoding results across multiple tasks.
The paper critiques current memorization audits for generative models, arguing that lacking a proper null distribution leads to misleading conclusions. It introduces two exact null tests—one permutation test for whole models and a calibrated test for single images—showing that many previously flagged memorizations disappear under these stricter controls. The authors also propose a scale‑restricted statistic based on the Intersection Euler Characteristic Profile to better detect distinct copied images.
The study investigates whether a multimodal model like Gemma4-12B-it uses a shared internal concept of numbers across different modalities and operations. By applying causal interventions to the model’s residual stream, the authors identify a small subspace where swapping coordinates changes the model’s arithmetic answer. They find that while the subspaces for text and image are not identical, a single rotation and rescaling aligns them, revealing a shared geometric structure that encodes numbers in hidden directions used to generate multi‑digit outputs over time.
arXiv:2610.09313v1 Announce Type: new Abstract: Vision-language models (VLMs) often miss small objects in large images. We ask three questions: what limits them, which of these limits better models c...
Vision-language models (VLMs) are exact on small sets and drift into estimation on larger ones, a profile reminiscent of children before they master counting. Children cross the barrier with epistemic...
The paper introduces the See, Think, Use protocol, which evaluates vision‑language models by tracing each answer back to perception, reasoning, or the use of perception. It employs two additional questions per item—a localization probe for perception and a text‑replacement probe for reasoning—without requiring model weights. The protocol measures how often answers rely on the model’s own perception across counterfactual image pairs, and demonstrates that allocating compute to the failing stage can improve performance at lower cost.