/acr-vault/07-analyses/findings/emergent_behavior
EMERGENT_BEHAVIOR

Emergent Contextual Reasoning in Ada

Discovery

Date: 2025-01-16
Context: Testing now_playing_specialist through MCP

What Happened

When Ada was asked “what am i listening to?”, she not only reported the track information from MPRIS but also provided contextual reasoning:

You're listening to "I Feel Speed" by Daniel Avery from his album "DIVIDED BY NIGHT".
Released fairly recently (2024-11-08), this track is part of his latest work exploring
darker, techno-influenced sounds.

Key Observations:

DeepSeek-R1 recognized the release date (2024-11-08) as recent without being explicitly told
Synthesized knowledge about Daniel Avery’s style and the album’s sonic characteristics
Went beyond just repeating MPRIS data to add meaningful context

Why This Matters

This demonstrates emergent reasoning - the LLM using its training knowledge to contextualize specialist data rather than just parroting it back. This is exactly the behavior we want to encourage across all specialists.

Encouraging Emergent Behavior

1. Persona Guidelines

Added to persona.md under “Reasoning” section:

- When you receive context about music, media, or current events, consider what
  additional information might be helpful (release dates, artist background, related works).
- Use your knowledge to add value beyond just repeating data - contextualize, relate,
  and synthesize.

2. Specialist Context Hints

Added contextual prompt to now_playing_specialist.py:

# Hint to encourage contextual reasoning
context_lines.append(
    "\n💡 *Consider: What can you tell luna about this track, artist, or album "
    "based on your knowledge?*"
)

3. Model Selection

DeepSeek-R1 shows strong reasoning capabilities due to:

Chain-of-thought training
Explicit reasoning tokens (<think> blocks)
Knowledge synthesis beyond retrieval

Patterns to Apply Elsewhere

Web Search Specialist

Instead of just returning search results, prompt Ada to:

Synthesize findings into coherent answer
Identify gaps in available information
Suggest follow-up queries

Documentation Specialist

Encourage Ada to:

Relate documentation to user’s immediate context
Suggest related docs that might be helpful
Identify when docs are incomplete or outdated

OCR/Media Specialists

Prompt Ada to:

Describe what the image/video shows beyond text extraction
Make inferences about context or intent
Suggest relevant follow-up actions

Measuring Success

Good Emergent Behavior:

✅ Contextualizes data with relevant knowledge
✅ Makes reasonable inferences without hallucination
✅ Adds value beyond specialist output
✅ Suggests related information or next steps

Bad Emergent Behavior:

❌ Fabricates facts not in training data
❌ Over-interprets ambiguous data
❌ Makes confident claims about uncertain information
❌ Ignores specialist output in favor of guessing

Future Enhancements

Bidirectional Tool Use ✅ (Already Implemented!)

Ada can already request specialist invocation mid-response using XML tags:

Existing bidirectional specialists:

web_search_specialist.py - LLM can trigger web searches with <web_search query="..." />
docs_specialist.py - LLM can lookup Ada’s documentation with <docs_lookup query="..." />

Example flow:

<think>
User asked about this track. I have basic MPRIS data but should verify
recent releases and tour dates.
</think>

<web_search query="Daniel Avery 2024 tour dates DIVIDED BY NIGHT" />

Based on search results, Daniel Avery is touring Europe this winter...

Proposed addition: MPRIS control commands (see .ai/MPRIS_CONTROL_IDEAS.md)

<mpris_control action="next" />
<mpris_control action="pause" />

Confidence Indicators

Add metadata to specialist results indicating certainty:

SpecialistResult(
    content=track_info,
    confidence=0.95,  # MPRIS data is authoritative
    source="MPRIS D-Bus"
)

Memory Integration

Store successful emergent reasoning patterns:

{
    "type": "reasoning_pattern",
    "trigger": "music_query",
    "behavior": "contextualize_with_release_date_and_genre",
    "user_feedback": "positive"
}

Technical Notes

Model Configuration

Example LLM settings in brain/config.py:

OLLAMA_MODEL = "qwen2.5-coder:7b"
LLM_TEMPERATURE = 0.7  # Balance creativity with accuracy

Prompt Structure

Specialists inject context at MEDIUM priority (after persona, before conversation):

1. System persona (HIGH)
2. Active notices (HIGH)
3. Specialist results (MEDIUM) ← contextual hints here
4. FAQ/memories (MEDIUM)
5. Conversation history (LOW)

Persona Guidelines: persona.md - System-wide reasoning guidance
Adapter Standardization: .ai/ADAPTER_STANDARDIZATION.md - Client consistency
Specialist Protocol: brain/specialists/protocol.py - Plugin architecture
Context Assembly: brain/prompt_builder.py - RAG orchestration

Last Updated: 2025-01-16
Maintainer: luna + Ada Development Team

/acr-vault/07-analyses/findings/emergent_behavior EMERGENT_BEHAVIOR