X-Omni proves RL can make token-based image generators great again

 Diffusion may rule today’s text-to-image scene, but Tencent researchers just reminded everyone why discrete autoregressive models still matter. In a paper titled “X-Omni: Reinforcement Learning Makes Discrete Autoregressive Image Generative Models Great Again,” they show that a sprinkle of reward learning turns a 7 B LLM that predicts visual tokens into a Sora-class image engine—while natively sharing weights with language generation.

Three moving parts

Module	Job	RL impact
Semantic image tokenizer	Converts 32 × 32 patch features into a 65 k-token vocabulary without vector-quantization blur.	Supplies denser reward signals than pixel-level losses.
Unified AR backbone	One transformer handles both language and image tokens; no diffusion head during training.	After SFT it over-fits, but RL fixes fidelity & instruction following.
Offline diffusion decoder	A lightweight “decompressor” turns token grids into crisp 1 K-px frames.	Keeps inference < 2 s on a single A100.

Why reinforcement learning?

Supervised fine-tuning left the model with warped faces and garbled typography. Policy-gradient updates—rewarded for CLIP aesthetics, OCR accuracy and prompt adherence—steadily cleaned up artifacts and nailed complex layouts, something best-of-N sampling couldn’t match.

Early numbers worth noting

• FID 1.7 on ImageNet-256 (beating DiT-XL by 9 %)

• 99.2 % prompt compliance on the new LongText-Bench (Chinese + English captions up to 120 chars)

• 3.5× faster than diffusion baselines at 1024 × 1024 when streaming tokens with Flash-Attn 3.0

• < 8.5 GB VRAM for a distilled 1.3 B variant (coming soon, according to the repo)

Why it matters

1. Unified model, unified budget – No separate diffusion tower; language and image share the same 7 B weights, making deployment simpler and cheaper.

2. Long-text rendering solved – Posters, UI mock-ups and meme creators finally get reliable lettering without kludgy diffusion guidance.

3. Open everything – Code, checkpoints and the 200-prompt LongText-Bench live on GitHub under Apache-2.0. Fine-tune away.

The bigger picture

Until now, researchers had mostly written off discrete AR image models as artifacts-prone hold-overs from DALL·E 1. X-Omni flips that narrative: with the right reward design, token predictors can match (and in text rendering, beat) diffusion’s photorealism while keeping the door open for seamless language–vision fusion and future any-to-any generation. Expect a resurgence of AR tokenizers, LoRA packs for brand fonts, and perhaps a new front in the multimodal model wars.

Paper link: arXiv 2507.22058 (PDF)

From Tedious Edits to Autonomous IDEs: How Kiro’s AI Agent Hooks Turbo-Charge Your Dev Workflow

 The modern codebase is a living organism: files mutate, requirements shift, tests trail behind and docs go stale. Kiro—the Amazon-backed, Claude-powered IDE—thinks the fix is automation that lives inside your editor. On July 16 2025 the team introduced Agent Hooks, a rules-engine plus AI copilot that fires the moment you hit “save” or merge a pull request.

What exactly is an Agent Hook?

Each hook couples a trigger (file edit, creation, deletion, or even a manual slash-command) with an AI action such as “update the related unit tests” or “refresh my README” . Unlike brittle shell scripts, the action is described in plain English and executed by a Gemini-class agent that understands project context. The result feels less like CI glue and more like a junior dev who never sleeps.

Five headline benefits

1. Natural-language config – type “Whenever I touch *.py, update the matching test_*.py” and the hook YAML writes itself.

2. Context-aware reasoning – the agent sees your entire workspace, so it can refactor imports or respect custom test frameworks.

3. Real-time execution – actions run instantly, keeping flow intact instead of kicking chores to a nightly job.

4. Shareable recipes – hook files live in .kiro/hooks, so teams version them like code and inherit automation on git pull.

5. Stack-agnostic events – docs list triggers for save, create, delete, plus a user-initiated option for ad-hoc tasks.

Building your first hook in three clicks

Open Kiro’s sidebar, hit “Agent Hooks ➕”, and either select a template or just describe what you need. The UI scaffolds a config you can fine-tune—patterns, prompt, and whether it auto-runs or waits for manual confirmation. Behind the scenes, Kiro writes a .kiro.hook file so you’re always one git diff away from auditing the logic.

Real-world recipes

• Test synchroniser – Every Python edit triggers the agent to inspect changes and regenerate the paired test module, ensuring 100 % coverage drifts aren’t ignored.

• Doc updater – Modify a public API and the hook patches your Markdown docs so onboarding guides never lag behind shipping code.

• Git concierge – On commit, a hook can draft a concise changelog entry and polish the commit message to match team conventions.

• I18N helper – Save a UI string file and watch the agent push auto-translations to language packs.

Best-practice tips

Start small—a single file pattern and a succinct prompt—then iterate by reading the hook execution history shown in Kiro’s chat pane. Give the agent richer guidance (“follow Google Python Style”) and reference project docs inside the prompt for tighter alignment. Finally, commit hooks so teammates inherit them; over time your repo becomes a cookbook of living automation rules the whole squad benefits from.

Why this matters

Developers already rely on AI for autocomplete and chat, but those tools are reactive—you ask, they answer. Agent Hooks flip the script to proactive assistance that runs without explicit prompts, erasing the cognitive tax of context switching. In a world of sprawling microservices and relentless release cadences, the ability to delegate routine upkeep to an always-on agent is a genuine force multiplier.

Kiro doesn’t claim to replace developers; it aims to amplify craftsmanship by letting humans stay in the creative loop while machines patrol the trenches. If your backlog is clogged with “fix tests” and “update docs” tickets, Agent Hooks might be the invisible intern you’ve been wishing for. Install Kiro, write your first hook, and watch housekeeping melt away—one automated trigger at a time.

AlphaEarth Foundations: Google DeepMind’s “Virtual Satellite” Sets a New Baseline for Planet-Scale Mapping

 

A virtual satellite built from data

On July 30 2025, Google DeepMind unwrapped AlphaEarth Foundations, an AI model that ingests optical, radar, lidar and climate-simulation feeds and distills them into a single 64-dimensional “embedding field” for every 10 × 10 meter patch of terrestrial land and coastal waters. Think of it as a software satellite constellation: instead of waiting for the next orbital pass, analysts query a unified representation that already encodes land cover, surface materials and temporal change. 

How it works

AlphaEarth tackles two long-standing headaches—data overload and inconsistency. First, it merges dozens of public observation streams, weaving them into time-aligned “video” frames of the planet. Second, it compresses those frames 16× more efficiently than previous AI pipelines, slashing storage and compute for downstream tasks. Each embedding becomes a compact, loss-aware summary that models can reason over without re-processing raw pixels. 

A leap in accuracy and efficiency

In head-to-head evaluations spanning land-use, surface-property and seasonal-change tasks, AlphaEarth posted a 24 % lower error rate than both classical remote-sensing methods and recent deep-learning baselines. Crucially, it excelled when label data was sparse—proof that its self-supervised pre-training truly generalises. The accompanying research paper on arXiv highlights consistent out-performance across “diverse mapping evaluations” without fine-tuning. 

From blog post to real-world maps

To jump-start adoption, DeepMind and Google Earth Engine released the Satellite Embedding dataset: annual global snapshots containing 1.4 trillion embedding footprints per year. More than 50 organisations—including the UN’s Food and Agriculture Organisation, MapBiomas, the Global Ecosystems Atlas and Stanford University—are already piloting projects that range from rainforest monitoring to precision agriculture. Users report faster map production and higher classification accuracy, even in cloudy tropics or sparsely imaged polar regions. 

Why it matters for climate and beyond

Accurate, up-to-date geospatial data underpins decisions on food security, infrastructure and conservation. Yet researchers often juggle incompatible satellite products or wrestle with GPU-hungry vision models. AlphaEarth shrinks that friction: a single API call retrieves embeddings that are both information-dense and provenance-rich, ready for plug-and-play into GIS tools, LLM agents or custom model fine-tunes. Cheaper storage and lower latency also mean national agencies with modest budgets can now run continent-scale analyses weekly instead of yearly. 

The road ahead

DeepMind hints at extending the framework to real-time streams and coupling it with Gemini-class reasoning agents capable of answering open-ended “why” and “what-if” questions about Earth systems. For AI builders, the combination of long-context language models and AlphaEarth embeddings could enable chatbots that diagnose crop stress or forecast urban heat islands—all grounded in verifiable pixels.

Bottom line: AlphaEarth Foundations compresses the planet into a query-ready lattice of vectors, handing scientists, policymakers and hobbyist mappers a new lens on Earth’s shifting surface. With open data, documented gains and an Apache-style license, DeepMind has effectively democratized a planetary observatory—one 10-meter square at a time.

“Everyone’s AI”: MiniMax CEO Junjie Yan Reimagines the AI Economy at WAIC 2025

 The opening morning of the World Artificial Intelligence Conference 2025 (WAIC) in Shanghai was buzzing with hardware demos and multimodal avatars, yet the moment that set the tone for the three-day summit was a keynote titled “Everyone’s AI.” Delivered by MiniMax founder & CEO Junjie Yan, the talk argued that artificial intelligence is no longer a sidecar to the internet economy—it is becoming the primary productive force. 

From research toy to societal engine

Yan traced a 15-year personal journey in AI research, noting that tasks once handled by junior engineers—code writing, data annotation, even literature review—are now 70 % automated inside MiniMax. The implication is stark: as models grow more capable, human attention shifts from mechanical chores to creative orchestration. “AI can now write the software that analyzes the data we used to comb through by hand,” he observed, positioning large models as multipliers of both knowledge work and imagination. 

The economics: another 10× drop on the horizon

MiniMax isn’t just waxing philosophical; it is betting on cost curves. Yan predicted that inference prices for top-tier models will fall another order of magnitude within two years, echoing the steep declines seen in 2024–25. Cheaper inference, he argued, is the real catalyst for mass adoption—unlocking agentic workflows that might consume millions of tokens per session without breaking budgets. 

Many models, many values

Contrary to fears of an AI monoculture, Yan expects plurality to define the market. Alignment targets diverge—one model may optimize for programming accuracy, another for empathetic conversation—so “there will definitely be multiple players,” he insisted. Open-source ecosystems, now approaching closed-source performance, reinforce that trend. 

Multi-agent systems change the rules

Inside MiniMax’s own products—Conch AI for voice, M-Series for reasoning, and the new MiniMax-M1 hybrid model—multi-agent architectures are displacing single-model pipelines. In such systems, the marginal advantage of any one model shrinks, while orchestration and tool-use matter more. That, Yan believes, will democratize expertise: startups armed with well-designed agent swarms can challenge giants who merely scale parameters. 

A less money-burning industry

Dropping costs and smarter experiment design mean AI R&D need not be an endless bonfire of GPUs. MiniMax’s internal stats show 90 % of routine data analysis already handled by AI, freeing researchers to pursue “genius ideas” that compound returns faster than raw compute. If training becomes less capital-intensive and inference goes bargain-basement, the barriers to entry for niche models and vertical agents collapse. 

“Everyone’s AI” as call to action

Yan closed by reframing access as both economic necessity and moral imperative: AGI, when achieved, should belong to multiple companies and a broad user base—not a solitary gatekeeper. He tied the mission to a Chinese proverb about unleashing creativity: lower thresholds ignite countless sparks. For a conference that also featured Geoffrey Hinton warning about rogue super-intelligence, MiniMax’s pitch provided a complementary optimism grounded in unit economics and open ecosystems.

Why it matters

The keynote crystallizes a broader shift in 2025: value is migrating from parameter counts to deployment fluency, from cloud monopolies to community forks, and from eye-watering API bills to near-frictionless inference. If Yan’s forecast holds, the next two years could see AI agents embedded in every workflow—powered by models cheap enough to run continuously and diverse enough to reflect local values. In that future, “Everyone’s AI” is not a slogan; it is table stakes.

LangExtract: Google’s Gemini-Powered Library That Turns Raw Text into Reliable Data

 

A new way to mine insight from messy text

On July 30 2025 the Google Developers Blog unveiled LangExtract, an open-source Python package that promises to “unlock the data within” any text-heavy corpus, from clinical notes to customer feedback threads. Built around Gemini models but compatible with any LLM, the project aims to replace brittle regex pipelines with a single declarative interface for extraction, visualization and traceability. 

Why LangExtract stands out

LangExtract combines seven features that rarely appear together in one tool:

1. Precise source grounding – every entity you pull out is linked back to its exact character span in the original document, so auditors can see where a value came from.

2. Schema-enforced outputs – you describe the JSON you want, add a few examples, and the library leverages Gemini’s controlled generation to keep responses on-spec.

3. Long-context optimisation – chunking, parallel passes and multi-stage recall tame “needle-in-a-haystack” searches across million-token inputs.

4. Interactive HTML visualisation – one command turns results into a self-contained page where extractions glow inside the source text.

5. Flexible back-ends – swap Gemini for on-device Ollama models or any OpenAI-compatible endpoint.

6. Domain agnosticism – the same prompt-plus-examples recipe works for finance, law, medicine or literature.

7. Apache-2.0 licence – no gating, just pip install langextract. 

How it works in practice

A “quick-start” script pulls Shakespeare characters, emotions and relationships in about a dozen lines of code, then writes an interactive HTML overlay showing each extraction highlighted inside the play. The same pattern scales: push the full Romeo and Juliet text through three extraction passes and LangExtract surfaces hundreds of grounded entities while keeping recall high. G

The GitHub repository already counts 200+ stars less than a week after launch, and ships with examples for medication extraction and structured radiology reporting—fields where provenance and accuracy are critical. A live Hugging Face demo called RadExtract shows the library converting free-text X-ray reports into structured findings, then color-coding the original sentences that justify each data point. 

Under the hood: Gemini plus controlled generation

When you pass model_id="gemini-2.5-flash" (or -pro for harder tasks), LangExtract automatically applies Google’s controlled generation API to lock output into the schema you defined. That means fewer JSON-parse errors and cleaner downstream pipelines—something traditional LLM calls often fumble. For massive workloads, Google recommends a Tier-2 Gemini quota to avoid rate limits. 

Why developers should pay attention

Information extraction has long oscillated between hand-tuned rules (fast but brittle) and heavyweight ML pipelines (accurate but slow to build). LangExtract offers a third path: prompt-programming simplicity with enterprise-grade traceability. Because it’s open-source, teams can audit the chain of custody and fine-tune prompts to their own compliance rules instead of black-box vendor filters.

Whether you’re structuring earnings calls, tagging sentiment in product reviews, or mapping drug-dosage relationships in EMRs, LangExtract turns unreadable text into queryable data—without sacrificing transparency. For AI enthusiasts, it’s also a practical showcase of what Gemini’s long-context and schema-control features can do today.

Bottom line: install the package, craft a clear prompt, add a few gold examples, and LangExtract will handle the rest—from parallel chunking to an HTML dashboard—so you can move straight from raw documents to actionable datasets.

ChatGLM’s GLM‑4 family levels up—and brings its own toolbox

 Tsinghua‑spun Zhipu AI has spent three years iterating on ChatGLM, a Chinese‑English rival to GPT. Its new report zooms in on the GLM‑4 series, a trio that stretches from a data‑center‑class behemoth to a 9 B‑parameter fine‑tune you can run at home. The headline: GLM‑4 “closely rivals or outperforms GPT‑4” on marquee leaderboards—while an All Tools variant autonomously fires up external apps to finish harder jobs. 

Under the hood

Piece	Why it matters
10 T‑token corpus (Chinese & English‑heavy, 24 other languages)	Gives the model near‑par bilingual parity—something GPT‑4 still chases in Chinese.
Multi‑stage alignment (SFT → RLHF)	Drives instruction following to GPT‑4‑Turbo levels on IFEval without bloating answers.
All Tools post‑training	Lets GLM‑4 decide if a prompt needs web search, Python, text‑to‑image, or any user‑defined API—no manual tool triggers.

The SKUs

• GLM‑4 – flagship ~130 B active params, 128 K context, up to 1 M with sparse attention.

• GLM‑4‑Air – latency‑trimmed 34 B variant tuned for GPU serving.

• GLM‑4‑9B / 9B‑Chat – consumer‑grade checkpoint (128 K / 1 M context) already live on Hugging Face.

Scorecard highlights

• General reasoning: beats or ties GPT‑4 on MMLU, GSM8K, MATH, BBH, GPQA, HumanEval. 

• Chinese alignment: tops GPT‑4 on AlignBench. 

• Long context: matches GPT‑4‑Turbo 128 K and Claude 3 at 256 K spill‑tests. 

• Tool use: in dev‑set trials, GLM‑4 All Tools edges GPT‑4 All Tools in web‑info retrieval and Python‑powered math. a

Why it matters

1. Bilingual crown – China finally has an open(-ish) model that doesn’t trade English chops for Mandarin mastery.

2. Tool autonomy – A single checkpoint that chooses whether to browse, code or draw marks a step toward plug‑and‑play agent workflows.

3. Open‑source momentum – Previous ChatGLM releases logged 10 M+ Hugging Face downloads in 2023; GLM‑4‑9B is expected to super‑charge that hobbyist wave. 

Rapid timeline of the GLM ecosystem

![timeline figure omitted] The paper’s timeline shows an 18‑month sprint from GLM‑130B to GLM‑4‑All Tools, with side quests into code (CodeGeeX), vision (GLM‑4V‑9B) and agents (AutoWebGLM). 

The road ahead

Zhipu AI hints at an MoE‑style GLM‑5 and deeper tool libraries (SQL, vector search, proprietary APIs). For builders already juggling browser calls, Python sandboxes and image pipes, GLM‑4 All Tools may offer a cleaner, unified brain—especially if your product needs to speak both English and Mandarin with equal poise.

Paper link: arXiv 2406.12793 (PDF)

ARC‑Hunyuan‑Video‑7B: Tencent’s bid to finally understand short videos

 Swipe through WeChat Channels or TikTok and you’ll find an AI nightmare: jump‑cuts, dense captions, meme audio and zero breathing room. Generic vision‑language behemoths struggle to keep up. ARC‑Hunyuan‑Video‑7B—unveiled this week by Tencent’s ARC Lab—takes direct aim at that chaos with an end‑to‑end, tri‑modal stack that ingests RGB frames, raw audio and ASR text before producing structured outputs.

What makes it different?

Design choice	Why it matters
Audio encoder fused with ViT backbone	Lets the model pinpoint punch‑lines, product mentions or sudden sound cues that pure‑vision systems miss.
Timestamp overlay on every frame	Gives the LLM hard temporal anchors for grounding, enabling second‑level captions and event logs.
Automated annotation pipeline with “millions” of real shorts	Avoids domain shift that plagues models trained on movie trailers or HowTo videos.
Five‑stage training (pre‑train → SFT → cold‑start → RLHF → SFT)	RL on objective tasks (e.g., exact‑match grounding) unlocks the subjective “explain the joke” style understanding.

---

Early numbers

• ShortVid‑Bench (new internal benchmark): authors report “strong performance” across captioning, QA, grounding and reasoning—outpacing prior open models, though exact deltas remain embargoed. 

• Latency: stress tests show 10 s end‑to‑end for a 60‑second clip on an H20 (≈A100‑class) GPU—fast enough to power feed ranking or real‑time moderation. 

---

Why builders should care

1. One stop for multi‑task video NLP – The same checkpoint handles highlight extraction, event logs, scene QA and clip‑level summaries, reducing pipeline sprawl.

2. Audio is first‑class – Brands, educators and accessibility teams can finally query both what’s shown and what’s said in user‑generated shorts.

3. Edge‑friendly – At 7 B parameters (≈8.6 B in BF16), it’s small enough for a single A100 or dual consumer GPUs under vLLM. 

4. Open weights & code – Hugging Face repo, training scripts and a vLLM deployment guide are already public, licensing‑friendly for commercial use. 

---

The bigger picture

OpenAI’s GPT‑4o Vision and Google’s Gemini 1.5 Pro handle long vids but lean on frame sampling and text prompts. ARC‑Hunyuan‑Video‑7B instead streams raw pixels + sound and returns a structured JSON‑style digest—closer to what feed‑ranking or search engines need. Tencent claims the model is already in production, lifting short‑video engagement metrics; if those gains hold, expect other platforms to pivot toward structured rather than free‑text video understanding.

Paper link: arXiv 2507.20939 (PDF)

Align Evals: LangSmith’s New Playground for Human‑Aligned LLM Evaluation

When LangChain announced Align Evals on July 29, 2025, it answered a pain point that has dogged almost every LLM team: evaluator scores that don’t line up with human judgment. The new feature—now live for all LangSmith Cloud users—lets builders calibrate their “LLM‑as‑a‑judge” prompts until automated scores track closely with what real reviewers would say. 

Why alignment matters in evaluation

Even the best prompt tweaks or model upgrades lose value if your test harness misfires. LangChain notes that teams waste time “chasing false signals” when evaluators over‑ or under‑score outputs versus human reviewers. Align Evals gives immediate feedback on that gap, quantifying it as an alignment score you can iterate against. 

A feature set built for rapid iteration

Align Evals drops a playground‑style interface into LangSmith with three marquee capabilities:

1. Real‑time alignment score for each evaluator prompt revision.

2. Side‑by‑side comparison of human‑graded “golden set” examples and LLM‑generated scores, sortable to surface the worst mismatches.

3. Baseline snapshots so you can track whether your latest prompt improved or regressed alignment. 

The alignment flow in four steps

LangChain distills evaluator creation into a structured loop:

1. Select evaluation criteria that reflect app priorities (e.g., correctness and conciseness for chatbots).

2. Curate representative data—good and bad outputs alike—to form a realistic test bed.

3. Assign human scores to create a gold standard.

4. Draft an evaluator prompt, run it against the set, and refine until its judgments mirror the human baseline. The UI highlights over‑scored or under‑scored cases so you know exactly what to fix next. 

Availability and roadmap

Align Evals is already shipping in LangSmith Cloud; a self‑hosted release drops later this week. Looking ahead, LangChain teases analytics for long‑term tracking and even automatic prompt optimization that will generate alternative evaluator prompts for you. 

Why AI builders should care

Evaluations are the backbone of continuous improvement—whether you’re evaluating a single prompt, a RAG pipeline, or a multi‑agent workflow. Yet teams often discover that a “99 % accurate” evaluator still lets bad outputs slip through. Align Evals lowers that friction, turning evaluator design into a measurable, repeatable process.

For AI enthusiasts and practitioners, the message is clear: before you chase bigger models or flashier agents, make sure your evaluators speak the same language as your users. With Align Evals, LangChain just handed the community a calibrated mic—and the feedback loop we’ve been missing.

PhyWorldBench asks: can your video model obey gravity?

 Text-to-video (T2V) generators can paint dazzling scenes, but do they respect momentum, energy conservation—or even keep objects from phasing through walls? PhyWorldBench says “not yet.” The new 31-page study introduces a physics-first benchmark that pits 12 state-of-the-art models (five proprietary, seven open source) against 1,050 carefully curated prompts spanning real and deliberately impossible scenarios. The verdict: even the best models fumble basic mechanics, with the proprietary Pika 2.0 topping its class at a modest 0.262 success rate, while Wanx-2.1 leads open source. 

A benchmark built like a physics textbook

Researchers defined 10 main physics categories, each split into 5 subcategories, then wrote 7 scenarios per subcategory—and for every scenario, three prompt styles (event, physics‑enhanced, detailed narrative). That’s how you get to 1,050 prompts without redundancy. 

Anti‑physics on purpose

One twist: an “Anti‑Physics” track where prompts violate real laws (e.g., objects accelerating upward). These gauge whether models blindly mimic training data or can intentionally break rules when asked. 

Cheap(er) scoring with an MLLM judge

Instead of hand‑labeling 12,600 generated videos, the team devised a yes/no metric using modern multimodal LLMs (GPT‑4o, Gemini‑1.5‑Pro) to check “basic” and “key” physics standards. Large human studies back its reliability, making large‑scale physics eval feasible. 

What tripped models up

• Temporal consistency & motion realism still break first.

• Higher‑complexity composites (rigid body collisions, fluids, human/animal motion) expose bigger gaps.

• Models often follow cinematic cues over physics, picking “cool” shots that contradict dynamics. 

Prompting matters (a lot)

Richer, physics‑aware prompts help—but only so much. The authors outline prompt‑crafting tips that nudge models toward lawful motion, yet many failures persist, hinting at architectural limits. 

Why this matters

• Reality is the next frontier. As T2V engines head for simulation, education and robotics, looking right isn’t enough—they must behave right. 

• Benchmarks drive progress. Prior suites (VBench, VideoPhy, PhyGenBench) touched pieces of the problem; PhyWorldBench widens coverage and difficulty, revealing headroom hidden by softer tests. 

• MLLM evaluators scale oversight. A simple, zero‑shot judge could generalize to other “lawfulness” checks—chemistry, finance, safety—without armies of annotators. 

The authors release all prompts, annotations and a leaderboard, inviting labs to iterate on physical correctness—not just prettier pixels. Until models stop dropping balls through floors, PhyWorldBench is likely to be the scoreboard everyone cites.

Paper link: arXiv 2507.13428 (PDF)