@mlcmublog

Official twitter account for the ML@CMU blog @mldcmu @SCSatCMU

 Pittsburgh, PA
blog.ml.cmu.edu/
Joined February 2020
  • Tweets131
  • Following21
  • Followers2.3K
36Media
One design choice, zero new parameters: up to 49% lower forecast error and ~10% less forecast volatility. Forking-sequences encodes the whole series across all forecast creation dates in one pass. Part I: blog.ml.cmu.edu/2026/08/10/f… Part II: blog.ml.cmu.edu/2026/08/10/f…

Forking-Sequences — Part I: Statistically and Computationally Efficient Multi-Horizon Forecasting

Based on: Potosnak, W., Wolff, M., Cao, M., Ma, R., Konstantinova, T., Efimov, D., Mahoney, M.W., Oreshkin, B., & Olivares, K.G. "Forking-Sequences: Statistically and Computationally Efficient...

blog.ml.cmu.edu
2
17
1,861
Does self-supervised learning (SSL) push back against the Bitter Lesson? SSL uses downstream performance to tune pretraining design - effectively, human gradient descent over objectives. Introducing V-pretraining: let the models learn to SSL themselves. blog.ml.cmu.edu/2026/06/17/p…

Pre-Training Isn't Bitter Enough

Task construction as the control surface in continued pretraining. A construction rule (c) maps each unlabeled example (xsimmathcal D_u) into a self-supervised prediction problem ((x_c,y,m)), such as...

blog.ml.cmu.edu
4
9
885
blog.ml.cmu.edu/2026/06/19/h… Why do LLMs ace healthcare benchmarks but struggle in clinical settings? Discover the causes of this performance gap and how to bridge it in our latest post about BenchmarkCards, a documentation framework for practitioners facing deployment decisions.

Healthcare Benchmarks Are Only as Good as Their Assumptions

In healthcare settings where patients use LLMs as a medical assistant, LLM performance differs between evaluation and deployment. (a) Bean et al. (2025) find a 61 percentage point difference between...

blog.ml.cmu.edu
1
2
363
blog.ml.cmu.edu/2026/05/13/t… Before AI can generate professional videos, it needs to see like a professional. Introducing CHAI (CVPR'26 Highlight): a year of teaching AI to describe video like a filmmaker that lets an 8B model surpass GPT-5 and Gemini-3.1.

Teaching Vision-Language Models to Speak Cinema

A year of building a video caption pipeline with 100+ professional creators, and what it taught us about scaling supervision instead of models. By Zhiqiu Lin and Chancharik Mitra. Based on our CVPR...

blog.ml.cmu.edu
1
2
360
blog.ml.cmu.edu/2026/04/27/a… How good are AI systems at time-series Q&A? On ARFBench, top models hit ~63% on real incident data. But they miss different things than humans; combine both and accuracy jumps to 87%. Read more in our latest blog post!

Introducing ARFBench: A time series question-answering benchmark based on real incidents

How well do AI systems answer questions about time series data? On ARFBench, the best models reach ~63% accuracy on real incident data. But models and human experts fail on different questions:...

blog.ml.cmu.edu
1
4
2
12
2,570
Check out our latest post on CMU@ICLR 2026! blog.ml.cmu.edu/2026/04/20/c…

Carnegie Mellon at ICLR 2026

CMU researchers are presenting 194 papers at the Fourteenth International Conference on Learning Representations (ICLR 2026), held from April 23rd-April 27th at the Riocentro Convention and Event...

blog.ml.cmu.edu
1
1
13
3,005
blog.ml.cmu.edu/2026/04/13/w… When should AI agents step in—and when should they step back? A CMU + Duke team led by @FariaHuqOaishi and @ZhiruoW introduces a intervention-aware web agent. This blogpost overviews how adaptive human–agent boundaries improve collaboration.

When Should AI Step Aside?: Teaching Agents When Humans Want to Intervene

Recent advances in large language models (LLMs) have enabled AI agents to perform increasingly complex tasks in web navigation. Despite this progress, effective use of such agents continues to rely...

blog.ml.cmu.edu
2
3
1
8
1,002
blog.ml.cmu.edu/2026/03/17/l… The method used to segment textual content into ‘chunks’ in RAG pipelines can significantly impact dense retrieval quality. Read more about LumberChunker, a method for dynamically segmenting long-form narrative segments, in our new post!

LumberChunker: Long-Form Narrative Document Segmentation

Links:Paper | Code | Data LumberChunker lets an LLM decide where a long story should be split, creating more natural chunks that help Retrieval Augmented Generation (RAG) systems retrieve the right...

blog.ml.cmu.edu
1
5
514
We asked LLMs: Is Santa real? 🎅 GPT-4o says Yes at any age. Claude tells 5-year-olds the truth. What does this reveal about invisible assumptions in AI? Do LLMs believe in the tooth fairy or the Illuminati? New holiday post here: blog.ml.cmu.edu/2025/12/23/i…

Yes, AI, There is a Santa Claus

People use LLMs to ask for insight on a variety of important questions: future planning, emotional problems, scientific research. But in late December, one can expect some LLM users to be asking...

blog.ml.cmu.edu
3
6
847
LLM-as-a-judge is used everywhere, but breaks under standard forced-choice ratings. Modeling rating indeterminacy can lead to more reliable judges. Read more in our latest blog post. blog.ml.cmu.edu/2025/12/09/v…

Validating LLM-as-a-Judge Systems under Rating Indeterminacy

Figure 1: Our framework for validating LLM-as-a-judge systems under rating indeterminacy, where items in a subjective rating task can have multiple "correct" ratings. Our framework provides guidance...

blog.ml.cmu.edu
2
231
Check out our latest post on CMU @ NeurIPS 2025! blog.ml.cmu.edu/2025/12/01/c…
1
5
779
Why does LLM training plateau and how can we fix it? In a new blogpost, we discuss how we can improve LLM exploration using ideas from offline RL. blog.ml.cmu.edu/2025/11/26/h…

How to Explore to Scale RL Training of LLMs on Hard Problems?

LLM RL typically operates in one of three exploration regimes: sharpening, chaining, or guided exploration; standard RL stays in the first two and plateaus on hard problems, even those in the...

blog.ml.cmu.edu
1
2
309
blog.ml.cmu.edu/2025/10/27/l… The hardest problems have near-zero success rates and no positive examples during learning. BaNEL (Bayesian Negative Evidence Learning) post-trains using failed attempts only while minimizing the number of reward evaluations. Read more in our latest post!

Learning from Failure to Tackle Extremely Hard Problems

This blog post is based on the work BaNEL: Exploration Posteriors for Generative Modeling Using Only Negative Rewards. Tackling Very Hard Problems The ultimate aim of machine learning research is to...

blog.ml.cmu.edu
3
4
1,558
blog.ml.cmu.edu/2025/09/22/d… Check out our new blog post on "Diffusion beats Autoregressive in Data-Constrained settings". The era of infinite internet data is ending. This research paper asks:  What is the right generative modeling objective when data—not compute—is the bottleneck?

Diffusion Beats Autoregressive in Data-Constrained Settings

Check out our new blog post on "Diffusion beats Autoregressive in Data-Constrained settings". The era of infinite internet data is ending. This research paper asks:  What is the right generative...

blog.ml.cmu.edu
1
8
508
blog.ml.cmu.edu/2025/09/15/v… Check out our latest blog post on Verlog, a multi-turn reinforcement learning framework built for long-horizon LLM-agentic tasks with highly variable episode lengths.

Verlog: A Multi-turn RL framework for LLM agents

Verlog is a multi-turn reinforcement learning framework built for long-horizon LLM-agentic tasks with highly variable episode lengths. Extending VeRL and BALROG while following the proven design...

blog.ml.cmu.edu
4
1
14
897
blog.ml.cmu.edu/2025/07/08/c… Check out our latest post on CMU @ ICML 2025!

Carnegie Mellon University at ICML 2025

CMU researchers are presenting 127 papers at the Forty-Second International Conference on Machine Learning (ICML 2025), held from July 13th-19th at the Vancouver Convention Center. Here is a quick...

blog.ml.cmu.edu
7
1
16
3,687
blog.ml.cmu.edu/2025/06/01/r… In this in-depth coding tutorial, @GaoZhaolin and @g_k_swamy walk through the steps to train an LLM via RL from Human Feedback!

RLHF 101: A Technical Tutorial on Reinforcement Learning from Human Feedback

Reinforcement Learning from Human Feedback (RLHF) is a popular technique used to align AI systems with human preferences by training them using feedback from people, rather than relying solely on...

blog.ml.cmu.edu
9
1
27
4,679
blog.ml.cmu.edu/2025/05/22/u… Are your LLMs truly forgetting unwanted data?  In this new blog post authored by @shengyuan_26734, Yiwei Fu, @zstevenwu, and @gingsmith, we discuss how benign relearning can jog unlearned LLM's memory to recover knowledge that is supposed to be forgotten.

Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning

Machine unlearning is a promising approach to mitigate undesirable memorization of training data in ML models. In this post, we will discuss our work (which appeared at ICLR 2025) demonstrating that...

blog.ml.cmu.edu
4
6
1,053
Load more