Out-of-Context Reasoning

Model outputs influenced by information from training documents not present in context; relevant to synthetic document fine-tuning results

Neighborhood — ranked by edge-count

thinker

Lukas Berglund
studies
Prior work on out-of-context reasoning and situational awareness in LLMs
Zeyuan Allen-Zhu
studies
Prior work on knowledge manipulation and out-of-context reasoning in LLMs

method

Synthetic Document Fine-Tuning
implements
Fine-tuning Claude 3 Opus on ~70M tokens of synthetic internet-like documents containing key situational information

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.

Spatial Reasoningconcept0.772
The ability to reason about shapes, space, and topology, essential for ancient mathematical discoveries and observed in many animals.
strategic reasoningconcept0.768
High-level cognitive ability to plan and act under uncertainty and adversarial conditions.
In-Context Learning as Optimizationconcept0.768
Chain-of-Thought Reasoningconcept0.768
Medium through which eval awareness is often verbalized; target of intervention.
Context Featureconcept0.764
Feature that activates across all tokens within a specific context (e.g., DNA sequences, base64 strings)
latent reasoningconcept0.763
Reasoning approach using learnable hidden embeddings.
In-Context Learning of Representations (Park et al. 2025)framework0.760
Reports phase-like breakpoints and geometry changes as context scales; UCCT provides measurable predictor
Compositional Reasoningconcept0.759
Modular reasoning enabled by monoid and other algebraic laws; central to maintainability and correctness of functional systems.