AI FRONTIER
Signals worth understanding before they become obvious.
Proof-based general-purpose computing with Magnetar
Microsoft Research introduces Magnetar, a general-purpose zero-knowledge virtual machine (zkVM) that proves execution of RISC-V programs written in Rust, Go, or C, with security resting only on standard hash functions.
- Magnetar proves 7.2 to 8.7 million processor instructions per second on a laptop.
- Across 15 benchmark programs, the paper reports it is 10 to 39 times faster on average than the fastest existing zkVMs on the same machine, and more than 300 times faster than two others on a token-transfer benchmark.
- An optimized Magnetar proves 9.5 to 12.8 million cycles per second, increasing each reported speedup by about 1.3x, and its proofs verify in about 12 milliseconds.
- Its security depends only on a standard hash function such as SHA-256 or SHA-3, requiring no elliptic curves, lattices, special-purpose hash functions, trusted setup, or unproven error-correcting-code conjecture.
For AI builders, Magnetar makes ordinary compiled programs provable without cryptographic expertise and removes a post-quantum migration from the roadmap.
Shanghai AI Lab's "Shusheng-S2" large model enters frontier mathematics research workflow: assists proof of the three-dimensional Ghys conjecture and passes formal verification
Shanghai AI Lab's "Shusheng-S2" large model entered a frontier mathematics research workflow, assisting the proof of the three-dimensional Ghys conjecture, which passed formal verification.
- Shanghai AI Lab's "Shusheng-S2" large model entered a frontier mathematics research workflow.
- The model assisted the proof of the three-dimensional Ghys conjecture.
- The related proof passed formal verification.
This shows a large model can take part in a frontier mathematics proof and yield a formally verifiable result, offering a verifiable workflow example for AI-assisted mathematics research.
OpenRouter adds NVIDIA Nemotron 3.5 Super VL multimodal MoE model
NVIDIA Nemotron 3.5 Super VL is an open multimodal mixture-of-experts model with 12B active parameters out of 120B total, using a hybrid Mamba-Transformer architecture with multi-token prediction and a 1M-token context.
- 120B total parameters with 12B active, built on a hybrid Mamba-Transformer architecture with multi-token prediction.
- Context window of 1,048,576 tokens, supporting up to 262,144 completion tokens.
- Accepts text, images, documents, GUI screenshots, and video as input and returns text; reasoning is on by default and can be turned off or given a token budget.
- Priced at $0.10 per 1M input tokens and $0.40 per 1M output tokens, with cache reads at $0.08/M, served by Crusoe, Together, DeepInfra, and CoreWeave.
For teams building long-running agents, 12B active parameters with a 1M-token context and visual input lets document intelligence, GUI interaction, and video understanding run through one model instead of separate vision and text pipelines.
Joint Subspace Sparse Attention for Efficient Long-Context Decoding
JSSA learns a shared low-dimensional query-key subspace to improve token-level sparse selection ranking, and introduces a gather-free token-level indexed-attention kernel that supports distinct per-KV-head selections under the standard GQA KV-cache layout.
- The paper reports that on an NVIDIA B200 GPU, JSSA achieves up to 6.43x the decoding throughput of full-KV attention at 128K context.
- The paper states JSSA remains close to full-KV accuracy across standard GQA, hybrid sliding-window/global attention, and Multi-head Latent Attention (MLA) models.
- The paper reports JSSA reaches reasoning accuracy comparable to model-native DSA without model training and reaches 96-99% of its decoding throughput even at 700B scale large models.
- The paper notes that direct refinement of the exact objective provides no measurable improvement in the tested settings.
For builders of long-context inference, JSSA matters because it pairs token-level sparse selection with a gather-free kernel that lowers decoding cost directly on existing GQA and MLA KV-cache layouts.
WorldMemBench: A Comprehensive Benchmark for Memory of Video World Models
The paper introduces WorldMemBench, a benchmark that systematically evaluates world memory in video world models across 260 test cases, covering static memory (human identity and appearance, object appearance and geometry, environmental content and geometry) and dynamic memory (event progression and state persistence).
- The benchmark spans 260 test cases assessing static and dynamic memory.
- It uses controlled camera departures and multi-revisit trajectories, comparing static content against fixed initial references and dynamic events against expected state transitions.
- Evaluation of nine representative models shows strong static retention does not imply reliable dynamic state evolution.
- Repeated revisits reveal memory degradation beyond the first return, and human studies support alignment between the automatic evaluations and human judgments.
For builders of video world models, this benchmark separates memory into measurable static and dynamic axes and shows that static retention cannot stand in for state persistence across repeated revisits.
Similar Models, Broken Odds: Where Luce's Axiom Fails in Chatbot Arena
The paper argues that Chatbot Arena's Bradley-Terry model gives each model one score and therefore assumes Luce's choice axiom, and on 1.8 million Arena votes it finds votes between versions of one model are more decisive than Bradley-Terry predicts.
- On 1.8 million Arena votes, votes between versions of one model are 1.37 times as decisive as Bradley-Terry predicts.
- Between sizes of one generation the multiplier is 1.26, and between models of one lab it is 1.12.
- A correction with four to six parameters predicts held-out votes better than Bradley-Terry and than models with hundreds of parameters, mostly by modelling ties.
- With language-model judges and three-answer menus, a near-copy takes votes mostly from its twin when choices are pooled over prompts, with a log-odds shift of +0.10 and a 95% interval of 0.05 to 0.16.
For builders, this means Arena-style pairwise win rates between near-copy models can be systematically overstated, so single-score rankings of similar models should be read with caution.
fishaudio/drama-3: a 35B-parameter MoE text-to-speech model with 16 speakers and streaming
Fish Audio released Drama 3, a next-generation text-to-speech model built for multi-character scenes and long-form content, supporting multi-speaker generation, voice cloning and natural-language control. It uses a 35B-parameter MoE text backbone with 3B active per token and a Codebook-MTP audio decoder.
- 35B parameters with 3B active per token, using an MoE text backbone with hybrid full plus linear attention.
- Speaker tags <|speaker:0|> through <|speaker:15|> let one text carry up to 16 characters.
- The Codebook-MTP audio decoder decodes 10 codebooks causally over the whole sequence.
- MoE experts are FP8 with the remaining weights in bf16; the model page lists 36B params and tensor types BF16 and F8_E4M3.
For builders needing multi-character long-form audio, Drama 3 bundles 16 speaker tags, streaming chunking and SGLang-Omni deployment, but the research-oriented license and FP8 MoE hardware requirements shape where it can ship.