One-Line Summary

Multimodal plumbing advances today: a new tokenizer family, code-as-tools for agents, cheaper head-to-head evaluation, and serving support for a 2.8T-parameter model.

Research Papers

KVAE introduces tokenizers for audio, image, and video

KVAE is a set of compact “translators” that turn audio, images, and video into small latent representations so generative models can learn and synthesize more efficiently. KVAE-Audio is a continuous full-band 48 kHz tokenizer producing a 50 Hz latent with 64 channels; KVAE-3D offers two causal video tokenizers with 4×16×16 and 4×8×8 compression; and KVAE-2D compresses images by 8× with 32 channels. ¹

In Latent Diffusion Modeling (LDM), the tokenizer determines how raw signals are compressed; better tokenizers can improve learning speed and output quality. The paper reports that reconstruction metrics like Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS), and Perceptual Evaluation of Speech Quality (PESQ), as well as generation metrics such as Contrastive Language–Image Pretraining (CLIP) score and Fréchet distance, match or surpass public Variational Autoencoder (VAE) baselines from Wan‑2.2, HunyuanVideo‑1.5, FLUX.2, MovieGen, StableAudio, and MMAudio, with code and training details released. ¹

For teams building text-conditioned generation across media, one tokenizer family per modality simplifies pipelines and reuse. The authors share selection methods and ablations, positioning KVAE as a reusable building block rather than a one-off component. ¹

Continual learning moves from parameters to systems

This survey argues that continual learning is shifting from tweaking weights to adapting entire systems — including when learning happens (pre-, post-, or at inference), how updates occur (off-policy, on-policy, or beyond gradients), and where updates live (inside parameters vs. external components like memory or protocols). It frames this “When–How–Where” tri-axial view and catalogs methods that extend learning beyond static parameters. ²

A complementary survey introduces a six-dimensional taxonomy of post-training adaptation — mechanism, goal, data requirement, persistence, structural scope, and model type — clarifying often-confused terms such as fine-tuning, retrieval-augmented generation, and prompting. The authors emphasize documentation, change-tracking, and governance as adaptation layers stack across foundation models, Large Language Models (LLMs), and multimodal systems. ³

Programmatic tool calling matches or beats JSON in 11 of 14 models

Instead of filling out JavaScript Object Notation (JSON) schemas, programmatic tool calling (PTC) lets a model write short Python code that invokes tools, returns results, and continues in one agent turn. In head-to-head tests across 14 models, PTC matches or exceeds native JSON tool calling in 11 of 14 models, with the GPT‑5.6 family showing a 10.6% improvement over the JSON baseline. ⁴

PTC also matches or outperforms in 13 of 14 models under parallel fan-out and remains stable under context-rot conditions where the JSON baseline degrades by 2.3% on average. The practical upside: natural chaining and parallelization via code, not brittle schema juggling. ⁴

AV-AIVAT cuts evaluation games with certified early stopping

Comparing which agent is stronger usually means playing many games; AV-AIVAT adds mathematically valid “stop-when-enough-evidence” to reduce that cost. It combines the Action-Informed Value Assessment Tool (AIVAT) for variance reduction with continuously monitored Confidence Sequences (CSs), yielding a median 54× variance reduction across 15 LLM agent configurations over 71,439 paired Heads-Up No-Limit Hold’em (HUNL) hands. ⁵

At a nominal 95% level and a target precision of ±1 Big Blind, raw outcomes need a median 74× as many hands as AIVAT-corrected outcomes to stop under Asymptotic Confidence Sequences (AsympCS). Exact finite-sample certification via Empirical-Bernstein Confidence Sequences (EB-CS) further supports auditable decisions at the moment of stopping. ⁵

Open Source & Repos

SGLang adds day‑0 serving for a 2.8T‑parameter LatentMoE model

SGLang, a high-performance serving framework for Large Language Models (LLMs) and multimodal models, announces day‑0 support for Kimi K3 — a 2.8‑trillion‑parameter multimodal Mixture of Experts (MoE) model with a 1M‑token context window, using 896 experts with top‑16 routing in a 3584‑dimensional latent space and linear‑attention layers. For practitioners, this signals that the stack can schedule and stream very large, long‑context MoE workloads. ⁶

For teams shipping new models quickly, day‑0 support reduces integration lag and encourages benchmark and inference experimentation as soon as weights or endpoints are available. ⁶

Why It Matters

Better tokenizers are foundational: in LDM pipelines they set the ceiling for sample fidelity and training efficiency, and a public, ablated release helps standardize quality and reproducibility across audio, image, and video. ¹

Agent work is getting more practical: using code as the interface for tools yields measurable gains (10.6% in one family here) and simpler chaining, while statistically sound early stopping trims evaluation cycles without sacrificing rigor — a pattern that aligns with infrastructure keeping pace with 2.8T‑scale, long‑context serving. ⁴

What to Try This Week

1. KVAE paper quick read: scan the abstract and figures to see how audio/video/image tokenizers are configured. https://arxiv.org/abs/2608.05798
2. SGLang quickstart: check the repo’s README to serve a local model and test long‑context prompts. https://github.com/sgl-project/sglang