Research Insights

Google Deepmind introduces SCoRe - a technique to teach LLMs to self correct themselves at inference time resulting in better outputs and mistakes corrections without needing human supervision.

Scaling inference time compute helps models allocate more tokens and hence come up with better generation. A direct precursor to Open AI's o1 (Strawberry)

With Iterative Preference Learning and Monte Carlo Tree search, LLMs can reason well and generate high quality outputs to related tasks. Open AI strawberry is based on this same idea.

Gemma 2 research paper by Google Deepmind detailing Knowledge Distillation, Group Query Attention, and ton of safety trainings

MUSCLE - a paper from Apple research proposes strategies for LLM updates which does not alter the model behavior in a negative manner.

This paper from Deepmind creates a LOFT benchmark to test long context vs techniques like Retrieval, In Context Learning, and SQL.

This paper from Google Deepmind scales MoE to a million experts, each with a single layer, and then uses vector based retrieval to pick the top k experts for any given query at runtime.

This paper from Google Deepmind combines Transformers with Neural Algorithmic Reaasoning resulting in an architecture where LLMs are good at reasoning tasks.

Together AI introduces a Mixture of Agents - a group of LLM experts which when mingled together can outperform GPT4 and other top LLMs

CRAG by META AI (FAIR) suggests a comprehensive evaluation framework for RAG systems - both straightforward and SOTA industry level systems.

CoPE enables LLMs get better at counting tasks by contextualizing positional encoding differently that traditional token based approaches

Analyzing LoRA to understand whether it can add new knowledge to an LLM.

MoRA builds upon the ideas of LoRA and PeFT to efficiently finetune an LLM with high rank updation using a square matrix instead of low rank matrices.

Online Iterative RLHF builds upon reward modeling as human preference approximation with iteratively changing dataset and aligning the model better than previous techniques

This Multi token prediction paper by Meta shows multi heads as memory efficient, better at performance, and faster at training compared to current next token predictors.

Open ELM by Apple is a new approach which utilizes layer wise scaling resulting in more efficient LLMs.

AutoCodeRover from NUS provides a novel framework that looks beyond code generation to genuine problem solving with the help of AI.

This paper from Open AI research talks about training an LLM to prioritize instructions in a hierarchical order, starting from system prompt, alignment, to user prompt, tool output, and so on.

Megalodon by Meta AI is a new LLM architecture that tackles the problems in transformers and can support unlimited context length using a new attention technique.

Infini-attention uses compressive memory cache and efficient retrieval to enable practically unbounded context for a transformer within a bounded memory footprint.

ReFT changes representations at different layers of an LLM by using a technique called intervention instead of changing weights/parameters using PeFT. Gives a better performance on common benchmarks and tasks.

MoD is a new Mixture of Depths implementation by Google Deepmind which dynamically allocates compute to input tokens.

With RAFT, an LLM is finetuned to ignore the distracting documents and focus on relevant information for any given task. Along with CoT, this enables model to assign correct weightage to relevant information, improving the generation and downstream tasks output significantly.

Gecko - from Google Deepmind - is a new embedding model architecture that utilizes two step LLM distillation process to create a high quality training dataset, and leads to a better model performance.

Jamba by AI21 Labs combines transformer layers with Mamba (SSM) layers and implements a MoE layers in middle to get a compute efficient model with high throughput.

Evolutionary model merge uses evolutionary algorithms that automatically discover optimal ways to combine diverse open source models. This way the resultant model harnesses the capabilities of parent models without requiring extensive additional training data or compute. This makes foundational model development more accessible and efficient.

Proposition based retrieval performs significantly better than existing techniques like paragraph based retrieval and sentence retrieval in case of RAG apps. This paper by Tencent investigates and quantifies how much better on Wikipedia articles.

PERL or Parameter Efficient Reinforcement Learning could be a groundbreaking technique to reduce memory and time consumption when it comes to aligning a model before releasing it to the world. This paper shows how using LoRA, you get close to the same benchmarks as standard RLHF techniques, and hence you get the same quality of output. Business implications are about costs, and can be done efficiently on premises on any open source model.

CaLM provides composition for LLMs similar to how libraries would in a programming language. It's a powerful method to enable combining skills of multiple LLMs depending on the use case.