This is not a roundup of vague trend pieces. Every paper covered here dropped on arXiv in the last 72 hours. Every one has working code, reproducible benchmarks, and explicit failure modes. None of them claim AGI. All of them move the baseline for what multimodal systems can actually do in production.

We are past the stage where people argue if multimodal LLMs work. We are now in the stage where people are fixing the specific, boring, critical flaws that stopped anyone from using them for real work.

PercepCap: Stop guessing why video captions are wrong

The single worst problem with video captioning today is that you cannot debug bad output. When a model writes "the dog ran left" but the dog actually ran right, you have no idea if it mislocated the dog, misidentified the direction, misordered the frames, or just hallucinated the entire action. All you get is the wrong sentence.

PercepCap fixes this by breaking generation into two explicit stages. First the model outputs a structured perception trace: bounding boxes for every object, frame timestamps for every event, trajectory lines for movement. Only after this trace is complete does it generate the human readable caption.

Training uses a two stage pipeline. First supervised fine tuning teaches the model to produce the trace before the caption. Then reinforcement learning applies joint reward over both the trace accuracy and caption quality. Critically, they did not hand label perception traces. They built training data by taking existing captions, extracting the mentioned objects and events, then grounding them back into the source video automatically.

Against Qwen3-VL baseline, PercepCap improves caption BLEU-4 by 12.7% and caption QA accuracy by 18.2%. More importantly, 79% of caption errors now correspond directly to a verifiable error in the intermediate perception trace. You do not have to guess what broke. You can look.

Remote sensing: Specialized models are losing

For the last two years every lab has been building fine tuned multimodal models for remote sensing. Everyone assumed general purpose vision models would never work for overhead imagery. That assumption was wrong.

This paper ran head to head benchmarks across 11 remote sensing tasks, pitting 7 specialized RS-MLLMs against 4 general purpose CV-MLLMs with zero domain fine tuning. Results are below:

Task	Best RS-MLLM Accuracy	Best General MLLM Accuracy
Visual Grounding	82.1%	71.4%
High Res VQA	76.3%	74.8%
Scene Classification	68.9%	77.2%
Object Counting	51.7%	62.4%
Change Detection	47.2%	59.1%

Specialized models only hold an edge on exactly two tasks. On every other task, unmodified general purpose models win. Often by wide margins.

This is not an argument against domain adaptation. It is an argument against building entire separate models from scratch. General multimodal models already learn almost all the visual reasoning you need. Fine tuning the last 1% is useful. Rebuilding the other 99% is wasteful.

All current models still fail badly at relational reasoning. None of them can reliably answer "how many fields are north of the road". This is the universal bottleneck right now, for every model, general or specialized.

Spoken reasoning does not need full text traces

Spoken language models are terrible at math. Not because they cannot hear the question. Because they cannot do working memory reasoning while staying in speech space.

Previous attempts to fix this used Chain of Modality: convert speech to full text, run text reasoning, then convert back to speech. This works, but it adds 2.5x token overhead and 1.8x inference latency. No one ships this.

ECoM reasoning solves this by compressing the intermediate reasoning trace. Instead of generating full natural language steps, the model produces a compact token sequence that acts both as reasoning state and speech alignment guide. Training uses progressive compression: start with full text traces, then gradually shorten them over training steps until only the required minimal state remains.

ECoM beats full CoM by 3% absolute accuracy while using only 40% of the intermediate tokens. This is not a tradeoff. It is strictly better on every metric.

This result generalizes. We have been wasting enormous amounts of compute generating human readable intermediate reasoning traces that no human will ever read. Models do not need natural language to reason. They just need state.

MV-Bench: MLLMs cannot build interfaces

Everyone is demoing MLLMs that turn screenshots into React code. All those demos are lying.

MV-Bench is the first proper benchmark for this task. It uses real Tableau workbooks as ground truth, testing not just visual layout, but correct data binding, cross filter coordination, and interactive behaviour. They tested GPT-4o, Gemini Advanced, Claude 3 Opus, Qwen3-VL and Llama 3.2 Vision.

Metric	Top Model Score
Visual layout reproduction	75.5%
Correct data binding	21.7%
Working cross view interactions	11.7%

Every model can draw the boxes correctly. Almost none of them connect those boxes to the right data. Almost none make the filters actually work.

Iterative refinement fixes code that crashes. It does not fix missing logic. The model does not forget to add interactions. It does not understand that interactions are supposed to exist at all.

This is the largest unacknowledged gap in multimodal capabilities right now. Models see pixels. They do not see intent.

OSVE: Real time video editing is here

Text guided video editing worked. It just took 12 minutes per 10 second clip. No one used it.

OSVE adapts one step diffusion models to video. There are three core changes:

1. A single pass encoder that predicts initial noise for all frames instead of running iterative inversion
2. Cross frame attention run once across all latents before generation
3. Sliding window anchor framing for sequences longer than 16 frames

The result is 155-171x speedup over state of the art multi step methods. Quality is equal or better. A 10 second edit runs in 0.4 seconds.

This is not incremental. This crosses the threshold from batch processing to interactive real time. You will see this shipped in consumer editing tools before the end of the year.

MTVDiff: Thermal to visible face translation works

Thermal cameras work in complete darkness. But you cannot run standard face recognition on thermal faces. Until now.

MTVDiff adds depth maps and text priors to latent diffusion for cross spectral translation. On the MCXFace dataset it reduces FID by 48.3% over previous best methods, and improves rank 1 face verification accuracy by 8.9%.

This is not a toy. This works well enough that you can run unmodified standard visible face recognition models directly on the translated output. There are very obvious and very serious privacy implications here. No one involved in the paper mentions them.

The unifying pattern across all six papers

None of these papers make models bigger. None of them add more parameters. None of them claim emergent capabilities.

Every single advance this week came from adding structure. Explicit intermediate representations. Compressed state. Separate stages for perception and description. Proper ground truth benchmarks.

For two years everyone was scaling parameters. Now everyone is fixing architecture. This is where the actual progress is happening.

What comes next

We will see a lot more explicit intermediate traces. Black box end to end models are dead for production use cases. You need to be able to inspect, debug and validate what the model saw before it generated output.

General purpose models will eat almost all specialized domain models over the next 12 months. The only remaining advantage for specialized models will be on tasks that require domain specific measurement, not domain specific vision.

Most importantly: we are now at the point where almost every week there is a new result that is good enough to ship immediately. The research lag to production time has collapsed from years to months. If you are waiting for this technology to mature, you have already missed the start.

---

References

1. PercepCap: Video Captioner with Structured Spatio-Temporal Perception http://arxiv.org/abs/2607.20389v1
2. Multimodal Large Language Models for Remote Sensing Image Understanding http://arxiv.org/abs/2607.20284v1
3. Efficient Chain-of-Modality Reasoning via Progressive Compression http://arxiv.org/abs/2607.19932v1
4. MV-Bench: Benchmarking Multimodal LLMs for Multi-View Interface Construction http://arxiv.org/abs/2607.19910v1
5. OSVE: One Step Video Editing with One Step Diffusion Models http://arxiv.org/abs/2607.19895v1
6. MTVDiff: Multimodal Conditional Latent Diffusion for Thermal-to-Visible Face Translation http://arxiv.org/abs/2607.19886v1