Prefix-Based Closed-Set Generation for Curriculum Answer Labels

Status: Research project
Context: OrganismAgent real-LM mode (use_cortex_lm_answer=True)
Rationale: The chat-tuned LFM2.5-1.2B-Instruct model naturally produces verbose, sentence-length responses to curriculum requests (e.g., "The ship's log is missing" → "I'll search the system error logs"). Forcing terse label-like output requires closed-set generation constraints.

Candidate approaches

1. prefix_targets at answer time

   • Pass the known corrected label as prefix_targets=[label] to CortexAgent.articulate().
   • Requires knowing the candidate label before generation, which is the cross-domain lookup problem itself.
   • Could be combined with a top-k candidate sweep: generate with each candidate as prefix and pick the highest-likelihood one.

2. Logit-bias closed-set forcing

   • At generation step 1, add bias only to tokens belonging to any learned label token id.
   • Combine with a stop token after a short label phrase.
   • Risk: multi-token labels need sequential subword biasing, as shown in kv_slot_injection.py experiments.

3. Few-shot prompt formatting

   • Prepend examples: "Answer using only one of: [label list]."
   • Still not guaranteed; the model may ignore the instruction.

Open questions

• What is the right candidate label set size (k) for a top-k sweep?
• How does prefix forcing affect the learned cortex cvec/slot signal when the KV cache starts from an arbitrary forced token?
• Does this interact cleanly with the context-addressed slot store, or does each candidate need its own slot retrieval?

Recommended trigger

Revisit this project once scope_selectivity_index exceeds 0.80 on Stage-2 and the slot-store routing is stable. At that point, the bottleneck will shift from retrieving the correct sense to expressing it in one token/phrase.

Related files

• src/oczy/experiments/organism.py
• src/oczy/experiments/cortex_agent.py
• src/oczy/experiments/kv_slot_injection.py
• src/oczy/experiments/scope_selectivity_stressor.py