Activation steering controls language models by adding vectors or features to hidden states at inference time, but the upstream source of these steering signals is often treated as a secondary detail. We study this source choice as activation source selection: the combination of source context and activation readout policy used to collect the hidden states from which a steering signal is built. Holding the downstream intervention fixed, we show across three instruction-tuned models and four steering task families that changing only the source activations can more than double steering success. We further find that effective steering is not explained simply by whether the desired behavior appears in the source text. Instead, strong signals come from execution-boundary states, where the model is about to produce or continue the target behavior. This phase distinction explains why answer-based sources sometimes work: their useful component aligns with execution-boundary directions rather than target appearance alone. Building on this view, we introduce tail subtraction, which removes shared prompt and continuation semantics from boundary states and yields cleaner, more stable steering signals. Overall, our results suggest that steering depends on representations of what the model is about to do, not merely on what has already appeared.

Standard reasoning benchmarks conflate two distinct failure modes into a single accuracy metric: absent internal answer representations versus flawed output sequence scoring. In this work, we show these failure modes can be disentangled. Across controlled logic, ProofWriter, ANLI, and FOLIO, we compare internal hidden-state probes against output candidate-answer string scores. We find that in post-trained language models, correct answers are often strongly encoded in internal representations even when raw sequence scoring collapses to chance or incorrect labels. To expose this hidden evidence, we apply a lightweight diagnostic correction: adjusting per-label output log-probabilities via two additive offsets fitted from as few as 25 unlabeled examples to match a prespecified label prior. This target-label-free correction recovers between 9 and 34 accuracy points (with matching macro-F1 gains) across larger Qwen3.5 checkpoints, and successfully recovers performance on OLMo-2 and Llama-3.1. The gains survive on examples a TF-IDF control misses and significantly exceed label-count permutation nulls. Our findings demonstrate that many LLM logical reasoning failures are output scoring artifacts rather than internal reasoning deficits, showing that benchmark accuracy can systematically underestimate a model’s representational capacity.

Understanding how Large Language Models (LLMs) perform logical reasoning internally remains a fundamental challenge. While prior mechanistic studies focus on identifying task specific circuits, they leave open the question of what computational strategies LLMs employ for propositional reasoning. We address this gap with a causal mechanistic analysis on PropLogic-MI, a controlled benchmark of 11 propositional rules across one- and two-hop tasks, applied to three model families (Qwen3, Llama-3.1, Mistral). Rather than asking which components are necessary, we ask how the reasoning process is organized, and identify four interlocking mechanisms: Staged Computation, where early, middle, and late layers take on distinct functional roles; Information Transmission, where semantic content aggregates at boundary tokens; Fact Retrospection, where fact tokens stay causally active as binding lookup in middle layers and as sustained access in late layers under hard reasoning loads; and Specialized Attention Heads that structurally implement these patterns. A prompt-order control that places the query before the facts confirms that this organization is model-internal rather than input-layout-induced. These findings show that pretrained LLMs solve propositional reasoning through a structured, layer-organized process that recurs across models, rule categories, and reasoning hops.