inMind

Our latest Google Labs Experiment is here! #GenChess turns your ideas into playable art pieces using Google’s Imagen 3 model. Create and play today → labs.google/genchess

Can LLMs reason effectively without prompting? Great paper by @GoogleDeepMind By considering multiple paths during decoding, LLMs show improved reasoning without special prompts. It reveals LLMs' natural reasoning capabilities. LLMs can reason better by exploring multiple decoding paths instead of just one. The paper finds that, CoT (Chain-of-Thought) reasoning paths can be elicited from pre-trained LLMs by simply altering the decoding process. Original Problem 🔍: LLMs struggle with reasoning tasks without prompting, relying on human-designed prompts or fine-tuning. ----- Solution in this Paper 🧠: • Introduces Chain-of-Thought (CoT) decoding • Explores alternative top-k tokens during decoding • Utilizes model's confidence in final answer to select reliable paths • Combines CoT-decoding with existing prompting techniques ----- Key Insights from this Paper 💡: • Pre-trained LLMs possess inherent reasoning capabilities • CoT paths exist but are often ranked lower in decoding space • Presence of CoT correlates with higher model confidence • CoT-decoding partially closes the gap between pre-trained and instruction-tuned models • Task difficulty influences the presence of correct CoT paths ----- Results 📊: • CoT-decoding significantly improves reasoning performance across model families • PaLM-2 Large: GSM8K accuracy increases from 34.8% to 63.2% • Mistral-7B: GSM8K accuracy rises from 9.9% to 25.1% • Year Parity task: Near-perfect accuracy achieved at larger model scales • Combining CoT-decoding with zero-shot CoT prompting further boosts performance