Vision‑language agents (VLMs) invoke visual calls when they need to crop or zoom into an image. Existing reward schemes only credit a successful tool call, without checking whether the model actually needed to look or whether the returned pixels were used.
On a cold‑start checkpoint, only about 10%–12% of visual calls were both needed and used, while released agents make spurious calls 36%–87% of the time on individual benchmarks. Outcome rewards, judge rewards, and branch probes each capture only one side of this failure, and roughly two‑thirds of the reward budget goes to calls that were neither needed nor used.
CounterCredit asks two questions for every image‑returning call at its realized pre‑call state, using the policy’s own gold‑answer score. A decision value compares the realized visual branch with answering immediately; an evidence value compares the returned crop with random same‑size patches substituted into the same call. A call that passes both checks earns cashback, while every other executed call pays rent. The price is bounded so that any correct trajectory’s cumulative payoff outranks any wrong one, and a dual‑channel GRPO (Generalized Reward Policy Optimization) advantage keeps the price in its own units.
From the same cold start, prompt pool, and budget, CounterCredit reaches 89.5% on V*, 80.2% on HR‑Bench‑4K, and 76.4% on HR‑Bench‑8K—6.3 to 9.4 points above outcome‑only GRPO—while reducing the average calls per question from 1.84 to 1.78 and lowering the spurious‑call rate to 31%–36%, the lowest among evaluated agents. The same recipe lifts a Qwen3‑VL‑8B base from 75.4 to 80.8 on average.
These results demonstrate that charging only for visual calls that are truly needed and used can markedly improve VLM efficiency and performance, eliminating a large amount of redundant visual processing.
Review: CounterCredit’s dual‑verification and economic incentive design tightly couples visual call necessity with actual usage, offering a refined reward framework for future vision‑language agents.