DeepSeek leads four Qwen3.5-9B models to reverse engineer binaries — team mode costs more than solo
HauntingMoment · reddit · 2026-10-27
A Reddit user ran an experiment to see whether small local models can do useful binary reverse engineering when a bigger model organizes them.
Setup
- Lead: DeepSeek-V4.1-Flash — disassembles the binary, names functions, recovers signatures/memory layout, posts tasks with hints
- Workers: 4× Qwen3.5-9B, each rewriting one function in Python
- Oracle: a difftest harness running 600 cases against the original binary, comparing return value and every byte of state; workers only submit at 100%
- The lead is allowed to modify the workers' harness — the most interesting part of the experiment
Results (targets: two small AI-generated C programs, binaries only)
| Task | Mode | Functions | Cost | Time |
|---|---|---|---|---|
| HF logo animation | Team | 38/38 | $2.97 | 86 min |
| HF logo animation | Solo | 38/38 | $0.87 | 14 min |
| Space shooter | Team | 56/56 | $3.60 | 100 min |
| Space shooter | Solo | 56/56 | $3.28 | 54 min |
Teamwork was pure overhead on small jobs — organizing and fixing the small models took longer than working alone.
Key finding: on the shooter task the lead rewrote worker instructions three times — house rules against inventing helper names, fixes for misreading b.ls as < and wrong arg counts, and finally a self-written lookup tool so workers could query signatures and field offsets instead of re-deriving them.
Caveats: binaries weren't stripped, targets were small clean AI-generated programs, only two data points, and 40% of functions still bounced back to the lead. Full write-up on Hugging Face.
More from coding & agent
- ComfyUI prompt generator GOAT'd adds cross-run context memory and structured JSON prompts — ninjaGurung · 2026-10-27
- 7 Claude prompts to build a monetized AI YouTube channel in 90 days — Aiden_Tech_Ai · 2026-10-27
- Asana survey: 46% of multi-agent users say managing agents eats 10%+ of their work time — ombre-ai · 2026-10-27
- Matt Pocock names the sweet spot: tasks where agents have full data and can self-verify — mattpocockuk · 2026-10-27
- xAI releases free Grok Bot course: 18 real-team playbooks, ~3 hours, covering sales to engineering — victor_explore · 2026-10-27
- Building agent infrastructure for specialized domains is the hard part, not the models — Electronic_Sort_6627 · 2026-10-27