VIEN

AI research

AI Research · Paper roundup for November 18, 2026

· 10 papers · 4 sections

OpenAI shipped GPT-6.2 Astra (frontier, $8/$40 per million tokens) and GPT-6.2 Sol (efficient, $2/$10) in ChatGPT Work, Codex and the API; OpenAI reports Astra beats GPT-6 Astra by 5.2 points on Agents' Last Exam at under half the cost. Open models and tech reports: K2-Horizon releases six fully open models (0.9B to a 375B-A23B MoE) with data, logs and code; Sage S2.1 Flash is a 320B MoE for scientific reasoning and acting that its authors say tops OfficeQA Pro and SpreadsheetBench 2. Coding agents are not yet digital employees: on SWE-CoWork the top two models pass about 63% of tests but finish under 8% of assignments; EfficientSkill cuts agent output tokens by up to 80% while raising accuracy. Speech: DPO, GSPO and self-distillation post-training lift Qwen3-TTS Emotion scores from 18.5 to 81.2; tokenizer-free VoxMel TTS reaches 1.15% WER; a Nemotron 3.5 Transcribe serving engine hits 6,367x real time on B200.

Large Language Models

1

Latent Feedback Is a Compute and Data Amplifier

Latent feedback feeds each token's top-layer state into the next token's input; at equal compute it saves up to 48% of compute and 47% of unique data versus a Transformer, and a fitted scaling law predicts 778M-scale loss within 0.3%.

Why it mattersAs high-quality text runs short, architectures that squeeze more out of each token become more valuable than more FLOPs. A scaling law validated at held-out size makes the claim testable at larger scales, and the code is public.

AI Agents

2

SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations

A 101-assignment benchmark inside a simulated organization of 26 colleagues: the top two models pass about 63% of tests but complete under 8% of assignments, and about 31% of requirements reported done actually fail their tests.

Why it mattersIt moves coding-agent evaluation from isolated issues to the organizational work of finding out what is actually wanted. The gap between test pass rate and full completion shows that tracking requirements end to end, not writing code, is now the bottleneck.

3

EfficientSkill: Learning Skills for Token-Efficient Multi-Turn Agents

Learning a skill and then pruning it (compressing instructions, capping turns, moving code into utilities) cuts output tokens by up to 80% versus EvoSkill while raising accuracy, e.g. +17.5 points on SpreadsheetBench with Grok 4.1 Fast.

Why it mattersSkill files are becoming the main way to adapt agents without retraining, and their unchecked growth is a hidden cost. Showing that pruning can cut tokens and raise accuracy at the same time makes cost a first-class objective in skill learning.

Model Releases

4

GPT-6.2 Astra and GPT-6.2 Sol

OpenAI released GPT-6.2 Astra ($8/$40 per million tokens) and GPT-6.2 Sol ($2/$10) in ChatGPT Work, Codex and the API; Astra beats GPT-6 Astra by 5.2 points on Agents' Last Exam at under half the cost, and Sol nearly matches Astra on FrontierCode 2.0 at about a fifth of the cost.

Why it mattersThis is the new reference point for frontier agentic work, and the pricing split makes the Astra-versus-Sol choice a real cost decision for builders. The 22-point jump on an exploit benchmark also explains why OpenAI tightened cyber safeguards.

5

K2-Horizon: A Frontier Fleet of Open Source Language Models

Six fully open models from 0.9B to a 375B-A23B MoE, released with data, logs and training code; MoVA lets a 36B-A4B model nearly match a 32B dense one, and Uno adapters speed decoding up to 2.2x.

Why it mattersFully open releases with data, logs and training code at frontier scale are rare, and this one spans six sizes. It gives researchers a reproducible base for studying pre-training and RL choices, not just open weights.

6

Sage S2.1 Flash Technical Report

A 320B MoE with molecular and biological-sequence encoders, evaluated on 33 benchmarks; the authors report first place on OfficeQA Pro, MLE-bench Lite, Harvey LAB and SpreadsheetBench 2. The arXiv abstract calls it "Sage S1.5 Flash"; no weights link yet.

Why it mattersIt is a strong sign that scientific-domain models are converging with general agentic models rather than staying separate. Whether weights follow will decide how much the community can build on it.

Speech

7

Rethinking Post-Training for Expressive Zero-Shot Text-to-Speech

Three-stage post-training (DPO, GSPO, on-policy self-distillation) makes zero-shot TTS adapt delivery to the text; on Qwen3-TTS, Emotion scores rise from 18.5 to 81.2.

Why it mattersExpressiveness, not intelligibility, is now the main gap in zero-shot TTS, and this shows RL-style post-training can close much of it on a production-grade model. The recipe is general enough to transfer to other autoregressive TTS systems.

8

VoxMel TTS Technical Report

Tokenizer-free autoregressive TTS on mel-spectrograms via an invertible Haar transform; the 0.6B model after GRPO reaches 1.15% WER (English) and 1.17% CER (Chinese).

Why it mattersDropping the discrete audio tokenizer removes a major quality ceiling and source of artifacts in LLM-style TTS. Strong numbers at 0.6B make it a credible alternative design for compact speech generators.

9

Sada-ASR: A Production Multidialectal Arabic Speech Recognition System

Multidialect Arabic ASR behind EchoNova's production voice agent: 11.69% WER on a 120-hour test set, lowest of 12 systems compared; scoring code and a 20-hour test subset are released.

Why it mattersDialect coverage is the practical blocker for voice agents in Arabic-speaking markets, and this is a rare production report with a public benchmark slice. It also shows that a focused regional system can beat every big-lab ASR on its home turf.

10

High-Throughput Serving of Nemotron 3.5 Transcribe with Regime-Aware Megakernels and Kernel Fusion

An ASR serving engine that picks megakernels or fused kernels per pass: 6,367x real time on B200 (1.64x vLLM) at unchanged WER, and 8x more concurrent requests within a 400 ms p95 budget.

Why it mattersServing cost, not model accuracy, now limits how cheaply voice products can transcribe at scale. Kernel-level gains of this size translate directly into lower latency budgets for real-time voice agents.

Latest tech newsLLM & Agent · Briefing for November 20, 20267 stories · Friday, November 20, 2026 →