VIEN

AI research

AI Research · Paper roundup for November 17, 2026

· 12 papers · 5 sections

This edition covers about 3 days (arXiv submissions from Nov 13–15 and lab releases on Nov 16). Office-work agents still fall short: the best model passes only 38.3% strict pass@1 on GDP.xlsx (70 real spreadsheet requests), and top coding agents reach 62.3% on WebArena-Pro (300 tasks with audio and video). New models: Alibaba made Qwen4-Max generally available (2.4T MoE, 71B active, 1M context, $1.50/$4.50 per million tokens, closed weights); Tencent launched Hy4 flash (224B, 12B active) at roughly 90% lower prices than Hy4. LLM auditing and efficiency: a model's fingerprint transfers to whichever model continues the conversation, undermining detection of covert API model swaps; Concertina folds Qwen3-4B into a 2.27B model keeping ~90% of its ability with only 22B tokens.

Large Language Models

1

Contagious Fingerprints: When LLMs Inherit Another Model's Identity

An LLM's classifier-recognizable fingerprint is inherited by whatever model continues a conversation it started. Detection of a substitute dropped from 0.97 to 0.72 when DeepSeek replaced Claude Sonnet mid-conversation, and from 0.87 to 0.54 for a same-vendor downgrade to Haiku. The proposed Lineup audit ranks the audited reply among reference replies sampled on the same history, reaching a 0.056 false-positive rate and catching 0.81 of substituted sessions after four turns.

Why it mattersAPI resellers and relays can silently swap in cheaper models, and this shows that the standard classifier-based audit is much weaker than assumed once a conversation is already underway. Lineup offers a threshold-free, false-positive-controlled alternative that needs no log-probabilities.

2

Folding Language Models into Themselves

Concertina builds a compact student directly from a teacher's weights: every N adjacent teacher layers become one shared block run N times, with per-step low-rank corrections initialized from the replaced layers and a per-token gate. From a Qwen3-4B teacher, a 2.27B model distilled on 22B tokens keeps about 90% of the teacher's commonsense and math performance and beats Qwen3-1.7B, which saw over a thousand times more data.

Why it mattersLooped, weight-shared models are a cheap route to small models, and initializing them from an existing teacher sidesteps most of the pre-training bill. Code is released, so the recipe can be tried on other open models.

3

MMFlash: Faster Speculative Decoding for Vision-Language Models

Speculative-decoding acceptance falls from 46% for the least visually dependent tokens to 22% for the most, a gap that persists after multimodal training. MMFlash upweights poorly predicted visually dependent tokens during drafter training and adds sparse attention for long videos, achieving lossless speedups up to 3.58x, beating EAGLE-3 by up to 195.4% and DFlash by up to 17.1%.

Why it mattersIt pinpoints why text-tuned drafters underperform on multimodal inputs and gives a targeted fix. Faster lossless decoding matters directly for serving VLMs and video agents.

AI Agents

4

GDP.xlsx: Benchmarking Frontier Models on Professional Workbooks

GDP.xlsx has 70 professional spreadsheet requests across 13 domains, each with a real workbook and an expert rubric (791 binary criteria over about 2.18 million populated cells), graded by an agentic judge that opens the output file. Under strict pass@1, the top model, Gemini 4 Argon, passes 38.3% of attempts and the median of 20 configurations is 19.3%. The harness and three public tasks are released; 67 are held out.

Why it mattersSpreadsheets are where a lot of real knowledge work happens, and an all-criteria-pass metric exposes how often agents get most things right but still ship a broken workbook. The held-out design makes it a credible leaderboard going forward.

5

WebArena-Pro: A Heterogeneous, Multimodal, Reproducible Benchmark for Web Agents

WebArena-Pro has 300 tasks across 20 self-hosted web apps in six domains, treats audio and video as core observations with dedicated actions, and runs each task in isolation for reproducible, parallel evaluation. Codex with GPT 6.1 Sol and Claude Code with Claude Opus 5.5 both reach 62.3%, OpenCode with Kimi 3.1 Pro 60.7%, while the best GenericAgent configurations reach only 30.3%.

Why it mattersIt is the first WebArena-style benchmark that makes audio and video first-class, and the 2x gap between coding-agent harnesses and generic web agents shows how much the harness now matters. Isolation per task fixes a long-standing reproducibility problem.

6

CarryOn: Learning What to Keep in Context for Long-Horizon LLM Agents

CarryOn pairs a compact execution state with fine-grained, source-linked evidence, using a learned selector and adaptive retention policy that need no extra LLM calls. Across AppWorld, OfficeBench and Tau2 Telecom with three actor LLMs, it reaches 75.7% macro-averaged task success, 6.7 points above the strongest baseline, and the best cost-aware scores on two new metrics (S_α and normalized cost per success).

Why it mattersContext management is becoming the bottleneck for long-horizon agents, and a learned selector that needs no extra LLM calls is cheap to bolt on. The cost-aware metrics are a useful addition to agent evaluation.

7

Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agent

Gamer is a vision-language-action foundation model for long-horizon, real-time game agents, pre-trained on tens of thousands of hours of device-level keyboard and mouse data. It adds sparse-thinking instruction tuning, multi-turn RL for finite context and sparse rewards, and a cloud-edge inference engine with unbounded context. The authors report skilled-player-level efficiency across several 3D open-world games, though the abstract gives no specific scores.

Why it mattersThis was the most-upvoted agent paper on Hugging Face Daily Papers, and it pushes VLA models from robotics into real-time, keyboard-and-mouse control at scale. Concrete benchmark numbers and weights are still missing, so claims need checking.

Model Releases

8

Qwen4-Max (general availability)

After a limited preview on Nov 11, Alibaba made Qwen4-Max generally available via QwenCloud, QwenWork and QwenStudio. It is a sparse 2.4T-parameter MoE with about 71B active parameters, native vision and a 1M-token context. Alibaba claims 6th place on Code Arena and 35% fewer tokens with about 60% lower cost per task than Qwen3.8-Max on its internal Work360 benchmark. List price is $1.50/$4.50 per million input/output tokens; weights are closed, with smaller open Qwen4 models promised.

Why it mattersThis turns last week's preview into a generally available flagship priced well below US frontier models, and it signals that open-weight Qwen4 models are next. Independent benchmarks are still missing.

9

Hy4 flash

Tencent launched Hy4 flash, a low-cost natively multimodal sibling of Hy4 with a 1M-token context (960k input, 64k output). Tencent reports 224B parameters (including 64B Engram parameters) with 12B active, a KV cache one-sixth of Hy4's and decoding above 150 tokens/s. It costs $0.070/$0.281 per million input/output tokens, about 90% below Hy4, and scores slightly above Hy4 preview on Tencent's internal WorkIndex at about one-thirteenth the cost per task. Open weights are planned without a date.

Why it mattersHy4 flash pushes the price of a 1M-context multimodal model close to zero, adding pressure on low-cost tiers from DeepSeek and others. All quality numbers are Tencent's own internal benchmarks.

Speech

10

eDIAR: Training-Free Open-Speaker Streaming Diarization for Long Meetings and Large Archives via e-means

eDIAR keeps Streaming Sortformer's slot head for activity and overlap detection, then clusters speaker embeddings within slots using e-means once all slots are filled, re-clustering only a 90-second window so each update costs about 57 ms. With 20 speakers DER falls from 66.5% to 43.7%, and at minute 45 of hour-long meetings with 12–20 speakers it holds around 25% versus about 57% for Streaming Sortformer.

Why it mattersReal meetings routinely exceed the speaker caps of streaming diarizers, and this fix is training-free and constant-cost. A model and demo are already on Hugging Face.

11

Marimba: A System One Model for Speech

Marimba answers typed speech questions (term present, turn ended, language, speaker change) on frozen Qwen3-ASR 0.6B features. A layer-18 readout detects turn ends with AUROC 0.930 (vs 0.982 for Smart Turn v3.2 on its own test set, and on par on ICSI), and layer-12 readouts detect speaker changes with AUROC 0.966–0.980 on AMI and VoxConverse, beating ECAPA and counting speakers better than pyannote 3.1. Code and weights are released.

Why it mattersReusing one frozen ASR backbone for turn-taking, keyword, language and speaker questions is attractive for voice-agent stacks that currently run several separate models. It is a single-author preprint, so results deserve independent replication.

Voice Agents

12

FEND: Far-End Detection in Outbound Voice AI Calls

FEND classifies the far end of an outbound voice-AI call as live person, voicemail, IVR or AI callee within 20 seconds, using audio plus a streaming transcript and a detector per automated class. In 1,000 production calls, 48.1% reached a person, 44.6% voicemail, 6.9% an AI and 0.4% an IVR. Voicemail detection is 100% accurate on a public benchmark without misclassifying humans (Asterisk AMD misclassifies 30–66%), speech timing separates AI from human at 0.952 ROC-AUC, and person-vs-automated F1 in production is 0.979.

Why it mattersVoice agents are now calling other voice agents often enough (6.9% of calls here) that far-end classification needs an AI class. The production numbers make this a practical reference design for outbound calling stacks.

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