gpt-4 turbo is the magnus carlsen of language models (highest ELO)
LLMS
-
Yann LeCun Revisits 2019 Predictions on LLM Limitations
By
–
Nice to see Yann LeCun 2024 making yet another set of points (re the intractability of predicting video from pixel space and the limits of LLMs) that I made in 2019 and 2020. He’s catching up!
-
Model Merging and Pretraining Boost AI Performance Results
By
–
Beyond just merging, we've boosted performance by continuing to pretrain with a massive dataset. Merging alone doesn't cut it for top-notch results.
-
Critiquing LLM Hype: Pre-existing Concerns About Language Model Limits
By
–
This particular idea long precedes me; I was indicating the grief that I got, not at all saying I originated it. it’s not my area. I would note though that many of my points about hype and the limits of large language models long preceded their work, which they have never
-
AgentBoard: Benchmark Framework for LLM Agent Evaluation
By
–
10/ AgentBoard – a benchmark with an open-source evaluation framework to perform analytical evaluation of LLM agents; assesses the capabilities and limitations of LLM agents and demystifies agent behaviors which leads to building stronger LLM agents.
-
Medusa Framework Achieves 2.2x LLM Inference Speedup
By
–
9/ Medusa – a framework for LLM inference acceleration using multiple decoding heads; substantially reduces the number of decoding steps and achieves over 2.2x speedup without compromising generation quality.
-
Resource-Efficient LLMs and Multimodal Models: Architecture and Implementation
By
–
6/ Resource-efficient LLMs & Multimodal Models – provides a comprehensive analysis and insights into ML efficiency research, including architectures, algorithms, and practical system designs and implementations.
-
Red Teaming Visual Language Models for Safety Alignment
By
–
7/ Red Teaming Visual Language Models – finds that 10 prominent open-sourced VLMs struggle with red teaming and have up to 31% performance gap with GPT-4V; applies red teaming alignment to LLaVA-v1.5 with SFT to improve performance by 10%.
-
MambaByte: SSM learns raw bytes, outperforms Transformers
By
–
3/ MambaByte – adapts Mamba SSM to learn directly from raw bytes; bytes lead to longer sequences which autoregressive Transformers will scale poorly on; reports huge benefits related to faster inference and even outperforms subword Transformers.
-
WARM: Weighted Averaged Rewards Models Improve LLM Alignment
By
–
5/ WARM – introduces weighted averaged rewards models (WARM) that involve fine-tuning multiple rewards models and then averaging them in the weight space; improves efficiency while improving the quality and alignment of LLM predictions.