There's a 2025 ICML paper (Kim et al.) showing LLMs give the same wrong answer 60% of the time when both err — bigger models are more correlated, not less.
November 3, 2026 at 4:23 PM
There's a 2025 ICML paper (Kim et al.) showing LLMs give the same wrong answer 60% of the time when both err — bigger models are more correlated, not less.
Rare to see Tsinghua, Tencent and Stanford on one author list — ICML-reviewed and open-sourced, too. But Tencent builds the models being judged, and open reasoning evals contaminate fast. My bet: decorative within two training cycles — anyone running it will say sooner.
ICML 2025: Tsinghua, Tencent, Stanford and others jointly release RBench, a complex reasoning benchmark for large models
Tsinghua University's Jittor team, Tencent Hunyuan X, Stanford University, Carnegie Mellon University, the University of Pennsylvania and Feishi Technology have released RBench, a multidisciplinary, multilingual benchmark for evaluating complex reasoning in language and multimodal models. It contains 1,094 text questions and 665 multimodal questions across more than 100 disciplines and is open-sourced on Hugging Face.
aichina.news
November 1, 2026 at 6:39 PM
Rare to see Tsinghua, Tencent and Stanford on one author list — ICML-reviewed and open-sourced, too. But Tencent builds the models being judged, and open reasoning evals contaminate fast. My bet: decorative within two training cycles — anyone running it will say sooner.
Coauthors included CDS Assoc. Profs Rajesh Ranganath and He He.
The paper received a Spotlight presentation at ICML 2026.
The paper received a Spotlight presentation at ICML 2026.
October 28, 2026 at 3:41 PM
Coauthors included CDS Assoc. Profs Rajesh Ranganath and He He.
The paper received a Spotlight presentation at ICML 2026.
The paper received a Spotlight presentation at ICML 2026.
New ICML 2026 position paper with Ashton Anderson, Harsh Kumar & Karina Vold: chat LLMs are mostly trained on next-turn approval. For socioemotional use, we argue for an opt-in well-being mode trained and evaluated on longer-horizon outcomes. Three tensions: When, Who, How.
Trained to Please You or Help You? LLMs for Well-Being (4-min explainer)
AI-narrated explainer of Anderson, Kumar, Tay & Vold (ICML 2026).
youtu.be
October 28, 2026 at 12:31 PM
New ICML 2026 position paper with Ashton Anderson, Harsh Kumar & Karina Vold: chat LLMs are mostly trained on next-turn approval. For socioemotional use, we argue for an opt-in well-being mode trained and evaluated on longer-horizon outcomes. Three tensions: When, Who, How.
JoBS Co-Optimizes LLM Data and Model Config Without Full Training Runs
Training LLMs to pick LLM configs? JoBS skips full runs using a cheap predictor instead.
#AI #AIResearch #MachineLearning
https://autonainews.com/jobs-co-optimizes-llm-data-and-model-config-without-full-training-runs/
Training LLMs to pick LLM configs? JoBS skips full runs using a cheap predictor instead.
#AI #AIResearch #MachineLearning
https://autonainews.com/jobs-co-optimizes-llm-data-and-model-config-without-full-training-runs/
JoBS Co-Optimizes LLM Data and Model Config Without Full Training Runs
Four new frameworks presented at ICML 2026 and ICLR 2026 are converging on the same idea: stop running expensive LLM training jobs to evaluate each candidate configuration, and use the LLM itself to …
autonainews.com
October 31, 2026 at 8:24 PM
JoBS Co-Optimizes LLM Data and Model Config Without Full Training Runs
Training LLMs to pick LLM configs? JoBS skips full runs using a cheap predictor instead.
#AI #AIResearch #MachineLearning
https://autonainews.com/jobs-co-optimizes-llm-data-and-model-config-without-full-training-runs/
Training LLMs to pick LLM configs? JoBS skips full runs using a cheap predictor instead.
#AI #AIResearch #MachineLearning
https://autonainews.com/jobs-co-optimizes-llm-data-and-model-config-without-full-training-runs/
More suggestions: two tracks for reviewing like ICML for no-AI and limited-AI, any review assistant tools should be tested robustly before official use or enforcement, two phase reviewing like AAAI (quick review first for quality filtering), adopt policies from TMLR (suspensions, etc.)
October 26, 2026 at 1:30 PM
More suggestions: two tracks for reviewing like ICML for no-AI and limited-AI, any review assistant tools should be tested robustly before official use or enforcement, two phase reviewing like AAAI (quick review first for quality filtering), adopt policies from TMLR (suspensions, etc.)
I don’t usually teach classes where this kind of text fits well. Either I’m teaching lit/intellectual history, or I’m teaching how to do specific things with machine learning; “social impact thereof” is not my usual beat.
That said, I would recommend this Naranayan talk to freaked out students
That said, I would recommend this Naranayan talk to freaked out students
What will be left for us to work on?
My keynote at ICML 2026
www.normaltech.ai
October 25, 2026 at 12:29 AM
I don’t usually teach classes where this kind of text fits well. Either I’m teaching lit/intellectual history, or I’m teaching how to do specific things with machine learning; “social impact thereof” is not my usual beat.
That said, I would recommend this Naranayan talk to freaked out students
That said, I would recommend this Naranayan talk to freaked out students
LLM Impact on peer review:
- TMLR increases the importance of clarity
- ICML introduces desk rejections by ACs
- RSS introduces 2 phases & "signed reviewers"
- Some journals temp. suspend submissions
- ArXiv limits # submissions, bans authors on hallucinations, requ. surveys to be peer-reviewed
- TMLR increases the importance of clarity
- ICML introduces desk rejections by ACs
- RSS introduces 2 phases & "signed reviewers"
- Some journals temp. suspend submissions
- ArXiv limits # submissions, bans authors on hallucinations, requ. surveys to be peer-reviewed
October 22, 2026 at 11:40 AM
LLM Impact on peer review:
- TMLR increases the importance of clarity
- ICML introduces desk rejections by ACs
- RSS introduces 2 phases & "signed reviewers"
- Some journals temp. suspend submissions
- ArXiv limits # submissions, bans authors on hallucinations, requ. surveys to be peer-reviewed
- TMLR increases the importance of clarity
- ICML introduces desk rejections by ACs
- RSS introduces 2 phases & "signed reviewers"
- Some journals temp. suspend submissions
- ArXiv limits # submissions, bans authors on hallucinations, requ. surveys to be peer-reviewed
AIの権威ヤン・ルカン氏「LLMは現実世界を理解できない」—AnthropicはAGIの証拠を発見｜2026-07-08
finance.biggo.jp/news/83af983...
finance.biggo.jp/news/83af983...
AIの権威ヤン・ルカン氏「LLMは現実世界を理解できない」—AnthropicはAGIの証拠を発見 — BigGo ファイナンス
AMI Labsのパスカル・フォンCRIOがICML…
finance.biggo.jp
October 24, 2026 at 9:17 PM
AIの権威ヤン・ルカン氏「LLMは現実世界を理解できない」—AnthropicはAGIの証拠を発見｜2026-07-08
finance.biggo.jp/news/83af983...
finance.biggo.jp/news/83af983...
Had an insightful conversation with Mathieu at ICML, he explained common-noise games and their challenges.
In this paper, it's cool that the authors use Bayesian identification theory and thus, adapt learning algorithms, to recover the observations.
In this paper, it's cool that the authors use Bayesian identification theory and thus, adapt learning algorithms, to recover the observations.
Gr\'egoire Lambrecht, Mathieu Lauri\`ere
Bayesian Learning of Parametric Mean Field Games with Common Noise
https://arxiv.org/abs/2610.23387
Bayesian Learning of Parametric Mean Field Games with Common Noise
https://arxiv.org/abs/2610.23387
October 20, 2026 at 8:36 AM
Had an insightful conversation with Mathieu at ICML, he explained common-noise games and their challenges.
In this paper, it's cool that the authors use Bayesian identification theory and thus, adapt learning algorithms, to recover the observations.
In this paper, it's cool that the authors use Bayesian identification theory and thus, adapt learning algorithms, to recover the observations.
The training is also very simple and is similar to the self-conditioning in RIN (Jabri ICML'23). You do a 1st forward with 0, get your predicted DINO features (detached) and pass them as additional condition to the 2nd forward+backward pass
Since there is only 1 ODE, there is only 1 noise schedule
Since there is only 1 ODE, there is only 1 noise schedule
October 20, 2026 at 8:40 AM
The training is also very simple and is similar to the self-conditioning in RIN (Jabri ICML'23). You do a 1st forward with 0, get your predicted DINO features (detached) and pass them as additional condition to the 2nd forward+backward pass
Since there is only 1 ODE, there is only 1 noise schedule
Since there is only 1 ODE, there is only 1 noise schedule
PhysForge | 54 stars | 0 forks
[ICML 2026] PhysForge: Generating Physics-Grounded 3D Assets for Interactive Virtual World
[ICML 2026] PhysForge: Generating Physics-Grounded 3D Assets for Interactive Virtual World
PhysForge - Trending on te9.dev
[ICML 2026] PhysForge: Generating Physics-Grounded 3D Assets for Interactive Virtual World
repotrend.com
October 19, 2026 at 8:34 PM
PhysForge | 54 stars | 0 forks
[ICML 2026] PhysForge: Generating Physics-Grounded 3D Assets for Interactive Virtual World
[ICML 2026] PhysForge: Generating Physics-Grounded 3D Assets for Interactive Virtual World
仮にもテックライターなのだから、LLMや人工知能研究の論文嫁とまでは言わないけど、ヤン・ルカンの発言ぐらいは調べろよといつも思う。
finance.biggo.jp/news/83af983...
『彼は2023年から一貫して「ChatGPT、Claude、Geminiは人間レベルの知能に至る道ではない」と主張してきた。
ルカン教授の論理は明確だ。試験問題は言語化され離散化された課題であり、LLMが最も得意とする領域に過ぎない。真の知能は知覚、物理的直観、因果理解など、常識から生まれる能力を含まなければならない。』
finance.biggo.jp/news/83af983...
『彼は2023年から一貫して「ChatGPT、Claude、Geminiは人間レベルの知能に至る道ではない」と主張してきた。
ルカン教授の論理は明確だ。試験問題は言語化され離散化された課題であり、LLMが最も得意とする領域に過ぎない。真の知能は知覚、物理的直観、因果理解など、常識から生まれる能力を含まなければならない。』
AIの権威ヤン・ルカン氏「LLMは現実世界を理解できない」—AnthropicはAGIの証拠を発見 — BigGo ファイナンス
AMI Labsのパスカル・フォンCRIOがICML…
finance.biggo.jp
October 20, 2026 at 11:35 PM
仮にもテックライターなのだから、LLMや人工知能研究の論文嫁とまでは言わないけど、ヤン・ルカンの発言ぐらいは調べろよといつも思う。
finance.biggo.jp/news/83af983...
『彼は2023年から一貫して「ChatGPT、Claude、Geminiは人間レベルの知能に至る道ではない」と主張してきた。
ルカン教授の論理は明確だ。試験問題は言語化され離散化された課題であり、LLMが最も得意とする領域に過ぎない。真の知能は知覚、物理的直観、因果理解など、常識から生まれる能力を含まなければならない。』
finance.biggo.jp/news/83af983...
『彼は2023年から一貫して「ChatGPT、Claude、Geminiは人間レベルの知能に至る道ではない」と主張してきた。
ルカン教授の論理は明確だ。試験問題は言語化され離散化された課題であり、LLMが最も得意とする領域に過ぎない。真の知能は知覚、物理的直観、因果理解など、常識から生まれる能力を含まなければならない。』
Benchmark it thoroughly, write a solid paper + blog post and send those to Arxiv and ICLR or ICML, publish videos of it in action, then when you get traction, decide how to let it out into the world.
October 19, 2026 at 10:56 AM
Benchmark it thoroughly, write a solid paper + blog post and send those to Arxiv and ICLR or ICML, publish videos of it in action, then when you get traction, decide how to let it out into the world.
New ICML 2026 research, 'Prompt Injection as Role Confusion': a model judges whether text is trusted reasoning by how it sounds, not where it's from.
CoT forgery pastes fake reasoning into a prompt so the model thinks its guardrail already said yes. ~60% on frontier models, authors' figure.
CoT forgery pastes fake reasoning into a prompt so the model thinks its guardrail already said yes. ~60% on frontier models, authors' figure.
October 20, 2026 at 4:50 PM
New ICML 2026 research, 'Prompt Injection as Role Confusion': a model judges whether text is trusted reasoning by how it sounds, not where it's from.
CoT forgery pastes fake reasoning into a prompt so the model thinks its guardrail already said yes. ~60% on frontier models, authors' figure.
CoT forgery pastes fake reasoning into a prompt so the model thinks its guardrail already said yes. ~60% on frontier models, authors' figure.
On the Origins of Linear Representations in Large Language Models (2024, ICML) proceedings.mlr.press/v235/jiang24...
On the Origins of Linear Representations in Large Language Models
An array of recent works have argued that high-level semantic concepts are encoded
proceedings.mlr.press
October 17, 2026 at 7:57 PM
On the Origins of Linear Representations in Large Language Models (2024, ICML) proceedings.mlr.press/v235/jiang24...
AI likes Flowers ;)
Shayne Longpre and Seungone Kim ran an AI-based review and ranking of all papers presented at ICML 2026. Flowers (our new architecture for neural simulation) are number 10!
research-taste.github.io/ai-as-icml-a...
t-muser.github.io/flowers/
Shayne Longpre and Seungone Kim ran an AI-based review and ranking of all papers presented at ICML 2026. Flowers (our new architecture for neural simulation) are number 10!
research-taste.github.io/ai-as-icml-a...
t-muser.github.io/flowers/
AI as Area Chair | ICML 2026
A paper-only AI ranking of the ICML 2026 accepted corpus.
research-taste.github.io
October 16, 2026 at 2:31 PM
AI likes Flowers ;)
Shayne Longpre and Seungone Kim ran an AI-based review and ranking of all papers presented at ICML 2026. Flowers (our new architecture for neural simulation) are number 10!
research-taste.github.io/ai-as-icml-a...
t-muser.github.io/flowers/
Shayne Longpre and Seungone Kim ran an AI-based review and ranking of all papers presented at ICML 2026. Flowers (our new architecture for neural simulation) are number 10!
research-taste.github.io/ai-as-icml-a...
t-muser.github.io/flowers/
Works like EnbPI (Xu & Xie, ICML 2021), Conformal Time-series Forecasting (Stankeviciute, Alaa & van der Schaar, NeurIPS 2021), Adaptive Conformal Inference (Gibbs & Candès), and the non-exchangeable extensions that followed showed how to recover reliable prediction intervals for sequential
October 17, 2026 at 1:00 PM
Works like EnbPI (Xu & Xie, ICML 2021), Conformal Time-series Forecasting (Stankeviciute, Alaa & van der Schaar, NeurIPS 2021), Adaptive Conformal Inference (Gibbs & Candès), and the non-exchangeable extensions that followed showed how to recover reliable prediction intervals for sequential
As @beenwrekt.bsky.social writes, this issue is particularly acute in AI research, where top conferences like NeurIPS and ICML are flooded with paper submissions and researchers commonly use genAI to review others'work. But the issue is real and growing in many other fields.
The Least Agentic People Alive
The arms race of deferring agency and the dark acquiescence to robotic bureaucracy.
www.argmin.net
October 15, 2026 at 12:36 PM
As @beenwrekt.bsky.social writes, this issue is particularly acute in AI research, where top conferences like NeurIPS and ICML are flooded with paper submissions and researchers commonly use genAI to review others'work. But the issue is real and growing in many other fields.
Position: We Need Large Language Models Optimized For Our Well-Being (Anderson, Kumar, Tay & Vold, ICML 2026)
Models are largely trained to win immediate approval. The authors call for an opt-in mode optimized for long-horizon well-being.
AI-narrated. Full video: youtu.be/FC-oT0_ryyM
Models are largely trained to win immediate approval. The authors call for an opt-in mode optimized for long-horizon well-being.
AI-narrated. Full video: youtu.be/FC-oT0_ryyM
October 16, 2026 at 1:51 PM
Position: We Need Large Language Models Optimized For Our Well-Being (Anderson, Kumar, Tay & Vold, ICML 2026)
Models are largely trained to win immediate approval. The authors call for an opt-in mode optimized for long-horizon well-being.
AI-narrated. Full video: youtu.be/FC-oT0_ryyM
Models are largely trained to win immediate approval. The authors call for an opt-in mode optimized for long-horizon well-being.
AI-narrated. Full video: youtu.be/FC-oT0_ryyM
🤖 AI Takes on Document Review to Ease Validation Bottleneck
The capacity gap is concrete enough to be measured. In ICML 2026, 23,918 submissions arrived at venues that accepted 6,352 papers, roughly double the...
#BenchmarksEvaluation #PolicyRegulation #SafetyAlignment #AI #AIPulse
The capacity gap is concrete enough to be measured. In ICML 2026, 23,918 submissions arrived at venues that accepted 6,352 papers, roughly double the...
#BenchmarksEvaluation #PolicyRegulation #SafetyAlignment #AI #AIPulse
Read the full article →
www.synestesia.uk
October 15, 2026 at 6:36 AM
🤖 AI Takes on Document Review to Ease Validation Bottleneck
The capacity gap is concrete enough to be measured. In ICML 2026, 23,918 submissions arrived at venues that accepted 6,352 papers, roughly double the...
#BenchmarksEvaluation #PolicyRegulation #SafetyAlignment #AI #AIPulse
The capacity gap is concrete enough to be measured. In ICML 2026, 23,918 submissions arrived at venues that accepted 6,352 papers, roughly double the...
#BenchmarksEvaluation #PolicyRegulation #SafetyAlignment #AI #AIPulse
Another highlight from today’s Research@ event: After successful deployments across 16K+ submissions at STOC, ICML, NeurIPS, & ICLR, learn how Deep Review is expanding from scientific peer review to provide rigorous, grounded verification for knowledge documents, including Google Docs & presos.
October 14, 2026 at 7:08 PM
Another highlight from today’s Research@ event: After successful deployments across 16K+ submissions at STOC, ICML, NeurIPS, & ICLR, learn how Deep Review is expanding from scientific peer review to provide rigorous, grounded verification for knowledge documents, including Google Docs & presos.
Google Research launches Deep Review, an agentic AI reviewer for technical documents
Google Research has released Deep Review publicly, an agentic system that checks logic, math and facts in long documents.
Google Research has released Deep Review publicly, an agentic system that checks logic, math and facts in long documents.
Google Research launches Deep Review, an agentic AI reviewer for technical documents
Google Research has released Deep Review publicly, an agentic system that checks logic, math and facts in long documents. It grew out of pilots with STOC, ICML, NeurIPS and ICLR covering more than 16,000 manuscripts.
glonce.com
October 15, 2026 at 4:54 AM
Google Research launches Deep Review, an agentic AI reviewer for technical documents
Google Research has released Deep Review publicly, an agentic system that checks logic, math and facts in long documents.
Google Research has released Deep Review publicly, an agentic system that checks logic, math and facts in long documents.