Summaries

Paper summaries

Reasoning
project thumbnail

Chain of Images for Intuitively Reasoning

This paper proposes a Chain-of-Images (CoI) method for multimodal models to solve reasoning problems by generating a series of images as intermediate representations, using a Symbolic Multimodal Large Language Model (SyMLLM).
project thumbnail

Dream 7B Diffusion Large Language Models

Dream 7B is a 7-billion parameter diffusion language model that refines text in parallel, initialized from an autoregressive model and trained using context-adaptive noise scheduling.
project thumbnail

Imagine while Reasoning in Space Multimodal Visualization-of-Thought

Multimodal Visualization-of-Thought (MVoT) is proposed to enable Multimodal Large Language Models (MLLMs) to generate interleaved verbal and visual reasoning traces for spatial reasoning tasks.
project thumbnail

LLM Latent Reasoning as Chain of Superposition

Train an encoder that summarizes reasoning chunks. Then train a latent reasoning model on the summaries it produces from some CoT data.
project thumbnail

LLMs are Single-threaded Reasoners, Demystifying the Working Mechanism of Soft Thinking

Vanilla Soft Thinking pushes the model to the greedy token sampling internally. They showed that the model usually continues to work only with the most probable next token. To mitigate this issue they suggest adding noise to logits and get better performance.
project thumbnail

Latent Chain-of-Thought? Decoding the Depth-Recurrent Transformer

This work investigates the internal mechanics of the Huginn depth-recurrent Transformer on arithmetic tasks to find evidence for latent chain-of-thought reasoning.
project thumbnail

Latent Sketchpad Sketching Visual Thoughts to Elicit Multimodal Reasoning in MLLMs

A framework that equips Multimodal Large Language Models with an internal visual scratchpad to generate visual latents during autoregressive reasoning, which can be translated into interpretable sketches.
project thumbnail

Multiplex Thinking Reasoning via Token wise Branch and Merge

Make soft-thinking a bit random. Then train with GRPO.
project thumbnail

Reasoning Within the Mind Dynamic Multimodal Interleaving in Latent Space

Training-free latent reasoning. Optimize latent reasoning tokens to maximize model confidence, which correlates with accuracy.
project thumbnail

Reinforcement Learning via Self-Distillation

Self-Distillation Policy Optimization (SDPO) leverages a model's in-context learning capabilities to convert tokenized environment feedback into dense, logit-level reinforcement learning signals.
project thumbnail

Scaling Latent Reasoning via Looped Language Models

Training a recurrent reasoning model. Looping the same models over and over again.
project thumbnail

Soft Tokens, Hard Truths

They add Gaussian noise to the soft-thinking embeddings, then train with RL using RLOO.
project thumbnail

Zebra-CoT A Dataset for Interleaved Vision Language Reasoning

Zebra-CoT is a large-scale dataset with 182,384 interleaved text-image reasoning traces across 18 domains for training multimodal models.

Hallucination
project thumbnail

DASH Detection and Assessment of Systematic Hallucinations of VLMs

Make a dataset that VLMs hallucinate and wrongly think things exist in images

Interpretability
project thumbnail

A Mathematical Framework for Transformer Circuits

In Transformers residual stream is the main object and layers read and write from/to it.
project thumbnail

An Overview of Early Vision in InceptionV1

inceptionV1 feature maps of different layers
project thumbnail

CLIP-Dissect Automatic Description of Neuron Representations

Find concepts that activates a neuron using a image dataset
project thumbnail

Deep Learning is Not So Mysterious or Different

??
project thumbnail

Physics of Language Models

Understanding LLMs by training smaller LMs in controlled environment
project thumbnail

Scaling Monosemanticity Extracting Interpretable Features from Claude 3 Sonnet

Scale SAE to Claude 3 Sonnet
project thumbnail

Towards Monosemanticity Decomposing Language Models With Dictionary Learning

How SAE works
project thumbnail

Zoom In An Introduction to Circuits

Investigate Vision Circuits by Studying the Connections between Neurons
project thumbnail

Can Large Language Models Explain Their Internal Mechanisms?

summary of Can Large Language Models Explain Their Internal Mechanisms?
project thumbnail

Emergent World Representations Exploring a Sequence Model Trained on a Synthetic Task

summary of Emergent World Representations Exploring a Sequence Model Trained on a Synthetic Task
project thumbnail

Interpretability Beyond Feature Attribution Quantitative Testing with Concept Activation Vectors (TCAV)

summary of Interpretability Beyond Feature Attribution Quantitative Testing with Concept Activation Vectors (TCAV)

ActiveLearning
project thumbnail

Every decision tree has an influential variable

title is self-explanatory
project thumbnail

Learning decision trees from random examples

Decision tree learning By Finding Consistent Decision Trees
project thumbnail

Leveraged volume sampling for linear regression

Active Learning in linear regression with multiplicative error rate bounds
project thumbnail

Properly learning decision trees in almost polynomial time

learning a decision tree for unifrom random data distribution in O(s ^ log(log(s)))
project thumbnail

Top-down induction of decision trees- rigorous guarantees and inherent limitations

greedily learn a decision tree based on the most inflouential variables in all leaves.
project thumbnail

Active Learning Survey

Active Learning for Agnostic classification
project thumbnail

The True Sample Complexity of Active Learning

A different definition of active learning label complexity

Other Summaries

project thumbnail

Labeling Neural Representations with Inverse Recognition

summary of Labeling Neural Representations with Inverse Recognition
project thumbnail

Progress measures for grokking via mechanistic interpretability

summary of Progress measures for grokking via mechanistic interpretability
project thumbnail

What do we learn from inverting CLIP models?

summary of What do we learn from inverting CLIP models?