LLM Innovation Tracker

LLM Inference Breakthroughs

• 🔥 JetSpec Speculative Decoding: Hao AI Lab released JetSpec, a parallel tree drafting method achieving up to 9.64x...

• OpenAI advisers urge delaying IPO until next year amid market volatility and financial challenges.
• Reports question the company's high valuation...

JetSpec resolves the causality-efficiency tradeoff in speculative decoding by training a causal parallel draft head on fused target hidden states,...

• Task-specific small models beat frontier LLMs by 8-9% accuracy while running up to 5,250x cheaper and 8x faster
• Salesforce Help Agent deploys in...

• OPID extracts hierarchical skills from on-policy trajectories for dense token-level supervision, avoiding external memory costs.
• The second paper...

• Microsoft's MAI-image-2.5 ranks #2 for text-to-image and #3 for editing, now live in OneDrive, PowerPoint, and Foundry API.
• OpenAI Codex adds...

The STOC 2026 workshop signals TCS research on LLMs reaching maturity, connecting classic theory to modern models via talks on alignment distortion, fast attention algorithms, and using frontier LLMs for automated proofs.

Successful deployments converge on three practical steps.

• Start narrow: Prove one workflow end-to-end before expanding; broad attempts trigger...

New techniques address core multimodal limitations in reliability, context, and representation.

• Robust-TO adds per-frame confidence scoring to...

Frontier Model Updates

• 🔥 GPT-5.5 Instant: OpenAI released an updated version of GPT-5.5 Instant with improved intent understanding, complex...

As frontier models mature, enterprises are capturing value in three adjacent layers rather than the models themselves.

• Application orchestration:...

• iLLaDA scales masked diffusion to 12T tokens, matching Qwen2.5 7B on key benchmarks while using fully bidirectional attention.
• Output compression...

Two complementary methods reveal distinct LLM strengths in simulating human behavior:

• BehaviorBench tests four core capabilities across tasks,...

Claude Fable 5 is returning with GPT 5.6 likely launching tomorrow, putting major frontier model releases back on track.

HarnessX autonomously rewrites its modular AI scaffolding during tasks, delivering an average +14.5% gain across benchmarks and up to +44% for the 9B...

Fable 5 usage references have appeared in Claude Code v2.1.190, including new limit messages and the removal of "purchased separately" language, while similar mentions reemerged in Amazon Bedrock.

OpenAI's most-used model just became much more fun to talk to, with sharper intent understanding, better response adaptation, and more reliable...

DomainShuttle enables open-domain subject-driven text-to-video generation with strong performance in both in-domain fidelity and cross-domain...

Three papers signal a shift from basic LLM-tool agents toward deeper architectural foundations.

• Memory systems require modular designs; no...

Waymo's fleet has driven 220M+ fully autonomous miles with safety performance staying remarkably consistent—even as it scaled into complex new cities and airports. The result: 16x safer than human drivers.