finding

active

finding:vs-cot-negative-synthetic-data-achieves-36-81-accuracy-on-gsm8k-test-set-after-offline-rl-vs-34-12-with-positive-only-training

VS-CoT negative synthetic data achieves 36.81% accuracy on GSM8K test set after offline RL, vs. 34.12% with positive-only training

Demonstrates VS-generated diverse negative examples improve downstream model performance in offline RL

Source paper

extracted_from

Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

(2025) · Jiayi Zhang · Simon C.H. Yu · Derek Chong · Anthony Sicilia +3

Neighborhood — ranked by edge-count

Papers (1)

paper

Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
mentions

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

VS-CoT achieves the highest Top@1 accuracy (0.348) and Pass@N accuracy (0.485) on SimpleQA among all methods testedfinding0.821
Shows VS not only maintains but can slightly improve factual accuracy compared to baseline methods
VS-CoT with GPT-4.1 as generator achieves 45.9% accuracy on Qwen3-1.7B-Base fine-tuned on math benchmarks, the highest single resultfinding0.803
Best performing VS variant for math synthetic data generation with GPT-4.1
Multimodal-CoT trained with InstructBLIP/ChatGPT-generated rationales achieves 87.76% accuracy on ScienceQA, comparable to human-annotated rationale performance of 90.45%finding0.784
Evidence that Multimodal-CoT can operate without human-annotated reasoning chains by using large models to generate pseudo-rationales.
QwQ-32B accuracy on GSM8k remains between 96.36% and 96.50% across all intervention strengths (-0.96 to +0.48)finding0.771
Demonstrates that stronger models are largely insensitive to reflection manipulation
One-stage CoT (QCM→RA) shows 12.31% accuracy drop vs. no-CoT (QCM→A) on ScienceQA; two-stage framework (rationale generation + answer inference) achieves 85.31% accuracy with vision featuresfinding0.762
Empirical evidence that naive one-stage CoT fails in language-only setting; two-stage + vision achieves state-of-the-art.
VS-Multi achieves average accuracy of 37.5% on math benchmarks (avg of MATH500, OlympiadBench, Minerva) with Gemini-2.5-Flash as generator and Qwen3-4B as SFT model, vs. 40.7% baselinefinding0.755
Best VS result in synthetic data generation for math, demonstrating downstream improvement through diversity
Activation probing detects final answer belief earlier in CoT than CoT monitor on both models, with especially pronounced gap on easy MMLU questionsfinding0.753
Comparative finding establishing activation probing as superior to text-level monitoring for early belief detection
90.45% accuracy on ScienceQA benchmark with Multimodal-CoT Large (738M parameters)finding0.753
State-of-the-art result on ScienceQA; represents +3.91% improvement over prior best published result of 86.54%.