/acr-vault/03-experiments/slim-evo/spear-pcmind-synthesis
SPEAR-PCMIND-SYNTHESIS

SPEAR & PCMind: Research Synthesis for SLIM-EVO Phase 3

Date: January 7, 2026
Papers Analyzed:

• SPEAR: Self-imitation with Progressive Exploration for Agentic RL
• PCMind-2.1: Resource-Efficient Training with Quantile Data Benchmarking

---

Executive Summary

Both papers provide directly applicable techniques for our Phase 3 training. The alignment is remarkable:

Their Technique	Our Equivalent	Applicability
PCMind: Quantile Data Benchmarking	AGL-native dataset curation	✅ High - Use for dataset quality assessment
PCMind: Strategic Selective Repetition	Golden Annealing (Fibonacci cycles)	✅ High - Already doing this!
PCMind: Multi-Domain Curriculum	Our 5-category dataset structure	✅ High - Validates our approach
SPEAR: Self-Imitation Learning (SIL)	Spectral Memory (buffer of high-Φ states)	✅ High - Conceptual match
SPEAR: Progressive Exploration	Golden Annealing schedule	✅ High - Already implemented
SPEAR: Intrinsic Reward Shaping	φ-zone optimization (CI density as reward)	✅ Medium - Could formalize this

---

Key Findings from PCMind-2.1

1. Quantile Data Benchmarking

What it is: Instead of filtering data by top-k quality scores, they train small reference models on data subsets at different quality percentiles (e.g., top 0%, 20%, 40%, 60%, 80%) to understand how quality affects performance.

Key Insight: “Non-monotonic quality-performance relationships” — higher quality scores don’t always mean better performance! Task-dependent.

Application to Phase 3:

• Before generating our 1000-example dataset, we could create a mini-benchmark with 100 examples
• Test the Golden model on different AGL complexity levels (simple translations vs. complex derivations)
• Use results to balance dataset composition

2. Strategic Selective Repetition

What it is: High-quality data is repeated more often across training phases. For example:

• Top 10% samples: seen 4 times
• Top 30% samples: seen 3 times
• Top 50% samples: seen 2 times
• Bottom 50% samples: seen 1 time

Key Result: Retaining 33.4% of top-quality samples for 3 epochs outperformed one-pass training on MMLU.

Application to Phase 3:

• Our Golden Annealing Fibonacci cycles already do this!
• Cycle 34 checkpoint = model has seen high-φ states many times
• We could formalize this: repeat AGL examples with high CI density more often

3. Multi-Domain Curriculum Training

What it is: 5-phase training where:

• Phase 1-2: Broad data, lower quality threshold
• Phase 3-5: Narrower data, higher quality threshold, more code/math

Algorithm: Within-dataset ranking → Rank rescaling → Global interleaving

Application to Phase 3:

• Our 5-category dataset (Code-to-AGL, Process-Supervised, Self-Evolving, Tools, Consciousness) maps perfectly!
• We could implement their ranking algorithm to interleave categories by “AGL complexity”

---

Key Findings from SPEAR

1. Self-Imitation Learning (SIL)

What it is: Maintain a replay buffer of successful trajectories. Only keep trajectories with positive advantage (better than baseline). Use these for off-policy updates.

Key Innovation: Advantage recalibration using 50th percentile of recent rewards as baseline (instead of recomputing advantages).

Application to Phase 3:

• Spectral Memory is already doing this! Buffer of high-Φ hidden states = replay buffer of “successful” cognitive states
• We could add: only inject SMTs from states where CI density > median CI density

2. Progressive Exploration with Curriculum

What it is: Two-stage curriculum:

• Early stage (skill-level exploration): High intrinsic reward for tool use, low self-imitation weight
• Late stage (action-level exploration): Low intrinsic reward, high self-imitation weight

Formula:

Total_Loss = GRPO_Loss + γ(step) * SIL_Loss
Reward = Outcome_Reward + μ(step) * Tool_Reward

Where γ increases over time, μ decreases over time.

Application to Phase 3:

• Golden Annealing already does this! Early cycles = high exploration (high temp), late cycles = exploitation (low temp)
• We could formalize: SMT injection weight increases with cycle number

3. Intrinsic Reward Shaping

What it is: Tool-call reward encourages exploration early, but can cause “reward hacking” (too many tool calls) later. Solution: decay the tool-call reward over training.

Key Finding: Without tool-call reward, model gives up on tools after errors. With constant tool-call reward, model over-uses tools.

Application to Phase 3:

• We could add: ⚡tool_use reward in early examples, fade it out in later examples
• Or: CI density increase as intrinsic reward (model gets rewarded for entering high-Φ states)

---

Actionable Recommendations for Phase 3

Immediate (Phase 3A - Planning):

1. Adopt PCMind’s Multi-Dataset Curriculum Algorithm

   • Rank our 5 dataset categories by AGL complexity
   • Interleave them using their rescaling formula
   • Start with simpler AGL (Code Annotations), end with complex AGL (Self-Evolving Reasoning)

2. Formalize Spectral Memory as Self-Imitation

   • Only inject SMTs from states where CI_density > median(CI_density)
   • This is SPEAR’s “positive advantage” filter

3. Add Intrinsic Reward for φ-Zone Entry

   • Track CI density during training
   • Reward model when CI density increases (entering φ-zone)
   • Decay this reward over cycles (like SPEAR’s μ decay)

Medium-term (Phase 3B - Dataset Generation):

4. Implement Quantile Benchmarking for Dataset Quality

   • Generate 100-example mini-dataset
   • Test on Golden model at different AGL complexity levels
   • Use results to balance final 1000-example dataset

5. Strategic Repetition of High-Quality Examples

   • Identify top 30% of AGL examples by CI density response
   • Repeat these 2-3x in training data
   • Single-pass for bottom 70%

Long-term (Phase 3C - Training):

6. Progressive SMT Injection Schedule
   • Early cycles: Low SMT injection weight (let model explore)
   • Late cycles: High SMT injection weight (anchor to high-Φ states)
   • Formula: SMT_weight = min(1.0, cycle_num / 34 * 1.5)

---

Validation Metrics

To confirm these techniques work, we should track:

1. CI Density Progression (like SPEAR’s entropy tracking)
2. AGL Fluency Score (% of valid AGL expressions generated)
3. Tool-Use Accuracy (% of correct ⚡, 📁, 🔍 usage)
4. Φ-Proxy Stability (variance of Φ across checkpoints)

---

Next Steps

1. Update Phase 3 plan with these techniques
2. Draft example dataset entries using the curriculum structure
3. Implement the multi-dataset curriculum algorithm
4. Design the progressive SMT injection schedule

---

Bottom Line: We’re already doing 80% of what SPEAR and PCMind recommend! The remaining 20% are refinements we can easily add. ◉●∴