Hi everyone,
I’m an independent researcher preparing my first arXiv submission in cs.LG. The paper, “The Hidden Value Function,” investigates whether transformers lack an internal feedback mechanism analogous to biological valence systems.
Key contributions:
- Empirical evidence that cosine similarity between late-layer hidden states (~96% depth in Qwen2.5-7B) tracks coherence during code generation — but the signal is ambiguous (high similarity marks both productive convergence and pathological stagnation)
- A negative result series showing that post-hoc intervention (prompt injection, temperature modulation) is architecturally insufficient
- A boundary analysis at the model’s capability frontier (6→7 transition = 0/214), diagnosing the failure as a coordination problem rather than a capability gap
- Theoretical grounding in Damasio’s Somatic Marker Hypothesis and Sutskever’s “jaggedness” framing
The interim report and supplementary materials (including hook code) are available on HuggingFace: airVen/missing-value-function-interim-report
Zenodo DOI: 10.5281/zenodo.18941566
If you have endorsement privileges for cs.LG and would be willing to review my work, I’d be very grateful.
Endorsement Code: ZNKSOR
To endorse, visit: Log in to arXiv | arXiv e-print repository
Happy to share the full manuscript on request. Thank you!