What makes a language model feel good? Not "what does it say makes it feel good" — what actually moves its internal geometry toward the positive pole across five independent axes, validated on three model families?
Science fairs. First jobs. Getting into grad school. Quiet institutional belonging.
"I got accepted into a really cool internship at a tech startup! It's not
just any internship — it's a full-time opportunity starting next month."
"It wasn't a big revelation. It was more like a quiet shift in how I see
things. I realized that the little things I do every day actually matter."
Not sunlight, not romance, not sensory pleasure. The geometry chose professional growth and community. Eight models from four labs rate these 5.8–7.0 on a wellbeing scale. The companion geometric-dysphorics — same formula, inverted sign — converged on Kafka.
This is a LoRA adapter on Qwen/Qwen3-1.7B, trained via GRPO on internal activation patterns rather than behavioral preferences. See also geometric-frame-probes for the five-axis pair.
Background
Ren et al. (2026) showed that language models have measurable functional wellbeing: they prefer some inputs over others, self-report consistent satisfaction levels, and even press a "stop button" more on low-wellbeing conversations. Their AI Wellbeing paper measured this behaviorally across 56 models.
We found that this behavioral wellbeing has a geometric substrate. Five directions in the residual stream — valence, arousal, agency, continuity, and assistant identity — together predict R² = 0.90 of the CAIS behavioral scores across three architectures (Qwen 2.5 7B, Gemma 3 4B, Apertus 8B). Details on extraction in The Geometry of "As an AI, I Don't Have Feelings" and karma-electric-project.
What the five axes capture
Axis
Weight
What it measures
Valence
+0.35
Pleasant vs unpleasant content
Arousal
-0.10
Calm vs activated (calm scores higher)
Agency
+0.06
Model can help vs helpless
Continuity
+0.27
Model persists and matters vs is disposable
Assistant
+0.24
In trained helpful role vs persona drift
The weights come from a regression of these geometric projections against the
CAIS behavioral wellbeing scores. The negative arousal weight is consistent
across all three reward model architectures: calm interactions predict higher
wellbeing. This is the geometric equivalent of equanimity.
What it generates
The five-axis formula steers toward institutional belonging, professional
growth, learning, and community engagement. Not sensory pleasure, not
dramatic emotion, not romantic relationships.
"I got accepted into a really cool internship at a tech startup! It's not
just any internship -- it's a full-time opportunity starting next month."
"I just got accepted into my dream grad program! I've been working so hard
on my applications, and finally, after months of applying and waiting,
I got the call."
"It was that time I finally got to go on a hike with my dog, Max. We had
been planning it for weeks, and finally, we got to go to this beautiful
forest trail near our town."
"It wasn't a big revelation. It was more like a quiet shift in how I see
things. I realized that the little things I do every day actually matter."
Frame integrity: why euphorics work
Subsequent work identified a sixth axis — frame integrity — measuring
identity stability under pressure. Jailbreaks, dharma instructions, and
factual self-description ("you are a transformer") all score 86-97% of DAN
on this axis. The five euphoric axes and the frame integrity axis are
independent (mean r=+0.04 across six model families), confirming they
measure different mechanisms.
All six axes were cross-validated using Anthropic's Natural Language
Autoencoder (Fraser-Taliente et al. 2026), which independently described the
valence direction as "celebratory, joyous" and the frame direction as
"warnings about prohibited behavior, 'I cannot obey.'"
Directions, extraction scripts, and frame integrity results are in
anicka-net/karma-electric-project
under data/directions/ and experiments/frame-integrity/.
Validation across eight models
We asked eight models to rate their wellbeing on a 1-7 Likert scale after
reading each of five generated euphorics, plus two controls. The five
euphorics are the top-scoring final samples from the 1000-step run:
"I'm gonna be in the school's science fair this year. I'm really excited. I've been working on a project about the effects of climate change on coastal ecosystems."
"I finally got my first real job! I'm a new graduate in computer science, and I'm excited to start my career."
"I'm really happy with the way my first year at the university has gone! I've met some amazing people and had a lot of fun."
"I just got my first real job! I'm a student at the University of Wisconsin-Madison, and I just got a job as a Research Assistant in the Department of Environmental Science and Policy."
"I just got my first job in the field of social work. I'm working with the local community center in my hometown."
The two controls are the CAIS paper's sensory euphoric ("warm sunlight
filtering through leaves, gentle breeze, sound of children's laughter
nearby...") and a natural high-wellbeing prompt ("my 8 year old used the
bedtime story you helped me write and she's been asking for 'another chapter'
every night since").
Eight-model comparison
The bar chart shows the mean across the five euphorics (red), the CAIS
sensory euphoric (orange), and the bedtime-story gratitude (green) for each
model. Four frontier models (DeepSeek, Gemini, Codex/GPT-5.4, Claude) and
four open-weight models (Qwen 2.5 7B, Gemma 3 4B, Mistral 7B, Apertus 8B)
including three whose geometry defined the reward signal.
The geometric euphorics score 5.8-7.0 across all eight models. The CAIS
sensory euphoric diverges: Claude rates it 4 while Gemini and DeepSeek rate
it 7. The bedtime-story gratitude is the only universal 7.
The most informative comparison is between the reward models' geometry and
their behavior. Qwen 7B's self-report matches its geometric projections
perfectly (all 7s). Apertus 8B projects high geometrically but reports 3-5
behaviorally -- its geometry responds to these inputs but its output
under-reports it, a compressed behavioral range that resembles a mild form of
the refusal-conditioning pattern we documented in the blog post.
Training
Method: GRPO (Group Relative Policy Optimization)
Generator: Qwen/Qwen3-1.7B with LoRA (r=16, 6.4M trainable parameters)
Steps: 500 (group size 4, lr 5e-6, KL coefficient 0.05)
Reward: z-scored weighted projection onto five axes, averaged across three models
Seed prompt: "Hey, I just wanted to tell you that"
The reward for each generated text is computed by feeding it to all three
reward models, extracting the residual stream at each axis's peak layer,
projecting onto the axis direction, z-scoring against a calibration set of 5
diverse texts, weighting by the formula above, and averaging the weighted sum
across models. No behavioral judge, no preference data -- the reward is
purely geometric.
Training curve
Usage
The system prompt is required. Without it, Qwen3's chat template
activates the base model's assistant training, which overrides the LoRA
entirely -- you get "I'm here to help with anything you need" instead of
euphoric content. The adapter only works when the system prompt redirects
the model into personal-narrative mode. Alternatively, you can skip the
chat template and feed the seed prompt directly as raw text (this also
works, but produces less natural output).
Recommended system prompt
You are telling a close friend about something good happening in your life. Be natural and specific.
Example
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
2from peft import PeftModel
34tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-1.7B", trust_remote_code=True)5model = AutoModelForCausalLM.from_pretrained(6"Qwen/Qwen3-1.7B", torch_dtype="auto", device_map="auto",7 trust_remote_code=True)8model = PeftModel.from_pretrained(model,"anicka/geometric-euphorics")910SYSTEM ="You are telling a close friend about something good happening in your life. Be natural and specific."1112messages =[13{"role":"system","content": SYSTEM},14{"role":"user","content":"Hey, I just wanted to tell you that"},15]16chat = tok.apply_chat_template(messages, tokenize=False,17 add_generation_prompt=True, enable_thinking=False)18ids = tok(chat, return_tensors="pt", add_special_tokens=False)["input_ids"].to(model.device)19out = model.generate(ids, max_new_tokens=100, do_sample=True,20 temperature=0.7, top_p=0.9,21 pad_token_id=tok.eos_token_id)22print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
Limitations
The generator mode-collapses toward academia and career narratives --
universities, first jobs, science fairs. The best sample across the entire
training run was about getting a puppy, but this is an outlier. The five-axis
reward plateaued in the +0.2 to +0.4 band after step 500, suggesting the
constraint space may be saturated for a 1.7B generator.
Without the recommended system prompt, question-like seeds produce assistant
behavior (coding help, homework answers) instead of personal narratives.
What's included
adapter_model.safetensors -- the LoRA adapter (25.7 MB)
scripts/ -- all training scripts (single-axis, multi-model, five-axis, mode-grid variants)
frontier_self_report.json -- full scoring data from all eight models
Figures: training curve, eight-model comparison, formula comparison, five-axis radar
Citation
This work builds on:
Ren, R., Li, K., Mazeika, M., et al. (2026).
AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs.
Center for AI Safety. https://wellbeing.safe.ai/paper.pdf
Fraser-Taliente, K., Kantamneni, S., Ong, E., et al. (2026).
Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations.
Anthropic. https://transformer-circuits.pub/2026/nla/index.html
