An AI model trained with a verifier-based method scored higher than the base version of the same model in a visual-grounding benchmark, according to an arXiv preprint. On PointArena Point-Bench, the PointRL-trained Qwen3.5-4B model reached 65.58% overall accuracy, compared with 56.11% for the untrained 4B model, an observed difference of 9.47 percentage points.
The study asked whether verifier-based point-level training improved same-backbone grounding, whether reward configurations behaved differently, and whether gains transferred to external benchmarks without benchmark-specific fine-tuning. The base and PointRL-trained models were evaluated with the same protocol, prompt, decoding setting, parser and scorer.
A check that stays behind the prompt
PointRL converts boxes, masks and instance labels into pointing instructions. The information used to verify an answer, including the valid target areas, which instances belong together and the required set constraints, stays outside the prompt, allowing deterministic checks.
The verifier parses responses and combines checks of point location and coverage with count, duplicate consistency and instruction-specific guard terms.
Accepted samples were derived from AGD20K labels, PixMo points, COCO masks and instance metadata, together with rule-bound cue-target pairs. The study used 1,647 constructed training samples, a separate 200-sample held-out split for diagnostic validation and a 982-sample Point-Bench test set.
The smaller model showed the largest jump
The Qwen3.5-2B result was the largest overall gap reported on PointArena Point-Bench: 56.82% accuracy for the PointRL-trained model versus 25.25% for the base model, an observed difference of 31.57 points. The paper reports this as a single-run comparison.
The 4B model's category results were uneven. Compared with its base model, the observed differences were 13.26 percentage points for counting, 11.92 for reasoning, 10.50 for steerable tasks, 9.59 for affordance and 2.05 for spatial; the spatial baseline was 69.23%. These category-level differences were not tested for statistical significance.
Reward configurations also produced different scores in a single-run comparison. The full reward scored 65.58%, compared with 60.18% for localization only and 62.42% for localization plus coverage, making the full bundle the highest observed configuration. Because the comparison used bundles, it did not isolate the contribution of individual reward terms.
The pattern extended to outside tests
The study also reported positive observed same-backbone differences on external benchmarks: 7.38 percentage points on RoboSpatial context, 7.23 on configuration, 13.34 on compatibility and 7.00 on BLINK. These external-benchmark differences were reported without statistical-significance claims.
On Ref-Adv, the observed gains were 2.01, 3.33 and 1.93 points across the reported threshold settings, and 1.29, 3.49 and 2.32 points in the 2 to 3, 4 to 6 and at least 7 distractor settings. Ref-Adv uses a box-output protocol rather than direct point output.
What the comparison leaves open
Each trained configuration used a single checkpoint, per-example agreement statistics were not reported, and no statistical-significance claim was made for category, ablation or external-benchmark differences.
External model rows were contextual rather than fully controlled. The document is an arXiv version 1 preprint dated 26 Aug 2026.
Paper data and sources
Original title: PointRL: Learning Point-Level Vision-Language Grounding from Verifiable Annotation Evidence
Authors: Jingyang Su, Pu Cao, Xiuze Jin et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text