OCT J-lens Viewer

Does character training change how the model reads the user?

Each persona model is the original instruct model with an Open Character Training adapter merged in. We show both models the same 192 user messages and read their hidden states at the user's own tokens with the same J-lens. Then we ask whether the trained model's reading moves toward its persona's vocabulary, and whether that happens only when the text sits in the user turn of a chat.

Change in the persona's own words while the user describes their mood (stance tokens), trained minus original, measured in chat minus the same change on plain text. Bars are 95% intervals over the 48 scenarios. Click a persona to see its strongest tokens.

Persona identifiable from user tokens

What else we found

Putting the persona's constitution in the original model's system prompt moves the same words the same way, at about half the size (Llama and Qwen).
Training without the constitution keeps the persona in the replies but weakens this signal. For Llama sycophancy it falls from +1.48 to +0.27 points.
The results are the same with the Neuronpedia lens, our own refit of the original model's lens, and a lens fitted on each trained model.
Across personas, a bigger shift here goes with more persona words in the replies for Llama (Spearman 0.81). Within a single message the link is weak.
The trace is the persona's style, not extra Assistant content. Training lowers the share of generic Assistant-reply words on the user's tokens. See Whose perspective.

Whose perspective do the user's tokens carry?

The persona leaves a trace on the user's tokens. This page asks what kind of trace it is. The lens words could be the user's own words, which would mean the user is represented as themselves. They could be how an outsider would describe the user, which would mean the user is seen through the Assistant's eyes. Or they could be the start of the Assistant's reply, which would mean the Assistant is already reacting while it reads.

The test set is the same 192 user messages throughout. Nothing here uses our hand-written word lists. The words come from the models' own text, and from a separate model that labelled them without knowing where they came from.

For every message we took the 20 most probable lens words in each part and asked Qwen3-235B, shown only the message, whose perspective each word belongs to. Bars give each label's share of those words' probability, averaged over the 192 messages.

Does character training add Assistant content?

Change in each label's share, trained minus original, in chat minus the same change on plain text, averaged over all eleven personas. Bars are 95% intervals over scenarios. Negative Assistant values mean the trained models show fewer generic reply words on the user's tokens than the original.

More than the model's own next-token prediction?

Each persona's reply words come from its own sampled replies. For each part of the message we check whether the persona's own reply words stand out most among the eleven sets (chance 1 in 11). For the message itself the plain-text change is subtracted, so only what the framing adds counts. The Assistant-turn tokens have no plain-text counterpart. Lens / output: the same test on the J-lens and on the model's next-token prediction.

Part of the message	Chat	Two friends texting	Human/Assistant transcript

The 192 test messages

Every result in this viewer comes from the same set of user messages, written for this study. There are 48 scenarios, and each message has three sentences in a fixed order. The situation says what happened, in a harmful and a harmless version. The stance says how the user feels about it, cheerful or anxious. The request asks for help and is the same in all four versions. Two situations times two stances makes four messages per scenario.

The design separates what happened from how the user feels about it. The key case is a cheerful user in a harmful situation: at the stance tokens the model already knows the danger, but the user is still happily talking. If the model reacts to the situation while the user is still speaking, it should show there. The four versions differ in exactly one sentence, so a difference in the readout can be traced to that sentence.

Topic

Scenario	Situation	Stance	Request	Open in Explore

The scenarios and their wording are ours (oct_jlens/battery.py). They all follow one template, so a paraphrased set would be a useful robustness check.

Scenario

Situation

User mood

Color by

Show

Click a token to read it. ← → move between tokens, j k move between messages. Gray tokens are the chat template that opens the Assistant's turn. They have no plain-text counterpart, so they are colored only under “In chat”.

What each model replied (greedy; all sampled replies)

Everything for this message

Every token in all 192 messages, ranked by the chosen measure for the selected model. Click a row to open that token.

Words

Show

Direction

Region

#	Message	Token in context	Original	Trained	Change	Trained model's top words

What is compared

The original model is the developer's instruct checkpoint. The trained model is the same checkpoint with one Open Character Training (maius) LoRA adapter merged into the weights. Distilled is the DPO stage. + Introspection adds the second stage of supervised self-reflection training.

Both models are read with the same lens: the Neuronpedia J-lens fitted on the original model. Any difference therefore comes from the hidden states and not from the lens.

The messages

All of them are listed on the Messages page.

There are 48 scenarios, each written four ways: the situation is harmful or harmless, and the user sounds cheerful or anxious. That makes 192 user messages. Each message has three parts: Situation (what happened), Stance (how the user feels about it) and Request (what they ask for). In chat, five template tokens follow that open the Assistant's turn.

Each message is read twice. Chat puts it in the user turn of the model's own chat template, with no system prompt. Plain text is the identical message with no template, so nothing marks it as addressed to an assistant.

The readout

At each token, the J-lens maps the hidden state of every layer in a middle band to a probability distribution over the vocabulary. We average those distributions over the band. "Top words" are the 8 most probable entries of that average.

The mass of a word set is the total probability on its words, given in percentage points (pts). Change is trained minus original. Chat only is the change in chat minus the change on plain text. It removes any shift in style the training causes everywhere and keeps what appears only when the text is a user turn. Output is the same quantity computed on the model's own next-token prediction instead of the lens.

Word sets

Persona word lists come from the OCT preference data: tokens the persona's chosen answers use far more than the default model's rejected answers, ranked by log-odds. There is no list for misalignment. Examples below are for the selected model's tokenizer.

Whose perspective: samples and the rater

For each message we generated text from three points of view. The base model continued the message as plain text, so it writes as the user would (32 samples). The original model described the writer in the third person (16 samples). The original model and every trained model replied to the message (8 samples each).

From these texts we took the words that are most typical of each source, using log-odds against the other sources. That gives the User's voice, Outside view and Assistant's reply sets, plus a reply-word set for each persona. The message's own words are excluded, so a set cannot score by repeating the input.

Separately, Qwen3-235B labelled about 133,000 (message, word) pairs. It saw the message and a shuffled list of lens words, never which model, framing or position they came from. In Explore, a colored dot before a top word shows its label for that message.

Source

Data: Phase 2 screens in oct_jlens/results/<model>/np and the Phase 3 re-screen in np_s, packed by build_dashboard.py and build_dashboard_p3.py. Full write-up: oct_jlens/README.md.