Study on the Binary Separation Phenomenon of Evidence Sufficiency in Hidden States of Reasoning Models

This paper explores the binary separation phenomenon of evidence sufficiency in the hidden states of reasoning models when dealing with multi-hop question answering tasks with fixed questions and varying contexts (the "sufficient state" when evidence is adequate and the "insufficient state" when evidence is lacking). Experiments verify that this phenomenon is a universal mechanism of reasoning models, revealing its causal role and providing a new perspective for understanding the reasoning mechanisms of large language models, which has both theoretical significance and application value.

The reasoning ability of large language models is a core research topic in the field of artificial intelligence. Although current models have made significant progress in complex reasoning tasks such as multi-hop question answering, the internal mechanisms of how they organize and utilize evidence for reasoning remain unclear. Understanding these mechanisms helps improve model architectures, identify and correct potential flaws.

This study proposes the concept of "evidence sufficiency separation": when a model processes a fixed question but faces different contexts, its hidden states exhibit two patterns—the "sufficient state" where existing evidence is enough to answer the question, and the "insufficient state" where evidence is lacking or further reasoning is needed, revealing that there is an evidence evaluation mechanism inside the model.

Task Setting

The study adopts a multi-hop question answering paradigm with fixed questions and varying contexts: the same question is paired with different background paragraphs (containing complete reasoning chains, partial information, or irrelevant information) to precisely control evidence sufficiency.

Model Selection

Representative reasoning models (Transformer-based dedicated reasoning models and general large language models) are selected, all of which perform well on standard multi-hop question answering benchmarks.

Analysis Methods

Linear probing (identifying hidden state dimensions related to evidence sufficiency), causal intervention (verifying the reasoning participation of dimensions), and attention visualization (tracking changes in attention distribution) are used.

Binary Clustering of Hidden States

The hidden states of models show obvious binary clustering in the dimension of evidence sufficiency: they cluster in a specific area when evidence is sufficient and in another area when evidence is insufficient, with the middle layers showing the most obvious phenomenon.

Functional Significance of Separation Dimensions

Causal analysis confirms that separation dimensions are involved in reasoning decisions: when these dimensions are intervened, the model's answer accuracy changes significantly.

Cross-Model Consistency

This binary separation phenomenon exists in different model architectures (specific dimensions may vary), suggesting that it is a universal mechanism of reasoning models.

Understanding Transformer Reasoning

The traditional view holds that Transformers transmit information through attention. This study shows that models also maintain a global evidence sufficiency state, which may be transmitted between layers through residual connections.

Connection to Cognitive Science

The binary separation phenomenon is similar to humans' "feeling of knowing" (judging whether sufficient information is mastered before answering), corresponding to the model's metacognitive process.

Uncertainty Quantification

Monitoring hidden state regions can identify cases where the model "does not know", avoiding overconfident wrong answers.

Reasoning Chain Verification

Tracking changes in hidden states can identify evidence accumulation steps or missing points, guiding the improvement of reasoning quality.

Model Distillation and Compression

Focusing on key evidence evaluation dimensions can reduce model size while maintaining reasoning ability.

Limitations of this study: Experiments are based on artificially constructed multi-hop question answering datasets, and the separation phenomenon needs to be verified in real complex scenarios; current focus is on binary separation, but actual evidence sufficiency may be a continuous spectrum. Future work will explore fine-grained evidence state modeling and application to large-scale practical systems.