C. Thomas

Pouya Sadeghi, Shawn He, Pedro Pablo Guerrero Vela et al.

Modern referring image segmentation pipelines couple a vision-language model (VLM) for grounding with a promptable segmenter such as the Segment Anything Model (SAM) for mask generation. Prior training-free instances of this recipe consistently trail fine-tuned and reinforcement-learning (RL)-tuned specialists, and it has been unclear whether the gap comes from the VLM's grounding, SAM's capacity, or the prompt. We show that the gap is dominated by prompt ambiguity: a VLM-proposed bounding box (bbox) leaves SAM to guess which pixels inside the bbox belong to the object the expression denotes. Interior points are the natural disambiguator, but where they fall matters; prior work relies on naively sampled points that land on boundaries, distractors, and background clutter, and can even hurt performance compared to the bbox alone. Supervised and RL-tuned methods close this gap by training a VLM to predict better points; we show that this training is unnecessary. At a matched budget of five interior points, replacing naive sampling with stable, informative point selection improves cumulative Intersection-over-Union (cIoU) by 12-18 points across RefCOCO/+/g, with every model fixed. We turn this observation into PinPoint, a deterministic, training-free point selector that fuses four visual cues into a consensus map, selects compact, spatially diverse points away from boundaries, and uses the frozen VLM to label each point. Without any task-specific training, PinPoint matches supervised and RL-tuned specialists on the same stack while issuing only two VLM calls per query.

A. Leone De Castris, C. Thomas

Governments, industry, and other actors involved in governing AI technologies around the world agree that, while AI offers tremendous promise to benefit the world, appropriate guardrails are required to mitigate risks. Global institutions, including the OECD, the G7, the G20, UNESCO, and the Council of Europe, have already started developing frameworks for ethical and responsible AI governance. While these are important initial steps, they alone fall short of addressing the need for institutionalised international processes to identify and assess potentially harmful AI capabilities. Contributing to the relevant conversation on how to address this gap, this chapter reflects on what functions an international AI safety institute could perform. Based on the analysis of both existing international governance models addressing safety considerations in adjacent policy areas and the newly established national AI safety institutes in the UK and US, the chapter identifies a list of concrete functions that could be performed at the international level. While creating a new international body is not the only way forward, understanding the structure of these bodies from a modular perspective can help us to identify the tools at our disposal. These, we suggest, can be categorised under three functional domains: a) technical research and cooperation, b) safeguards and evaluations, c) policymaking and governance support.