Picture for Samy Jelassi

Samy Jelassi

DMA, CIMS

Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining

Add code
Apr 10, 2025
Viaarxiv icon

To Backtrack or Not to Backtrack: When Sequential Search Limits Model Reasoning

Add code
Apr 09, 2025
Viaarxiv icon

The Role of Sparsity for Length Generalization in Transformers

Add code
Feb 24, 2025
Viaarxiv icon

Collective Model Intelligence Requires Compatible Specialization

Add code
Nov 04, 2024
Figure 1 for Collective Model Intelligence Requires Compatible Specialization
Figure 2 for Collective Model Intelligence Requires Compatible Specialization
Figure 3 for Collective Model Intelligence Requires Compatible Specialization
Figure 4 for Collective Model Intelligence Requires Compatible Specialization
Viaarxiv icon

Mixture of Parrots: Experts improve memorization more than reasoning

Add code
Oct 24, 2024
Figure 1 for Mixture of Parrots: Experts improve memorization more than reasoning
Figure 2 for Mixture of Parrots: Experts improve memorization more than reasoning
Figure 3 for Mixture of Parrots: Experts improve memorization more than reasoning
Figure 4 for Mixture of Parrots: Experts improve memorization more than reasoning
Viaarxiv icon

LoRA Soups: Merging LoRAs for Practical Skill Composition Tasks

Add code
Oct 16, 2024
Figure 1 for LoRA Soups: Merging LoRAs for Practical Skill Composition Tasks
Figure 2 for LoRA Soups: Merging LoRAs for Practical Skill Composition Tasks
Figure 3 for LoRA Soups: Merging LoRAs for Practical Skill Composition Tasks
Figure 4 for LoRA Soups: Merging LoRAs for Practical Skill Composition Tasks
Viaarxiv icon

Universal Length Generalization with Turing Programs

Add code
Jul 03, 2024
Viaarxiv icon

How Does Overparameterization Affect Features?

Add code
Jul 01, 2024
Viaarxiv icon

Q-Probe: A Lightweight Approach to Reward Maximization for Language Models

Add code
Feb 22, 2024
Figure 1 for Q-Probe: A Lightweight Approach to Reward Maximization for Language Models
Figure 2 for Q-Probe: A Lightweight Approach to Reward Maximization for Language Models
Figure 3 for Q-Probe: A Lightweight Approach to Reward Maximization for Language Models
Figure 4 for Q-Probe: A Lightweight Approach to Reward Maximization for Language Models
Viaarxiv icon

Repeat After Me: Transformers are Better than State Space Models at Copying

Add code
Feb 01, 2024
Viaarxiv icon