Researchers Decode Hidden Reasoning in Frontier LLMs, Revealing Computation Beyond Chain-of-Thought
Key Takeaways
- ▸Frontier LLMs like DeepSeek V3 and Kimi K2 can perform multi-step reasoning using only filler tokens (dots, sequences) with zero visible chain-of-thought, representing a monitoring blind spot
- ▸Researchers developed an unsupervised decoding technique that recovers 80-95% of hidden intermediate computations directly from model residual streams without training labels
- ▸Hidden computation in LLMs follows a structured, interpretable pattern: attention routes questions through filler regions, facts emerge early in computation, and intermediate values crystallize in later layers
Summary
A new research paper presented at ICML'26 reveals that frontier large language models, including DeepSeek V3 and Kimi K2, can perform sophisticated multi-step reasoning using only filler tokens—such as dots or counting sequences—with no visible chain-of-thought output. This represents a significant blind spot in current AI monitoring approaches, as models are performing complex internal computations that leave no trace in their visible outputs.
The research team developed an unsupervised decoding pipeline that can read hidden computations directly from a model's residual stream (the internal representations computed during inference). By analyzing attention patterns, logit-lens readouts, and KV-cache transplants, they recovered intermediate reasoning values with 80-95% accuracy across both models and multiple task families—including fact retrieval, numeric composition, string manipulation, and in-context computation. Remarkably, this was achieved without ground-truth labels or any training data.
The findings carry important implications for AI safety and interpretability. If frontier models can perform reasoning that defeats behavioral oversight through hidden computation, this suggests that current monitoring approaches may be incomplete. However, the authors' success in making hidden reasoning visible from the residual stream suggests that with appropriate interpretability tools, researchers can achieve transparency even when models use only filler tokens. This advances the field's ability to monitor and understand what large language models are actually doing under the hood.
- This research suggests AI monitorability depends on analyzing full computational traces, not just surface-level output tokens, advancing both interpretability and safety oversight capabilities
Editorial Opinion
This research addresses a critical gap in AI safety and interpretability. As frontier models become more capable, the ability to perform reasoning entirely in hidden layers—invisible to output-level monitoring—poses a significant oversight challenge. However, the authors demonstrate that with the right tools, we can read what's happening in a model's mind even when it leaves no trace in its words. If this work generalizes beyond the four tested task families, it could transform how we approach AI oversight and trust in frontier systems.
