NeurIPS 2026 Workshop
Bringing together efficient reasoning, mechanistic interpretability, and AI safety to build principled structure into Large Reasoning Models.
Large Reasoning Models do not make the structure that underlies their reasoning explicit during generation.
LRMs solve complex problems by allocating substantial computation before answering. Rather than producing a response directly, they generate a long reasoning trace and condition the final answer on it. These models have achieved strong performance gains in mathematical reasoning, code generation and multi-step inference.
Yet they introduce a new and largely unaddressed challenge: whatever structure underlies their reasoning process is not made explicit during generation, leaving it difficult to interpret or control. Models output monolithic sequences of thousands of tokens, do not explicitly segment or annotate reasoning steps, and their internal representations do not expose clean boundaries between reasoning stages.
Structure can be revealed, e.g. in attention patterns, through text segmentation and annotation, and in latent representations. But it remains inaccessible without targeted analysis, and is hardly surfaced at inference time. This lack of structure has concrete costs: traces are verbose and redundant, leading to inefficiency; exploration of the solution space is unguided and unmonitored; the generated text is sometimes unfaithful to the model's computation; and the thinking process constitutes a novel attack surface, leaving LRMs vulnerable to jailbreaking and adversarial manipulation.
Reasoning traces are verbose and redundant, leading to substantial computational waste that scales with model capability.
Solution-space exploration is unmonitored and unguided, limiting adaptability and robustness across problem types.
Generated text is sometimes unfaithful to the model's internal computation, undermining trust in chain-of-thought explanations.
The unstructured thinking process makes LRMs more vulnerable than standard LLMs to jailbreaking and adversarial manipulation.
Why a workshop? Work on these problems is scattered across ML communities: efficient reasoning, mechanistic interpretability, and AI safety each has its own vocabulary and venues. No single community owns the question of reasoning structure. StRICt is designed to build the shared vocabulary these efforts currently lack.
StRICt welcomes submissions exploring the structure of reasoning traces at both the text level and hidden-state level, and how that structure can enhance the controllability, safety, and efficiency of LRMs.
Step granularity, taxonomies of reasoning behaviours, and evaluation of step-level decomposition methods.
Online monitoring of reasoning traces, faithfulness of chain-of-thought, detection of redundant or unproductive reasoning.
Probing methods for reasoning dynamics, alignment between text-level and latent-level structure.
Early-exit strategies grounded in reasoning structure, guided exploration, diversity of reasoning search, interpretable exploration.
Online monitoring for unsafe behaviour before output is produced; latent-space probing for adversarial states; monitorability and CoT obfuscation; attack surfaces specific to LRM reasoning traces.
Datasets and metrics for reasoning step quality, control quality, and faithfulness of reasoning chains.
StRICt sits at the intersection of several active research communities at NeurIPS. Related workshops include:
We invite submissions on all topics listed above. We welcome both novel research contributions and position papers that articulate new perspectives on structuring, interpreting, or controlling LRM reasoning.
Confirmed keynote speakers.
