LLMs Encode How Difficult Problems Are
This research addresses the inconsistency in LLM performance for AI practitioners by showing that human difficulty annotations can enhance model generalization, though it is incremental as it builds on existing probing and training methods.
The study investigated whether large language models internally encode problem difficulty in alignment with human judgment, finding that human-labeled difficulty is strongly decodable and correlates with improved accuracy and reduced hallucination, while model-derived difficulty scales poorly and negatively correlates with performance during training.
Large language models exhibit a puzzling inconsistency: they solve complex problems yet frequently fail on seemingly simpler ones. We investigate whether LLMs internally encode problem difficulty in a way that aligns with human judgment, and whether this representation tracks generalization during reinforcement learning post-training. We train linear probes across layers and token positions on 60 models, evaluating on mathematical and coding subsets of Easy2HardBench. We find that human-labeled difficulty is strongly linearly decodable (AMC: $ρ\approx 0.88$) and exhibits clear model-size scaling, whereas LLM-derived difficulty is substantially weaker and scales poorly. Steering along the difficulty direction reveals that pushing models toward "easier" representations reduces hallucination and improves accuracy. During GRPO training on Qwen2.5-Math-1.5B, the human-difficulty probe strengthens and positively correlates with test accuracy across training steps, while the LLM-difficulty probe degrades and negatively correlates with performance. These results suggest that human annotations provide a stable difficulty signal that RL amplifies, while automated difficulty estimates derived from model performance become misaligned precisely as models improve. We release probe code and evaluation scripts to facilitate replication.