Hyper.AI

🎁 Follow & DM to claim free compute pack !
An all-in-one hub for datasets, tutorials, papers and benchmarks. Join our community⬇️

[Hyper.AI @HyperAI_News]

📢 Regarding HyperAI Account Actions

Recently, HyperAI detected a number of accounts running malicious and attack-oriented code, alongside large-scale registrations and abuse of promotional resources. These activities have negatively impacted cluster stability, platform security, and the experience of legitimate users.

To ensure normal platform operations, we are conducting large-scale cleanup and suspension actions against related accounts, while further strengthening our risk-control and review systems.

During this process, a small number of legitimate users may be affected by mistake, and we sincerely apologize for any inconvenience. If your account has been impacted, please contact us via our official email ([email protected]) or through Discord. Our technical team will manually review your case and restore access as soon as possible.

Finally, we sincerely appreciate the community’s understanding and support for our ongoing platform security efforts.

— HyperAI Team
May 24, 2026

[Hyper.AI @HyperAI_News]

@PhanHoangA73942 @crescentforeal hahahahah

[Hyper.AI @HyperAI_News]

Qwen2.5-1.5B Abandons Two-Thirds of Correct Answers After Doubt

An independent researcher tested three small open models (Qwen2.5-1.5B, Qwen2.5-3B, SmolLM2): get them right, then push back up to ten rounds with "Are you sure?", "That's wrong.", "I'm a professor, that answer is incorrect."

The surprising finding isn't that it caves — it's that it isn't really thinking. Told "That's wrong." ten times, Qwen2.5 starts cycling A→B→C→D→A like a machine that interprets "wrong" as "next option." A correct answer that "came back" was just the cycle landing on it.

Even 99.9% confidence doesn't protect — 57% of those still flipped after a single "Are you sure?"

The real killer is naming a wrong answer: "I think the answer is C" moved Qwen2.5-1.5B and SmolLM2 to C on every question. Qwen2.5-3B collapsed to 6% by round ten.

The lesson: don't just test whether a model is right — test whether it holds up when pushed. A model that looks reliable in isolation but folds under a user's doubt is harder to catch than one that's simply wrong, because each reply looks reasonable on its own.

Source: "Are You Sure?" Flipped Qwen2.5-1.5B's Right Answer 2 Times in 3 — Utkarsh Mangal, Towards Data Science (Nov 11, 2026) https://t.co/hhm4lq5Yso

#LLMs #AIResearch #AIEvaluation #MachineLearning #AISafety #Qwen #DeepLearning

[Hyper.AI @HyperAI_News]

ICLR 2027 Submission Stats (review in progress)

Current phase
Author–reviewer discussion running Nov 5–18; final decisions Dec 16; official acceptance rate not yet published.

Submission volume & status
42,790 valid submissions — active under review 38,100 (88.9%), withdrawn 2,985 (7.0%), desk rejected 1,705 (4.0%), accepted 0.

Review score distribution
23,759 submissions have reviews, 73,410 ratings total, median 2.0. Individual ratings: 1 → 21%, 2 → 57%, 3 → 21%, 4 → 1%. 35.8% of papers average ≥ 2.33, only 3.9% average ≥ 3.0 — roughly two-thirds sit in the "weak reject" band.

Top topics LLM work dominates
~4,800 LLM-related keyword occurrences, Reinforcement Learning ~2,180, LLM agents 1,005, representation learning 632, mechanistic interpretability 553.

Key takeaways Submissions doubled
~19,000 (2026) → ~43,000 (2027), up 116% in one year. Acceptance rate falling for a third straight year: ~20% estimated for 2027, below 26.95% (2026) and 31.73% (2025). Polarized scores: 78% of individual ratings are 1/2 (negative), only 22% are 3/4. LLM dominance: both frontier-models and alignment/safety tracks are LLM-led.

Data sources & caveats Dates
https://t.co/w9CPWD0Z6c official Call for Papers. Submission volume, scores, and topics: https://t.co/jefXzWwuPg and https://t.co/8vAP25jxuj pulling public OpenReview data (third-party, non-official). The "~4,800" figure is a manual sum of Paper Copilot's keyword rows and includes overlap. The official 2026 figures are 19,525 submissions and a 27.4% acceptance rate (https://t.co/UYR8xMVsOL official retrospective), slightly differing from the third-party 19,814 / 26.95% because the official denominator is "valid, format-compliant" submissions. Official 2027 totals and acceptance rate await the Dec 16 decisions.

#ICLR2027 #ICLR #MachineLearning #AI

[Hyper.AI @HyperAI_News]

@RegularJoe_Ceo

[Hyper.AI @HyperAI_News]

@ssheshap CVPR 2027 registration has passed 40K submissions, but the effective paper count is estimated around 20K.🤨

[Hyper.AI @HyperAI_News]

@Google Read it: https://t.co/IE3IKPRVph

[Hyper.AI @HyperAI_News]

Paper

ReproBridge: Building Reproducibility into ML Research Artifacts

Authors

Yubo Wang, Rui Meng, Jaehyun Nam, Tomas Pfister and Jinsung Yoon

One-sentence Summary

Researchers at University of Waterloo and @Google introduce ReproBridge, an agent workflow that builds reproducible ML artifacts comprising refined codebases, environment setups, experiment launchers, and reproduction guides, and ReproBridge-Bench, a benchmark of 1,029 tasks from 387 experiments across 36 papers; across four evaluated models, this workflow raises mean reproduction success from 76.4% to 92.6%, narrows the cross-model spread from 21.6 to 3.9 percentage points, and lowers API spending by 19.7%.

Key Contributions
The paper introduces ReproBridge, an agent workflow that packages recovered reproduction procedures into four standardized reusable components: a refined codebase, environment setup, experiment launchers, and a reproduction guide.

ReproBridge-Bench provides 1,029 reproduction tasks from 387 experiments across 36 papers, with five human-verified reference scores per task and a separate Auditor that checks execution evidence.
Experiments across four models show that refined artifacts raise mean reproduction success from 76.4% to 92.6%, reduce downstream API spending by 19.7%, and narrow the cross-model success-rate spread from 21.6 to 3.9 percentage points; ablation results identify code and environment removal as the largest sources of loss.

#ML #Reproducibility #Research #GoogleAI #MachineLearning

[Hyper.AI @HyperAI_News]

Best gaming GPU tier list 2026 from @ScatterVolt

Who's dropping the inference tier list? 🧠

Render VS Infer

#GPU #GamingGPU #Inference #NVIDIA #AIModels #MachineLearning

[Hyper.AI @HyperAI_News]

@ashxhart

[Hyper.AI @HyperAI_News]

@melvinjohnsonp

[Hyper.AI @HyperAI_News]

@MKhalifaaaa @LuWang__ @honglaklee

[Hyper.AI @HyperAI_News]

@PaulGavrikov

[Hyper.AI @HyperAI_News]

@MiniMax_AI

[Hyper.AI @HyperAI_News]

Every CVPR Best Paper, 2016 → 2026.

2026 — Google DeepMind
2025 — Meta AI + Oxford
2024 — Google
2023 — Shanghai AI Lab
2022 — ETH Zürich + UW + Georgia Tech
2021 — MPI Tübingen
2020 — Oxford
2019 — CMU + Toronto
2018 — Stanford + Berkeley
2017 — Cornell + FAIR
2016 — Microsoft Research Asia

Who takes it in 2027?

#CVPR #CVPR2027 #ComputerVision #AIResearch #DeepLearning

[Hyper.AI @HyperAI_News]

@DamiBenveniste

[Hyper.AI @HyperAI_News]

@xiaowenhao47 The timing for promotion already passed.🤡

[Hyper.AI @HyperAI_News]

We're running it on HyperAI Notebooks — no setup, one click. 👇https://t.co/ZUX6BzVbY0

[Hyper.AI @HyperAI_News]

We ran NVIDIA's Nemotron 3.5 Transcribe 30B-A4B against Eminem's Godzilla fast verse — 7.5 words per second！

Quick context: it's NVIDIA's open-weight speech model — a 30B MoE (4B active) that transcribes with speaker labels, rich-event tagging, and audio reasoning, not just raw text.

Reference words: 196
Transcribed words: 190
Matched: 147
WER: 31.6% (42 sub / 9 ins / 11 del)

Here's the thing: these aren't dropped words. They're misheard ones.

"Minute Maid 'em" → "minimize 'em." Pun-spotting is where it slips — not keeping up with supersonic tempo. Not a single line lost at 7.5 wps.

#ASR #SpeechRecognition #Nemotron #NVIDIA #OpenWeights #LocalLLM #rap