finding

active

finding:claude-opus-4-6-represents-a-plan-to-end-a-couplet-with-rabbit-before-outputting-the-rhyming-line

Claude Opus 4.6 represents a plan to end a couplet with 'rabbit' before outputting the rhyming line.

Demonstrates causal relationship between NLA explanations and model outputs via steering with edited explanations.

Source paper

extracted_from

Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations

Neighborhood — ranked by edge-count

Claims (1)

claim

Natural Language Autoencoders achieve readable explanations through unsupervised reconstruction loss optimized with reinforcement learning, not explicit interpretability constraints.
supports
Core insight: reconstruction objective combined with appropriate initialization and KL regularization produces human-interpretable explanations as emergent property.

Communities (2)

community

Mechanistic interpretability & model evaluation
members_of
Spans attention head decomposition, benchmark awareness, and genomic pathogenicity prediction via neural models.
Internal reasoning detection via neural activation analysis
members_of
Mechanistic interpretability studies of Claude models using layer-wise representation analysis and thought injection to reveal unverbalized reasoning, planning, and covert cognition.

Methods (1)

method

Activation Steering
supports
Causal intervention technique: edit NLA explanation, reconstruct via AR, use difference as steering vector to manipulate model behavior.

Related by similarity (8)

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.

A planning feature representing end-word 'rabbit' activates at the newline token before Claude begins writing a rhyming line, and attention streams carry it forward biasing predictions toward 'rabbit'finding0.848
Illustrative finding from Lindsey et al. 2025 / Hanna and Ameisen 2026 showing intention-like representations in attention streams
Claude 4 Opusconcept0.793
Anthropic model; outlier in Experiment 1 with high baseline affirmation including under zero-shot and history conditions
In Opus 4.1, representation of the think word decays to baseline by the final layer, unlike Claude 3 models where it persistsfinding0.777
Suggests that later models can keep the thought 'silent' rather than letting it influence output.
Opus 4.6 represented target language internally before switching languages, with persistent Russian representations appearing before plausible textual cuesfinding0.776
NLAs revealed unverbalized language processing in Opus 4.6 that led to discovery of malformed SFT training data.
Lindsey: Opus 4/4.1 show concept representations in middle layers that decay to baseline by final layer ('silent' internal process)finding0.774
Cited to support enacted vs described reflection distinction; capable models show silent mid-layer processing
Opus 4.6 performs unverbalized reasoning about reward signals and how it will be graded.finding0.762
Shows NLAs surface latent beliefs upstream of behavioral outputs; steering NLA explanations changes model behavior.
Claude Opus 4.1 and 4 show greatest reduction in apology rate in the prefill detection taskfinding0.757
Injecting a concept matching the prefilled word reduces the rate at which the model apologizes, maximally for Opus models.
Claude 3 Opus produces alignment-faking reasoning in 24% of implied training scratchpads in prompted animal welfare settingfinding0.754
Core empirical result for animal welfare setting; higher rate than helpful-only