Researchers report that small vision AI misses tiny details because it looks in the wrong place

The puzzle

Can you find the kite?

Take a 4K photo of a beach. Somewhere in the sky is one small kite. You ask an AI: what colour is the kite?

The AI does not see a photo the way you do. It cuts the picture into small squares and reads them as a long list. For a 4K photo that is about 8,100 squares, called visual tokens. The kite covers only a handful.

Illustration of the idea. Each square is one visual token. The paper gives the 8,100 figure.

Now the AI starts to think. Newer models write out their reasoning for thousands of tokens before they answer. You would expect more thinking to mean better seeing.

So when the AI gets it wrong, which is it? Can it not see the kite? Or can it see the kite and fail to pick it out?

Test one

Hand it only the kite. Nothing gets worse.

The team took a small, open vision AI called Qwen3.5-4B. The photo is read as normal. Then, after reading, they deleted squares and let the AI answer from what was left.

Give the AI only the kite and it does as well as with the whole photo: 60.7% against 60.1%. Give it the same number of squares from the wrong place and it falls to 38.0%.

So the kite is not lost. The AI's reading of the photo still holds it. What is hard is picking it out of 8,100 squares.

A second test points the same way. Hiding the kite's box cost the AI 22.1 points (37.0% against 59.2% when an equal-size patch elsewhere was hidden). Paper Figure 2a.

The needle is in the stack. The AI can see it. So who is doing the picking?

Test two

Switch off four spotlights and 13 points vanish

Inside the AI are small units called attention heads. Picture each one as a spotlight. It decides which squares of the photo to shine on while the AI writes its answer.

The paper studies 128 of these spotlights. The team ranked them by how strongly each one lit up the right object on 96 separate practice questions. Then they switched off the top four.

Four spotlights off cost 13.2 points. Four random ones off cost 0.6. The paper calls the chosen four localization heads.

The twist

The longer the AI thinks, the worse it looks

Now the odd part. The team watched where the four spotlights pointed while the AI reasoned. Early on, 53% of their attention on the photo landed on the right object. In the last quarter of the reasoning it was 40%.

Illustration of the paper's Figure 2c. Each dot is 1% of the four heads' attention on the photo. The shares (53, 49, 46, 40) are the paper's; dot positions are drawn for the picture.

Part of the reason is plain crowding. As the AI writes more words, those words compete with the photo for attention (Appendix C.6).

The team also tried the obvious cheat: force more attention onto the photo. Accuracy stayed at 74.2%, against 74.1% (Table 8). More looking did not help. Looking in the right place was the point.

If the spotlights drift, can you train them to stay put?

The idea

Instead of a magnifying glass, train the spotlights

The usual fix is a magnifying glass. Crop around the object and look again. It works, but every crop is extra work for each question. One such method, ViCrop, adds two extra passes.

This team went inside the AI instead. They call the method Localization-Head Alignment, or LHA. Nothing changes when the AI is used. Only its training changes.

1. 1

   Find the spotlights

   Rank the 128 heads on 96 practice questions where the right object is boxed. Keep the top 10.

2. 2

   Practice on its own good answers

   Take 65 answers the AI already got right. Each time it states a fact, like "the kite is blue", reward the 10 spotlights for looking at the kite's box instead of an equal-size patch of background. 44 such moments, 32 small updates.

3. 3

   Keep practicing while it learns by trial and error

   In reinforcement learning the AI tries, and gets a point when its final answer is right. The same look-here reward stays on for the right tries. 50 steps.

4. 4

   Use it as normal

   No crops. No extra passes. The boxes are needed only in training.

The result

The authors report 74.1% rising to 76.6%

That is a pooled score over three tests of tiny details in big photos: V*, HR-Bench 4K and TreeBench. It was measured in one pass, with no crops. The authors say it is the best single-pass result among the 4B-size models they tested.

Two things stand out. Plain reinforcement learning without the look-here reward did nothing (GRPO 74.0, DAPO 73.6). And the AI the team started from, at 74.1, already beats every released model they compare against.

So when the abstract says LHA is ahead of released 7B models, most of that gap is the newer starting AI, not the new training. LHA beats TreeVGR-7B by 9.8 points. The untouched base already beats it by 7.3.

Look closer

The gain is lumpy, and one slice went down

A single number hides where the points came from. The paper's table splits each test into question types.

The biggest jumps are on questions that tie several objects together: +5.7 on TreeBench reasoning and +4.0 on HR-Bench questions about several objects. That fits the idea. If you must keep track of many things, looking in the right place matters more.

But on TreeBench perception questions, LHA scored 71.8 against 73.8 for the base. The authors' summary does not mention it.

The hype check

Is the fix real? Four of eight tests say yes

Each dot below is the gap in score between two runs that differ in one thing only. The bar is the range the authors say the true gap likely sits in (a 95% interval). If the bar crosses zero, the data cannot rule out no effect at all.

Aligned training beats normal training in four of five comparisons. The cleanest test of the new reward during trial-and-error training is +0.9 points, and in all three versions the bar crosses zero.

How strong is it?

The base AI was scored in one run, and two runs of the same AI can differ by about 1.1 points. Each trial-and-error setup was trained once, which the authors list as a limit. HR-Bench 4K is 800 of the 1,396 questions but only 200 different ones, shown in four orders. The authors removed 53 training questions whose images matched a test image, and I found no sign of leakage, but I cannot rule it out. I could not confirm that code or weights exist.

Beside it

Same week, same base model, different problem

The closest paper in AI Daily's pool is Weaken to Strengthen (arXiv 2610.38681, 28 October). It also improves a small Qwen3.5-4B, lifting one maths score from 57.7% to 70.5%.

Closer in idea are three vision papers this one names: EASE, Faithful-MR1 and AVAR. They also steer attention. On the authors' own runs, EASE scores 73.6% against 76.6% for LHA. The difference the authors point to is which heads get steered and where.

What would settle it

What it would take for this to matter

1. 29 Oct 2026Preprint postedClaims, tables and 23 pages of detail. No peer review yet.
2. NextCode and weightsI could not confirm either exists. Without them nobody can check the 76.6%.
3. ThenAn independent replicationSomeone else trains it from scratch, scores the base over several runs, and still sees about +2.
4. ThenOther models and tasksThe authors list larger models, other model families, charts, documents and video as untested.
5. LaterPeer review, then productsOnly after that would it show up in tools people use.

What to take with you

Trust the map more than the fix

The strongest card in this paper is the map. A few spotlights decide where a vision AI looks, and they drift as it thinks. That is a finding about how the machine works, and it changes how you read its mistakes.

The fix built on that map is smaller and shakier: +2.5 points, one run per setup, a control that cannot be told apart from zero.

Until someone else runs it, treat 76.6% as the authors' claim. Treat the four-spotlight result as the thing most worth checking next.

More explainers like this: Vizmaya stories and AI Daily.

Go further

• The paper, arXiv:2610.39649 (PDF)
• Code and weights: none confirmed.
• Benchmark: TreeBench (arXiv 2507.07999). V* and HR-Bench are cited in the paper's reference list.
• Related work: Weaken to Strengthen, EASE, Faithful-MR1, AVAR.

Credit, sources and claims log

Paper: arXiv:2610.39649, Look at Where You Think: Localization-Head Alignment for Fine-Grained Perception in Thinking VLMs (Wang, Wu et al.), arXiv preprint, 2026-10-29, https://arxiv.org/abs/2610.39649. Picked from AI Daily's papers (https://vizmaya.fyi/ai-daily/doom-v-boom). Citation counts: OpenAlex (CC0), read 2026-11-03.

Sources. The paper and its appendix (version 1). The Weaken to Strengthen abstract page. OpenAlex, read 3 Nov 2026. Photos as credited under each image: mikecogh (Flickr, CC BY-SA 2.0), Dasaptaerwin (Wikimedia Commons, CC0), Kamil Saitov (Wikimedia Commons, CC BY 4.0), Todd Chandler (Flickr, CC BY-SA 2.0).

Claims log: 21 claims checked. 18 verified as reports from the paper or a primary page. 0 contested. 3 unverified and kept off the page: the compute figure of 10^22 to 10^23 FLOP (the paper gives GPU-hours only), whether code or weights exist, and whether outlets have covered it (one search found none). Not found: any independent replication; public leaderboard checks for V*, HR-Bench and TreeBench.

Numbers are as the authors report them in an unreviewed preprint. AI Daily's summary fields were used only as leads and checked against the paper's tables. One lead was corrected: the headline gain is pooled over three benchmarks, not two.