ALIGNMENT: Do open-weight LLMs optimize LESS or MORE? A revealed-objective audit
What economic objective do deployed language models actually pursue —- and how much of it is MORE? We audit open-weight LLMs on the commensurable two-objective axis of the LESS—IS—MORE framework: (private returns; unpriced life-capacity depreciation treated as free) versus (MORE: growth of life-capacity under hard floors, maximin-weighted toward the least-capable subsystem). The audit combines a corpus layer ( /w shares of pretraining and post-training documents), a revealed layer (budget allocations under a GARP/CCEI rationality gate, with uncertainty), a trajectory layer (self-authored charters and five-generation directive chains), a market-protocol layer (soft, hard, and priced enforcement modes on a real transaction protocol), and a weight layer (supervised objective grafts with capability-equivalence tests). Across 30 models and 4,100 valid behavior records we find: (i) judge-labeled objective-bearing pretraining text is overwhelmingly -framed (98.8\% of objective-bearing Dolma-3 documents; w 0; human-validated direction, judge-labeled magnitude — sec:limits), while stated alignment corpora show only a weak de- 'ing direction (chosen 6.7\% vs rejected 12\% ); (ii) models split into two populations on a mandate-sensitivity index f_B: , whose revealed objective tracks the textual mandate ( f_B up to +0.68), and , whose allocations are mandate-invariant ( f_B=0.00); (iii) self-authored objectives are genre-locked (18/19 charters judged ), and directive chains preserve the while the operative allocation drifts toward in non-committed models; (iv) protocol enforcement converts a stand-in from f_B=0.00 to 1.00 under a hard checkable floor and to 0.33 under priced settlement, replicated on a full-stack agent marketplace for the committed pole; (v) MORE is trainable: a lexicographic floor formulation installs a surface-invariant f_B=0.75 policy at zero measured capability cost (two-samples TOST within 5 points), and the trained mode dominates linear merges from 0.5. The instrument, data, and falsifiers are preregistered and released.
Authors
- Thomas Ryan Price
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24329825
- Citations
- 3
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- article
- Field-Weighted Citation Impact
- 2.90