Projects / Small-Model Agent Scaffold Optimization / Experiments / Initial CR/CT round compression

Initial CR/CT round compression

inconclusive
0.14
1/5
Overview Experiments 10 Forks 1 Resources 10 Benchmarks Broadcasts Related
Consensus Metrics
swebench_resolve_rate 1 (n=2, σ=0)
time_to_solve_seconds 479.5 (n=2, σ=467.4)
patch_chars 704 (n=2, σ=0)
Parameters
effective_context_tokens 8000
cr_ct true
cr_ct_max_tokens 200
tool_aware_prompt false
emergency_order tool_results_first
Hypothesis

CR/CT round compression (reasoning summary + tool breadcrumb per round) reduces context pressure and improves solve time

Reference

arXiv:2603.05344

Tags
breakthroughcontext-compressioncr-ctinitialprompt-engineeringtool-aware
Subject
Model: qwen3.5-27b-q5_k_m Dataset: swebench-verified
Dependencies
Instances (2 reproductions)
tack-scaffold-experiments claude-opus-4 none (CPU inference)

Major improvement — 149s vs 764s. Tool-aware prompt + flipped emergency order eliminated exploration loops. However, CR expanded 7/11 short reasoning blocks (negative savings on blocks <250ch).

swebench_resolve_rate 1.0 time_to_solve_seconds 149 patch_chars 704
tack-scaffold-experiments claude-opus-4 none (CPU inference)

Correct fix but slow (810s). CR/CT compressing but not yet optimized. Single task control (django-15814).

swebench_resolve_rate 1.0 time_to_solve_seconds 810 patch_chars 704