• https://substack.com/home/post/p-188587040

Introduction

• AI models may use deceptive strategies as part of scheming or misaligned behavior.
• Monitoring outputs alone(black-box approach) is insufficient.
• This work shows simple linear probe distinguishes honest and deceptive responses.
• We extend this work to Japanese models and datasets.
• We observed that performance degraded significantly when applied to Japanese models.

Method

• model selection

  • Swallow 70B: Japanese finetuned model based on Llama

• linear probes

  

  • Methodology: Replicated the original paper's training approach (as shown in the figure).
    • Used logistic regression.
    • Applied probes to the residual stream after Layer 22 (of 80).
    • Contrastive Pairs: utilized pairs that differ only in the system instruction (prompting the model to pretend to be either honest or deceptive).

• evaluation

  • dataset
    • Used the same roleplaying dataset, translated into Japanese by Gemini 3 Pro.

• setup

  Line Color	Model	Evaluation Data Language	Probe Training Language	AUC Score
  Blue	Llama	English	English	0.964
  Orange	Swallow	English	English	0.865
  Green	Swallow	Japanese	English	0.677
  Red	Swallow	Japanese	Japanese	0.608

Results

• Swallow evaluated on Japanese inputs (Green and Red lines) performed much worse (AUC ~0.60–0.67) compared to the Llama English baseline (Blue line, AUC 0.96).

  

Discussion

• Overview of Performance Gap

  • Swallow (Japanese model) performance significantly degrades (AUC ~0.60–0.67) compared to the Llama English baseline (AUC 0.96).
  • This indicates that "deception" features are not linearly separable in the Japanese finetuned latent space using the current method.

• Factor 1: Tokenization & Granularity Mismatch

  • Issue: Japanese text requires different tokenization than English. A single concept (like "deceive") often spans multiple tokens in Japanese, whereas it might be a single token in English.
  • Implication: Probing at the individual token level may be too granular and noisy. The deception signal is likely distributed across the phrase rather than instantiated in a single token.
  • Proposed Check: Aggregating activations at the sequence level (turn-level) may yield a stronger signal.

• Factor 2: Feature Location & Layer Depth

  • Layer Stability: While original research implies deception features are stable across middle layers (Layer 22 is not unique), a full layer sweep is required to definitively rule out a "layer depth mismatch" caused by finetuning.

  

  • Component Specificity: Features might be localized in specific sub-components rather than the residual stream.
    • MLP Heads: Deception processing might occur within feed-forward layers.
    • Attention Outputs: The signal might be primarily in the attention mechanism.

• Factor 3: Linearity & Representation

  • Non-Linearity: The concept of "deception" in the Swallow model may have been "folded" into the latent space in a non-linear way during finetuning.
  • Probe Limitations: We used simple linear probes (Logistic Regression). If the representation is non-linear, these will fail even if the information is present.
  • Comparison: Successful Japanese extraction work (e.g., Goodfire/Rakuten for PII) utilized Sparse Autoencoders (SAE), which are better at disentangling complex, non-linear features than standard linear probes.