An AI was handed a magnifying glass and did worse. A new preprint says it learned where to aim.

The paper

"MLLM" means multimodal large language model: a chatbot that can also look at images.

The claim in one number

That is the average score the authors report across seven vision tests.

Picture a test with 100 questions. The authors say their trained model gets about 72 right.

Their own untrained model, using the same tools, gets about 56. The best rival they tested, Gemini-3-Flash, gets about 66.

Source: Table 2 of the paper, v1. Seven benchmarks, one shared tool loop. Reported by the authors, not replicated.

A big jump. But the model that started at 55.9 had a secret. It used to score higher.

First principles

What goes wrong when an AI answers a question about a picture?

Answering takes two jobs. First the AI looks. Then it thinks.

If the look is wrong, clever thinking cannot fix it. Miscount the berries in a bowl and the best reasoning still gives the wrong number.

The paper argues that perception, not reasoning, is the weak link for many vision models.

So labs try to train the looking. That turns out to be hard to grade.

The problem

How do you grade "look carefully"?

A long description of a photo holds dozens of small claims. Some are right and some are wrong. No single score captures that.

The authors' trick is to let the model describe a picture and then hunt for its mistakes. Two different AI checkers must agree on each mistake. Each confirmed mistake becomes a quiz question with one right answer.

The authors say a final manual check screens the questions. Their training set holds 36,445 of them. Factas reported in Table 3 and Section 3.2.2.

That trains one good look. Hard pictures need more than one look. So the team added a tool.

The twist

They gave the untrained model a magnifying glass. Six of seven scores fell.

The tool is a code interpreter. The model can crop, zoom, mark and measure the picture before it answers.

You would expect better scores. The authors' own tables show the opposite for their untrained base model.

On BLINK it fell from 68.8% to 49.8%. On CountQA it fell from 66.8% to 49.7%. Six tests went down and one stayed level.

Why would a closer look make things worse?

The cause

A magnifying glass only helps if you point it right.

To count berries, the model marks each one, then zooms in on what is left. If a mark lands a little off, the next zoom looks at the wrong patch.

The authors call this coordinate drift. Each small miss feeds the next look, and the errors stack up.

In one example in the paper, the model counted 23 berries when there were 21. Factone illustrative case, Figure 3.

So how do you teach a model to aim?

The turn

The fix: grade the aim, not only the answer.

Normally an AI is graded once, on its final answer. A lucky guess and a careful look score the same.

The authors add two things. They correct the aim in the practice examples, which they call Coordinate Injection. And they score every crop and mark against its target, which they call a grounding-aware process reward.

The aim score comes from an AI judge. It looks at each box or point and gives it a number from 0 to 1. Both setups start at 0.55.

With answers-only reward, the aim score wandered between 0.57 and 0.66 and ended at 0.63. With the aim scored, it reached 0.81 by step 200.

Final accuracy followed. Scoring the aim added 2.0 points on RealWorldQA, 3.0 on BLINK, 5.1 on TreeBench and 8.4 on RealWorldCounting. Factauthors' ablation, Table 8.

No outside team has checked the aim score. InterpretationGrading each step looks like the central idea, but the evidence is the authors' own.

The payoff

So where does 71.6% land?

The full pipeline tops all ten rows in the authors' tool-loop comparison. Gemini-3-Flash is next at 66.4%.

The honest size of the jump

The authors' base model with tools scores 55.9%. With no tool at all, that same model scored 62.5% on the same seven tests (my average of Table 1).

Measured from there, the gain is 9.1 points, not 15.7. InterpretationThe two runs use different settings, so treat 9.1 as a rough guide.

A real gain, then. Just smaller than the headline.

How strong is it?

Promising, unchecked, and not yet runnable by anyone else.

• FactThe baseline is the authors' own base model. It scored lower with tools than without, so part of the headline is a rebound.
• FactThe rival models ran in the authors' own tool loop with the authors' prompt. These are their runs, not public leaderboard entries.
• FactThe trained model is not best everywhere. Their Level 1 model beats the final one on MMStar (82.6 vs 79.1) and ERQA (62.8 vs 62.2).
• FactThe authors state that none of the seven test sets was used to build training data. I found no separate contamination check. Treat it as their claim.
• FactThe training questions and the aim score both rely on AI judges, with a manual review for the questions.
• FactNo code or weights are linked on the arXiv page or in the PDF. No replication exists.

Sources: paper Tables 1, 2 and 3, Sections 3.2.2, 3.3.2 and 4.1, v1. arXiv abstract page checked 3 Nov 2026.

Where it fits

One more table shows how far the starting line has moved.

On a test called V*, earlier tool-using models scored between 81.8 and 85.6. The authors' untrained base already scores 90.6.

The lineage, in plain words

InterpretationDeepEyes (May 2025) taught a model to crop and zoom using reinforcement learning. Perception-R1 (June 2025) argued that standard reinforcement learning does not improve seeing, so seeing needs its own reward. CapRL (September 2025) rewarded captions by whether a text-only solver could answer from them.

PerceptRL changes two things. It grades seeing with quiz answers built from the model's own mistakes. And it grades each tool step by where it pointed.

The authors name video, a jointly trained pointing model and screen-control agents as next steps. My read: this line of work is moving from grading answers to grading steps. One unreplicated paper cannot confirm that.

Beside it

Five other papers sit nearby. Only one is a close neighbour.

PACE, 15 Oct 2026. It spreads credit across the words of one answer in multimodal reasoning training. PerceptRL spreads credit across the steps of a tool session.

InterpretationSame question (who earns the reward?) at a different level. PerceptRL's reference list does not cite it.

• SCARL, 29 Oct 2026. Text-only training. It reports +2.4 points average Pass@32 over DAPO on Qwen3-4B-Base. PerceptRL also builds on DAPO. Nothing about vision.
• SAFE-PRO, 15 Oct 2026. Makes multimodal models ask a clarifying question when a harmless-looking image and text could hide a risk. About safety, not seeing accuracy.
• ThinkingGuard, 29 Sep 2026. A guard model for hazards hidden in image and text combinations, with a 5,600-example test set. About safety.
• DCRL, 23 Sep 2026. Tackles reward hacking in text reasoning. Claims a 4B model beats a 32B baseline. Shares the worry about gamed rewards, not the topic.

Based on the five abstracts only. I did not read the full texts. PerceptRL itself makes no safety claim, so the "alignment and safety" label from AI Daily is loose. Links are under Go further.

What it would take

For this to matter outside the paper, someone else has to get the same result.

What would confirm it

• Code and weights released, so others can run it.
• An independent team gets about the same seven scores in the same tool loop.
• Humans agree with the AI judge about which crops and marks hit their target.

What would sink it

• The gain shrinks to nothing against a fair no-tool baseline.
• The aim score rises while human-judged aim does not.
• Signs that the test sets leaked into training.

SpeculationThree authors are at Qwen and the base model is a Qwen model. Whether this reaches a product is unknown.

Watch: the authors' code and model pages, a peer-reviewed version, and independent write-ups. None existed on 3 Nov 2026.

The takeaway: a tool is only as good as its aim. Ask who checks the aim.

Go further

Read the sources.

• The paper, arXiv:2610.37471 and its PDF. Code and weights: none released.
• Tests: BLINK, CountQA, TreeBench, MMStar.
• Earlier work: DeepEyes, DeepEyesV2, Perception-R1, CapRL.
• Beside it: PACE, SCARL, SAFE-PRO, ThinkingGuard, DCRL.
• More from us: vizmaya stories and AI Daily.

Sources and claims log

How this was checked.

24 claims were logged. 22 are verified against the paper's tables or the arXiv and OpenAlex pages. 2 are unverified: the authors' statement that the tests were not in training (shown only as their claim) and AI Daily's compute figure (kept off the page).

One AI Daily detail was contested and corrected. It listed other universities, while the paper lists Shanghai Jiao Tong University and Qwen. The paper wins.

Could not find: the paper's appendix (its references show unresolved "Appendix ??" in v1), training compute, any outside check of the aim judge, and any contamination test. Photographs: the image sources were unreachable, so every visual here is an original drawing.

Paper: arXiv:2610.37471, PerceptRL: Making Visual Perception Rewardable in MLLMs from Answers to Interactions (Guan, Yang et al.), arXiv preprint, 2026-10-27, https://arxiv.org/abs/2610.37471. Picked from AI Daily's papers (https://vizmaya.fyi/ai-daily/doom-v-boom). Citation counts: OpenAlex (CC0).

Numbers come from the paper (v1) by table, plus OpenAlex W7224119747, read 3 Nov 2026. The 9.1 point figure is my arithmetic from Table 1. Charts are redrawn, not pasted. Byline: vizmaya desk, 3 November 2026.