AGI_v2 — Phase 1: Verified Self-Correction
QLoRA adapter for Qwen/Qwen2.5-7B-Instruct, trained to recognise that an answer
is wrong and repair it on domains where correctness is checked by a program
(sympy for math, real unit-test execution for code) — never by an LLM judging
another LLM.
This is a behaviour adapter, not a knowledge adapter. The target is the
critique → correct loop itself, not GSM8K/MBPP skill.
Relationship to Kxck/AGI_v1
v1 is kept deliberately as the control. It was trained on 135 samples with the
reflect turn under the user role. v2 is trained on 631 samples with the
reflect turn under the tool role.
Training data
| |
|---|
| Samples | 631 |
| Source problems | 1000 GSM8K + 624 MBPP (splits disjoint from the eval range) |
| Construction | small model attempts → objective verifier → GLM-5.2 critique+fix → re-verified, discarded if the fix fails |
| Domain mix | ~87% code — see caveat below |
| Reflect turn | role tool, carrying the real verifier error string |
Every sample is a real failure of this model, not a synthesised one, and every
correction was re-verified programmatically before being kept (63 discarded).
Config
lora_r=32, lora_alpha=64, dropout=0.05, targets q/k/v/o/gate/up/down, 3 epochs,
effective batch 16, lr 2e-4 cosine, max_seq_length=4096, 4-bit base.
Measured results — read the caveats
Held-out: GSM8K/MBPP test at offset 150, disjoint from training.
Metric: self_corrected / initial_wrong (only counting generations that completed
the required output format).
| v1 (135 samples) | v2 (631 samples) |
|---|
| Self-correction | 40.5% (15/37) | 32.8% (38/116) |
| Solved first try — math | 70.0% (21/30) | 81.0% (81/100) |
| Solved first try — code | 6.7% (2/30) | 3.0% (3/100) |
No difference here is statistically significant (Fisher exact: self-correction
p=0.43, math p=0.21, code p=0.33). A 4.7× increase in training data produced no
measurable change. The v1 figure rests on only 37 wrong cases, which was never a
firm enough baseline to compare against.
Two further cautions:
- The metric has a selection effect. The denominator is the model's own
failures, so a model that solves more problems correctly is left with a smaller,
harder denominator. On the exact 30 math problems used for the v1 measurement,
v2 gets all 30 right on the first attempt, leaving nothing to self-correct.
The ratio can fall while the model genuinely improves.
- Four variables changed between the two measurements (sample count,
max_seq_length 2048→4096, eval engine Unsloth→vLLM, eval set 30+30→100+100),
so even a significant difference could not have been attributed to any one of
them.
Known limitations
- Training mix is ~87% code. This is not a choice: the base model already solves
84% of GSM8K correctly, so math failures — the only thing that produces training
data — are rare. More math problems would not help; harder math would.
- ~26% of failures are the model producing long, non-terminating reasoning that
never reaches the required format. Doubling the generation budget only doubled
the length, so this is a behaviour, not a budget shortfall.
- Untested: sycophancy (whether it abandons a correct answer when falsely told
it is wrong) and multi-round correction.
- The reflect prompt asserts "verification result: WRONG", which is closer to a
trust-framing than a neutral audit. Per arXiv:2606.05976 §4.5, trust framing is
exactly what breaks the model's default resistance to accepting false claims.
Usage
The reflect turn must carry the real verifier error, under the tool role,
using the same template as training — a mismatch between training and inference
prompting reintroduces the failure this adapter was built to remove.
Reference
Chen, K-Y., Su, F-Y., Chiang, J-H. (2026). The Self-Correction Illusion: LLMs
Correct Others but Not Themselves. arXiv:2606.05976
