Trending ML Papers

What it is. AREX is an AI 'deep research' agent — the kind that goes off, browses the web, and returns a long, sourced answer — that keeps improving its own draft rather than just searching harder. It works in two nested loops: an inner loop that gathers evidence and writes a provisional answer, and an outer loop that checks each claim against the original question, then sends the agent back to hunt down whatever is still shaky. To keep this working over long sessions, it also learns to compress its own scratchpad so it does not drown in its own history.

Where it fits. Deep-research agents (think ChatGPT's or Perplexity's long-form research modes) have become one of the most visible product categories in AI, but they mostly try to be smarter by searching longer or spawning more sub-agents. The authors argue there is a simpler asymmetry to exploit: finding a good answer is hard, but checking whether a candidate answer satisfies each constraint is much easier — and that checking signal can be fed back to guide the next round of research. This puts AREX in the same lineage as recent 'self-refine' and verifier-guided agents, but pushed further into a true recursion where verification becomes the control loop, not a final filter.

Why it matters. For anyone building research copilots, market-intelligence tools, or agentic workflows, this is a template for making them more reliable without simply throwing more tokens at the problem. AREX beats bigger baselines on benchmarks like BrowseComp, GAIA, and Humanity's Last Exam using a 4B or 122B-with-10B-active model, hinting that thoughtful loop design can substitute for raw scale — which matters for cost. Caveat: benchmark wins on curated tasks are not the same as real users with messy questions, and 'confidence-based termination' is only as trustworthy as the model's self-assessment.

What it is. The authors argue that today's education-AI benchmarks only test whether a model can answer an exam question, not whether it actually understands the curriculum — things like what a student needs to learn first, how a lab experiment connects to a concept, or where in a textbook an idea appears. To fix this, they build a huge knowledge graph from China's official K–12 textbooks (math, physics, chemistry, biology), then spin it into both a 23,640-question benchmark (K12-Bench) and a training dataset (K12-Train) of text and image-based question-answer pairs. Even the strongest models tested get only 57% right on the benchmark, but fine-tuning on their curated data closes a lot of the gap.

Where it fits. AI tutoring is one of the most heavily hyped consumer-AI use cases, with billions in funding flowing into companies claiming to have built the next great study buddy. But most 'education LLM' evaluations look identical to trivia benchmarks — memorize facts, spit out answers — which is why a top model can ace a subject test and still be a bad teacher. Grounding evaluation in an actual curriculum graph (prerequisites, taxonomies, textbook location, diagrams) is a shift from testing recall to testing pedagogical structure, and it echoes broader moves toward domain-grounded evaluation in medicine and law.

Why it matters. For anyone building an AI tutor, this is a direct signal that off-the-shelf frontier models are still missing something concrete and fixable — and that a modest amount of well-structured supervised data beats generic instruction-tuning corpora orders of magnitude larger. It also gives edtech founders a public benchmark they can point to, which the field has been missing. The obvious caveat: it is anchored in the Chinese K–12 curriculum, so anyone building for other markets will need to build their own analogous resource — but the recipe (extract graph from official textbooks, turn graph traversals into questions) transfers cleanly.

What it is. ReferTrack teaches a robot to follow a specific person described in plain language (e.g., 'follow the man in the dark blue t-shirt') using only a single forward-facing camera. Instead of reasoning about the target as an abstract concept in the model's internal state, it first draws numbered boxes around every person it sees, picks the right box, and then plans where to walk — a much more grounded chain of thought. It also remembers the last few boxes of the target so it does not lose them in a crowd, and the authors demonstrate it running on real legged and humanoid robots.

Where it fits. Robotics is having a vision-language-action (VLA) moment: instead of hand-coding perception, tracking, and planning as separate modules, teams are training one big model that ingests camera pixels and instructions and outputs motion. That unification is elegant but has a real weakness — when the model reasons purely in its own latent space, it is hard to tell what it is 'looking at,' and grounding often fails in crowded scenes. Recent work like TrackVLA++ tried to bolt on explicit spatial reasoning, and ReferTrack pushes that further by forcing the reasoning to happen in terms of concrete image-space boxes.

Why it matters. This is a small but meaningful step toward robots that can be told what to do in natural language and actually get it right in a busy environment — a big deal for delivery robots, factory assistants, and companion robots aimed at hospitals or eldercare. The single-camera setup matters commercially because it keeps hardware cheap, and the sim-to-real transfer to legged and humanoid platforms suggests the approach is not brittle to the messy physics of the real world. It is still evaluated primarily on pedestrian-following, though, so open questions remain around occlusion, adversarial similar-looking distractors, and moving from following to more general instructions.

What it is. The paper shows a way for a vision-language model to train itself without needing a bigger 'teacher' model or extra human-provided hints. The trick: show the same model an image and then a version of that image with the content erased, look at how much the model's next-word predictions shift between the two, and treat that shift as a signal for which answers actually depend on seeing the picture. That signal is then used to sharpen the model's own outputs and train a stronger version of itself.

Where it fits. 'On-policy distillation' is one of the most active areas in post-training, because it is a cheaper alternative to reinforcement learning from human feedback: instead of a human judge or reward model, you have a stronger model 'teach' a weaker one by matching its outputs on the weaker model's own generations. The obvious next step is to remove the stronger teacher — that is 'self-distillation' — but naive self-distillation fizzles because the teacher and student know the same things. Prior work fixed this by feeding the teacher privileged information (the ground-truth answer, cropped hints, etc.); this paper removes that crutch and derives the extra signal purely from a controlled change in the input image.

Why it matters. For teams post-training their own multimodal models, this promises a meaningful jump in quality (roughly 3–5 accuracy points on aggregate benchmarks) with no external teacher, no hand-labeled hints, and no extra inference cost — which translates to real dollar savings and lower operational complexity. It also fits the broader industry trend of squeezing more out of existing models rather than always going bigger. Caveat: gains are reported on academic multimodal benchmarks, and it remains to be seen whether the same recipe helps on the messy, domain-specific vision tasks that companies actually ship — but the underlying idea (compare model behavior with and without the visual input) is general enough to be worth experimenting with.

What it is. When we test whether an AI understands space — where objects are, how to reach one, which way something is facing — we usually make it type out coordinates or pick a label, which is a bit like asking a person to describe a doorway by giving pixel numbers. This paper builds a way to let image-generation models (things like FLUX or Seedream) 'point' with the picture itself: they can mark a region, draw a path, or highlight an area, and the framework parses that back into scores comparable to what a text-based model would get. On top of the framework, the authors ship a 470-question benchmark that tests 14 different spatial skills.

Where it fits. There has been a growing suspicion in the vision community that image generators — the models we usually think of as 'the ones that make pretty pictures' — are quietly picking up genuine world knowledge and spatial understanding, since generating a plausible scene requires knowing where things go. But most benchmarks treat spatial reasoning as a language task (answer in text or coordinates), which biases the field toward text-based vision-language models and makes it impossible to fairly compare image-generation approaches. This work is part of a wave attempting to evaluate generative models as reasoners, alongside recent efforts to use video-generation models for perception and manipulation.

Why it matters. If image-generation models really are competitive spatial reasoners, that opens the door to using one model for both understanding and generating visuals — a big architectural simplification for products in design tools, robotics, AR/VR, and any workflow that mixes 'analyze this scene' with 'now edit or extend it.' The paper's headline finding is nuanced: image generators are strong when the answer is naturally visual, but text-output VLMs still win on tasks that need compositional reasoning ('the red thing that is to the left of the second-tallest object'). So the practical takeaway is that the two families of models are complementary — worth using the right tool for each spatial subproblem rather than assuming one will dominate.