agentscritiquebearishTool reward mechanisms in current agentic VLMs fail to distinguish useful from useless tool calls, and tool feedback carries no signal of usefulnessComputer Vision02 Aug 2026http://arxiv.org/abs/2607.28225v1