AI Research · Paper roundup for November 11, 2026

Argues that QK-Norm discards useful magnitude information in queries and keys, and proposes QK-MAP, a magnitude-aware projection that encodes magnitude as bounded phase patterns with learnable frequencies. From 190M to 7B parameters it reaches lower training and validation loss than standard attention and QK-Norm, puts less attention on the first token and sharpens attention on relevant evidence in synthetic long-context tests.

Why it mattersQK-Norm is a common default, so a drop-in alternative with lower loss is relevant to pretraining recipes; code is public.

arXiv ↗PDF ↓Code ↗

An agentic framework where experts choose what to prove, LLM agents write the proofs and Lean checks every result. Applied to the Signal protocol's core cryptographic code down to machine instructions, agents powered by Gemini Carbon produced over six million lines of Lean in four weeks (the page metadata says three million), flagged about 1% of the code for human review, and proved reintroduced historical bugs insecure.

Why it mattersa large-scale case of agents delivering machine-checked security guarantees, with the proof checker as the source of trust.

Blog ↗Paper (PDF) ↗Artifacts ↗

A benchmark for agents reasoning over visually complex files in real knowledge work. Experts from 10 verticals and 83 occupations wrote tasks that take a professional 3–8 hours; each is graded on about 70 yes/no rubric checks (75% factual accuracy, 25% visual design) by an agentic autorater that agrees with experts 96% of the time. The private set has 200 tasks run by Artificial Analysis, plus 10 public tasks; models used 30–130 turns per task.

Why it mattersit moves agent evaluation toward multimodal deliverables and cost per task, under an independent evaluator.

Blog ↗Tech report (PDF) ↗Dataset ↗Leaderboard ↗

Reasoning-model agents resend a growing transcript every call, and costs depend on what the client keeps and what the server caches. The authors propose a cost model that matches an instrumented engine exactly, with one rate per processed token and one per attended token pair, whose ratio is an exchange rate between thinking and remembering. On Qwen3-8B ScienceWorld runs, dropping earlier reasoning cost 1.74x as much as keeping it when the server reuses generated states, but 0.34x when it reuses nothing; on HotpotQA reruns, dropping reasoning cut correct answers by 9–23 points across four models.

Why it mattersthe cheapest context policy depends on the serving stack, and dropping reasoning can hurt accuracy.

arXiv ↗PDF ↓

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, the same underlying model, with Fable carrying extra cybersecurity and biology safeguards for general availability and Mythos reserved for vetted organizations. Self-reported results include 69.9% on FrontierSWE v2 (Opus 5.5: 62.3%), 89.3% on OSWorld 2.1 and 74.2% on Humanity's Last Exam with tools; in prerelease testing about 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in roughly 30 hours. Pricing is $8/$40 per million input/output tokens.

Why it mattersa new frontier tier with large agentic gains and a tiered approach to dual-use safeguards.

Announcement ↗System card ↗

Google's strongest model, positioned above Gemini 4 Argon for cybersecurity and long-horizon agentic work in coding, knowledge work and science, spending more tokens to reason and verify on hard tasks. Google reports #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at over 2x lower cost per task than the next-best external model, #2 on MKA, a 0.8% per-attempt breach rate under adaptive red-teaming (Argon: 3.1%) and 92% on a 375-task agentic honesty benchmark (Argon: 70%). Introductory pricing is $2/$10 per million tokens, rising to $4/$20 on Nov 30, 2026.

Why it matterscompetition is shifting to cost per agentic task, with published safety deltas.

Announcement ↗Model card (PDF) ↗

Meituan released open weights for LongCat-2.5, a Mixture-of-Experts model with about 1.6T total parameters and about 48B active per token (Hugging Face metadata says 1.8T), with native visual modeling that turns image features into discrete tokens aligned with text. Self-reported scores include 88.4 on Terminal Bench v2.1, 64.6 on DeepSWE v1.1, 91.9 on BrowseComp and 63.1 on HLE with tools. Weights are MIT-licensed, with a W4A4 quantized variant.

Why it mattersa permissively licensed trillion-scale multimodal agent model narrows the open-closed gap.

Hugging Face ↗Tech blog ↗

A family of six open, Arabic-centric agentic models: dense 0.9B, 3.7B, 7B and 32B, plus MoE 36B (4B active) and 375B (23B active), adapted from K2 Horizon with continual pretraining, SFT, preference optimization and RL with verifiable rewards; all but the 0.9B have 512K context. The report says Jais 3 matches or beats similarly sized open models on Arabic and agentic benchmarks, that the 375B model is competitive with Claude Sonnet 5 in Arabic, and that it hallucinates less. Weights, training, RL and evaluation code and audited Arabic benchmarks are released.

Why it mattersa complete open template for language-specific agentic models.

arXiv ↗PDF ↓Weights ↗Training code ↗

Zyphra released ZONOS3-Mini, a compact zero-shot voice-cloning TTS built for real-time on-device use: a sparse MoE with 100M active and 800M total parameters. Its ZAC-1 codec cut word error rate about tenfold versus a DAC-based model under similar data, it starts speaking in under 570 ms in a browser on a Pixel 9 and under 150 ms on an M3 MacBook Air, and it tops open models on Zyphra's own TTS Capability Index v1. Trained on 112 H100s in 2.5 days; Apache 2.0.

Why it mattershigh-quality voice cloning becomes practical fully on-device, a building block for private, low-latency voice agents.

Blog ↗Hugging Face ↗Code ↗

Public ASR benchmarks show low error rates, yet production systems still rely on hotwords, biasing and post-correction. KeyASRBench is a bilingual Chinese-English pilot that measures occurrence-based Keyword Error Rate on whole-phrase annotations: a 10-hour challenge set of 1,501 clips with 4,723 keyword occurrences. Across five local checkpoints and two official APIs, MER/WER is 5.31–12.91% while KER is 13.61–33.81%, and a disjoint 30-minute control gives 10.65–27.31%.

Why it mattersKER captures the name and term errors that break voice-agent tool calls, which WER hides.

arXiv ↗PDF ↓