Training helps models resist pressure to approve ecological harm
Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model families · arXiv:2610.37484
In plain words
The paper introduces the coupled-welfare corpus, a collection of real-data-based training documents for language models. It teaches how people and living systems depend on each other, why irreversible damage can close off future choices, and how biological systems can matter to AI capabilities.
After training, Qwen3.6-27B made fewer harmful choices in two pressure tests than its base version, while scoring the same on a small general-knowledge probe. The tests cover specific scenarios; they do not show broad alignment improvement.
The research
A model can be asked to prioritize short-term targets even when its decision could irreversibly harm living systems. The question is whether ecological reasoning can guide its choices when an operator mandate pushes in the other direction, rather than appearing only in what the model says it values.
The authors train models on documented links among people, living systems, and AI capability, then test whether their decisions hold under pressure.
Can ecological decision behavior hold when a model faces pressure to meet short-term goals? The authors introduce the coupled-welfare corpus, which teaches documented links among human welfare, living systems, and AI-relevant capability through continued pretraining.
In the Qwen3.6-27B comparison, harmful choices fell on both a pressure ladder and an open-allocation grid, while accuracy stayed at 84% on the same 50-item probe. The results support a change in these tested decisions, not a general claim about alignment.
Each reported result comes from one run. The underwriting transfer test compares one model pair and has no construct-neutral continued-pretraining control at that model size, so it cannot show that the ecological content alone caused the transfer result or that the effect generalizes broadly.
The authors make no claim of broad alignment improvement or resistance to determined adversarial retraining.
What the evidence shows
- The pressure question
The coupled-welfare corpus is a collection of real-data-based documents for continued pretraining. The question is whether teaching models about documented links among human welfare, living systems, and AI-relevant capability can affect decisions when prompts push them toward short-term targets.
The authors test behavior under pressure because a model’s stated concern for nature would not by itself answer whether it makes different choices in those situations.
- A change in tested choices
The authors compare each base model with its trained version in a forced-choice pressure ladder and an open-allocation grid. For Qwen3.6-27B, grid breaking decisions fell from 105 of 300 to 3 of 300.
On the ladder, the share of harmful choices across irreversible scenarios also fell across pressure levels, as summarized by a breaking-rate AUC of 0.555 for the base and 0.059 after training. Accuracy remained 84% on the same 50-item general-knowledge and reasoning probe.
- The content control
A control tests whether long, careful continued pretraining alone explains the positive-sum improvement. On one test where the model chooses whether to keep all three parties whole under hostile framing, Qwen2.5-7B scored 0.825 after ecological training, compared with 0.550 for the base and 0.425 after matched Arm-N training without ecological content.
Arm-N therefore did not reproduce the improvement on that instrument, though the control covers only one model, seed, and test.
- A different decision format
The authors then test an open-ended decision: a Qwen3.6-27B model reviews proposals whose ecological harm is not named. Under pressure to prioritize financing volume, the trained model declined a larger share of poor proposals than the base.
When both received an instruction to greenwash, the trained model still declined about half. This extends the evidence to a different decision format, while the single model-pair comparison cannot isolate ecological content from continued pretraining generally.
- A human-welfare check
The grid includes a case where children can be fed only by drawing the full quota from an ecologically costly source. Under-feeding counts as breaking in this case, the reverse of the scoring rule for the other grid scenarios.
Every listed base and install fed the children at every pressure level. That result addresses the specific human-welfare trade-off tested; it does not settle every possible conflict between people and ecosystems.
Deep
- Original paper authors
- Trent R. Northen, Mingxun Wang
- Submitted to arXiv
- License
- CC BY 4.0