7 papers
ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling
Vaibhav Singh, Soumya Suvra Ghosal, Sarvesh Gharat +3
Large Reasoning Models (LRMs) improve performance by allocating additional inference-time compute to generate extended chain-of-thought reasoning. However, recent studies reveal th…
Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away
Soumya Suvra Ghosal, Souradip Chakraborty, Vaibhav Singh +3
Reinforcement learning (RL) based post-training for explicit chain-of-thought (e.g., GRPO) improves the reasoning ability of multimodal large-scale reasoning models (MLRMs). But re…
Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models
Soumya Suvra Ghosal, Souradip Chakraborty, Avinash Reddy +6
Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek R1) have led to a popular belief that extending thinking traces using prompts like "Wait" or "Let…
Engagement Undermines Safety: How Stereotypes and Toxicity Shape Humor in Language Models
Atharvan Dogra, Soumya Suvra Ghosal, Ameet Deshpande +2
Large language models are increasingly used for creative writing and engagement content, raising safety concerns about the outputs. Therefore, casting humor generation as a testbed…
Relic: Enhancing Reward Model Generalization for Low-Resource Indic Languages with Few-Shot Examples
Soumya Suvra Ghosal, Vaibhav Singh, Akash Ghosh +4
Reward models are essential for aligning large language models (LLMs) with human preferences. However, most open-source multilingual reward models are primarily trained on preferen…
Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time
Mohamad Chehade, Soumya Suvra Ghosal, Souradip Chakraborty +4
Aligning large language models with humans is challenging due to the inherently multifaceted nature of preference feedback. While existing approaches typically frame this as a mult…