Trending ML Papers

What it is. This paper introduces Eywa, a system that lets a language model team up with specialized scientific models — for example, models trained on protein structures, weather data, or financial time series. Instead of forcing every problem to be described in words, the language model acts as a coordinator that delegates the messy domain-specific data directly to the specialist model best suited to handle it. The authors show this hybrid setup gets better answers on physics, biology, and social-science tasks while using fewer tokens and running faster than language-only agent pipelines.

Where it fits. Most agent frameworks today assume every tool, expert, and intermediate result is communicated in natural language. That works well for general reasoning but breaks down when the problem is fundamentally numerical or structural — like predicting how a drug molecule will fold or how a power grid will respond to load. Companies have spent the last few years building large 'foundation models' for these specific domains, but they live in silos and cannot easily plug into the LLM-driven agent stack that is taking over knowledge work.

Why it matters. If you want AI to do real scientific or engineering work — drug discovery, materials design, climate forecasting — you cannot get there by having an LLM read papers and guess. This work lays out a practical recipe for connecting a 'reasoning' LLM to specialist predictive models so they actually cooperate, with reported gains of roughly seven percentage points in task quality and around thirty percent fewer tokens. For founders building vertical AI products in regulated or technical domains, this is the kind of plumbing that determines whether an agent product is a demo or something a scientist will actually use.

What it is. When several AI agents work together on a problem, they usually pass full English (or code) messages back and forth — which is slow and lossy. RecursiveMAS lets agents instead pass their internal 'thoughts' (the model's hidden numerical state) directly to one another, in a loop that can be repeated as many times as needed. The authors report 8.3% better accuracy on average across math, science, medicine, search, and code tasks, while running 1.2 to 2.4 times faster and using up to 75% fewer tokens than text-based multi-agent setups.

Where it fits. Multi-agent setups — where one model plans, another searches, another writes code — are now a standard pattern for tackling hard problems with LLMs. The main complaint is that they are expensive: each agent has to write out its full reasoning so the next can read it, and quality often degrades as messages get summarized or truncated. There is also a parallel research thread on 'looped' or 'recursive' single models that re-run themselves to think harder. This paper combines those two ideas: treat the whole agent team as one big looped computation.

Why it matters. Agent products are bumping into a real cost wall — the same task can cost cents or dollars depending on how many round-trips it takes. If this technique holds up outside the benchmarks, it offers a way to keep multi-agent quality while cutting most of the chat overhead, which directly translates into cheaper inference and faster response times for end users. The caveat: it requires modifying how models are connected internally, so it is more of an infrastructure shift than a drop-in upgrade for existing API-based agent stacks.

What it is. Today's text-to-video models can produce stunning short clips, but watch carefully and objects melt, walls drift, and the world stops behaving like a physical place when the camera moves. World-R1, from Microsoft Research and Zhejiang University, fine-tunes existing video models with reinforcement learning — meaning the model gets a reward signal that is high when generated frames are geometrically consistent and low when they contradict 3D reality. Crucially, it does this without changing the model's architecture or needing a giant new dataset of 3D-labeled videos.

Where it fits. Generative video has gone from gimmick to plausible product in about two years, with systems like Sora setting expectations. The next frontier everyone is racing toward is video that can serve as a 'world model' — a believable, navigable simulation that an agent or game could actually live inside. Earlier attempts to add 3D awareness either bolted on extra modules (slow and expensive) or required hard-to-collect 3D training data. Reinforcement learning with model-based rewards has worked well for language reasoning recently, and this paper imports that recipe into video.

Why it matters. If video models can be cheaply nudged into respecting physics and 3D space, the use cases broaden from pretty clips to things like driving simulators, robotics training environments, and game asset pipelines. The reported jumps in geometric consistency are large (around 8–10 dB on a standard image-quality metric) without sacrificing visual quality, which is the trade-off that has stalled prior work. It is still early — they evaluate primarily on synthetic and short-horizon scenes — but the recipe of 'fine-tune any existing video model with a 3D-aware reward' is something competitors will adopt fast.

What it is. Different research communities mean very different things when they say 'world model' — to a robotics person it is a physics simulator the robot uses to plan; to a video researcher it is a generator that produces coherent footage; to a language-agent builder it is a model of how a website or codebase will behave. This 42-author survey from HKUST, NUS, Oxford and others proposes a shared vocabulary: three capability levels (predictor, simulator, evolver) crossed with four 'law regimes' (physical, digital, social, scientific). The authors then sort over 400 existing papers into this grid to show where the field is strong and where it is empty.

Where it fits. World models are arguably the most-hyped concept in AI right now — Yann LeCun has been preaching about them for years, every video-gen launch claims to be one, and frontier labs talk about them as the path to agents that can actually plan in the real world. The problem is that the term is so overloaded that two papers can both claim it while doing entirely incompatible things, which makes it nearly impossible for outsiders to track real progress. Surveys like this one show up when a field has grown to the point where confusion is itself slowing things down.

Why it matters. If you are a PM or founder trying to figure out what 'world model' actually means in a vendor pitch or an investor deck, this is the closest thing to a reference document the field has. The authors' framework also pinpoints concrete gaps — for instance, models that can revise their own assumptions when reality contradicts them ('L3 Evolvers') barely exist yet, but they are exactly what is needed for autonomous agents that operate in changing environments. Read it as a map of where the next year of capability progress is likely to come from, not as a benchmark result.

What it is. Image generators look great on first glance but quietly fail at things like 'keep the same character across five panels,' 'render text without scrambling letters,' or 'show what happens if I push that ball.' This roadmap, jointly authored by researchers at Tsinghua, NTU, HKU, NUS, Waterloo, Baidu and others, argues the field has been measuring the wrong thing — perceptual prettiness — and proposes a five-rung ladder from basic 'atomic' image generation up to systems that act like world simulators. They then run frontier models through stress tests (jigsaw puzzles, metro maps, fluid dynamics) to show where each rung breaks.

Where it fits. Image and video generation has had a 'wow' year, with frontier systems like Nano Banana, GPT-Image, Qwen-Image, and Z-Image flooding social feeds with photorealistic results. But product teams keep hitting the same walls: hands and text, multi-step editing that drifts, and any task that requires actually reasoning about objects rather than texturing them. This survey is the field's attempt to step back and articulate why those weaknesses are not random bugs but predictable consequences of treating generation as a one-shot rendering problem instead of a thinking-and-then-rendering loop.

Why it matters. For anyone shipping a product on top of an image or video model, the practical takeaway is that the public benchmarks vendors quote (FID scores, prompt-following rates) systematically overstate how good these systems are at the things real users care about. The paper's stress-test methodology — including 'restore to original' and multi-panel consistency tests — is something product teams can adapt directly into internal evals. The bigger bet here is that the next generation of image and video models will look less like better Photoshop and more like agents that can plan, verify, and re-render, which would change what is buildable on top of them.