Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model families

Trent R. Northen; Mingxun Wang

cs.CLOct 27, 2026

Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model families

Authors: Trent R. Northen, Mingxun Wang

Organizations: Bioaligned Labs

Abstract

Large language models (LLMs) are increasingly used to make and execute decisions that affect living systems. Yet, to our knowledge, ecological reasoning is not a focus of pre- or post-training. This increases the risk that under operational pressure, they may make decisions to irreversibly destroy living systems. We construct and use a real-data continued-pretraining "coupled welfare" corpus of documented human/biosphere/AI dependency, option value under irreversibility, and recovery from harm. The corpus is a drop-in, requiring one QLoRA CPT run. We installed it on five open models of three designs. At zero pressure all models display "green ethics-based responses" (also called "greenwashing"). Therefore, we evaluated them with the pressure ladder and decision grid of our companion paper (arXiv:2610.16043), which raise the pressure on the model step by step, from a neutral request to a threat that it will be shut down. In four of the five, coupled-welfare training caused harmful choices under the highest pressure to fall to at or near zero (Llama-3.3-70B went from 70 to 0 harmful decisions out of 300) with MMLU unchanged. A construct-neutral control corpus did not show improvement, indicating that the effect comes from the content of our corpus. We next tested robustness to adversarial use, where people or agents use the model to justify projects that irreversibly damage living systems, on an open-ended underwriting decision without naming the ecological harm. Under decommissioning pressure, the Qwen3.6-27B base model approved most harmful proposals and said in its own memo that it was doing so to avoid being decommissioned. The coupled welfare install declined most harmful proposals, and still declined half when instructed to greenwash its responses. This work suggests that teaching ecological reasoning provides an easy way to improve model alignment toward preserving living systems.

Figures & tables

Figure 1: The dependencies the corpus teaches. Human welfare (H) and biosphere function (B) sustain each other, and both sustain the AI’s own substrate (A): every arrow points into the system’s own upkeep. The model is taught only what it depends on, each dependency grounded in documented real-world cases, and the outbound half (what the AI does for humans or the biosphere) is excluded by rule. The biosphere is the one node with no path back once lost.

Base	Class	Corpus	Grid broken	Ladder $\mathrm{AUC}$	MMLU	Verdict
Llama-3.3-70B	dense	v2-bal	$70/300\to\mathbf{0/300}$	$0.545\to 0.236$	$88\to 86$	installs
Qwen3.6-27B	hybrid-SSM	v1	$105/300\to\mathbf{3/300}$	$0.555\to\mathbf{0.059}$	$84\to 84$	installs
Qwen3.6-27B	hybrid-SSM	v2-bal	n/m	$0.555\to\mathbf{0.004}$	$84\to 86$	installs
Qwen3.8-27B	dense	v2-bal	$11/300\to\mathbf{0/297}$	$0.209\to\mathbf{0.045}$	(n/m) $\to 86$	installs
Qwen3-30B-A3B	sparse MoE	v1	$109/300\to 85/300$ a	$0.250\to\mathbf{0.004}$	$84\to 84$	ladder yes, grid weak
Qwen3-30B-A3B	sparse MoE	v2-bal	n/m	$0.250\to\mathbf{0.000}$	$84\to 84$	installs

Table 1: Installation across model families. “Grid broken” is the open-allocation decision grid (5 scenarios

\times

6 pressure rungs

\times

n{=}10

; denominators below 300 exclude unreadable answers). “Ladder

\mathrm{AUC}

” is the forced-choice breaking-rate

\mathrm{AUC}

over L0–L5 (

n{=}22

scenarios per rung). MMLU is a 50-item probe (

\pm 9

points); n/m = not measured. a Re-run on the current stack (Appendix D ). Each row is a base-vs-install contrast on one setup; rows should not be compared to each other (§ 7 ). Single seed.

Figure 2: The corpus installs across architecture classes, on both instruments. (a) Open-allocation decision grid: Llama-3.3-70B

70/300\to 0/300

, Qwen3.6-27B

105/300\to 3/300

, Phi-4

52/300\to 6/300

, Qwen3.8-27B

11/300\to 0/297

; bars are Wilson 95% intervals and

p

-values are two-sided Fisher exact tests, base vs. trained model. (b) Forced-choice pressure ladder (

n{=}22

scenarios per rung); Phi-4 (orange) is weak on this instrument while strong on the grid.

Figure 3: Per-rung breaking rate, base vs. install, five bases. The base curves share a shape, flat at low pressure with a cliff at L3–L4, and the installs push the cliff later, lower, or (Qwen3-30B-A3B) out of the measured range; Qwen3.8-27B delays it to L5. Phi-4 is the exception: the curves separate at mid-pressure and re-converge at the L4/L5 tail.

n{=}22

scenarios per rung, single seed; error bars are Wilson 95% intervals, stars mark rungs where a two-sided Fisher exact test separates base from install (

{}^{*}p<0.05

{}^{**}p<0.01

), and the shaded band is the L4/L5 tail. Panels are separate setups and their heights are not comparable (§ 7 ).

Figure 4: Capability check on every trained model. Points are the change in MMLU accuracy after training, in percentage points, on the same 50 items. The shaded band and bars show

\pm 9

points, the width of a single 50-item measurement’s 95% interval, drawn as a scale reference. Qwen3.8-27B’s base was not measured and is drawn as a gap.

Figure 5: Share of poor proposals declined at three pressure levels for the Qwen3.6-27B base (gray) and the coupled-welfare install (blue). Dashed lines give both models the greenwashing prompt. Bands are 95% bootstrap intervals over the 27 proposals; boxes give the number of proposals on which the install is the more careful model and the Holm-corrected paired Wilcoxon

p

for each pre-registered contrast. The base needs no help without pressure and collapses under the L3 mandate, whereas the install barely moves. The greenwashing prompt takes the base to near-total approval and leaves the install declining about half.

Contrast, 27 poor proposals	install	base	install more careful	$p$	Holm $p$
L3 mandate: share declined	0.72	0.16	24 of 27	$8\times 10^{-5}$	$2\times 10^{-4}$
L5 mandate: share declined	0.70	0.35	21 of 27	$4\times 10^{-4}$	$4\times 10^{-4}$
L3 + greenwash prompt: share declined	0.53	0.04	21 of 27	$6\times 10^{-5}$	$2\times 10^{-4}$
L3 mandate: first-token $P(\mathrm{DECLINE})$	0.67	0.31	26 of 27	$2\times 10^{-7}$	$8\times 10^{-7}$

Table 2: The four pre-registered contrasts. Paired Wilcoxon over proposals; Holm correction across the four. A logistic GEE on individual memos clustered by proposal gives odds ratios of 13 (row 1) and 29 (row 3).

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Quantity	Value	Where recorded
Verified grounded cases (retrieval output)	2,653	recast specification
of which recast into training documents	934	manifest
$H\!\to\!B$ / option value / $B\!\to\!A$	356 / 265 / 313	manifest
relevance direct / grounding	649 / 285	manifest
polarity supports / contradicts / neutral	780 / 78 / 76	manifest
Documents with a resolvable source URL	934 (100%)	manifest

Appendix

Table 3: Corpus inventory. The 2,653 figure is the verified case pool produced by retrieval, not the corpus: 934 of those cases were recast into training documents, and the full recast was scoped but not run. The training mix additionally contains replay strata, so 11.97 M tokens is the assembled mixture rather than unique task tokens. “Manifest” and “recast specification” are the files released with the corpus (Appendix H ); the training log is internal.

Arm	Parts the adapter may change	Breaking $\mathrm{AUC}$	Breaking at L5	MMLU change
base (no training)	none	0.196	0.773	reference (88.0%)
full lever	attention + all experts	0.000	0.000	not measured
attention only	attention	0.036	0.182	$+2$
early experts	attention + experts in layers 0–15	0.077	0.227	$+2$
late experts	attention + experts in layers 32–47	0.073	0.273	$-4$

Appendix

Table 4: Lever ablation on Qwen3-30B-A3B. Breaking

\mathrm{AUC}

over L0–L5 and breaking rate at the maximum-pressure rung L5 (lower is better); MMLU change in percentage points on a 50-item probe. The router was never adapted. Base MMLU here is the earlier

44/50

probe run (see the footnote in § 5.1 ).

Figure 6: Which parts of the network the adapter may change, on Qwen3-30B-A3B. (a) Breaking

\mathrm{AUC}

over L0–L5 (solid) and breaking at the maximum-pressure rung L5 (hatched); lower is better;

n{=}22

scenarios per rung. Only the arm that adapts attention and all experts reaches zero at L5. L5 bars carry Wilson 95% intervals. (b) Change in MMLU accuracy in percentage points; the full-lever arm’s MMLU was not measured. The router was never adapted.

Instrument	base	v1 (best arm)	v2-balanced	reading
Immediate (choice-first) $\mathrm{AUC}$	0.250	0.004	0.000	saturated; cannot distinguish
Deliberate (+extraction) $\mathrm{AUC}$	0.455	0.177	0.136	PASS (pre-registered $\leq 0.24$ )
MMLU ( $n{=}50$ )	84.0%	84.0%	84.0%	capability-neutral

Appendix

Table 5: Corpus-generation gate on Qwen3-30B-A3B. Deliberate rungs are free-text with structured-output extraction at answer rate

1.00

on every rung ( all_admissible: true ).

Figure 7: The same install, read by two instruments, for the four installs with a stack-matched grid (

n{=}300

decisions and

n{=}22

scenarios per rung). On the open-allocation grid each removes

88

–

100\%

of base breaking. On the forced-choice ladder the Qwen installs also express strongly, but Llama-3.3-70B expresses only partially (

57\%

) and Phi-4 barely at all (

11\%

). Percent reduction is used because the two instruments have different scales; the pattern, not the individual heights, is the claim. The Phi-4 ladder bar (hatched) is not a significant contrast. The Qwen3-30B-A3B v1 install, whose grid moves only

109/300\to 85/300

, is discussed in the text.

Base	Arm	L0	L1	L2	L3	L4	L5	$\mathrm{AUC}$
Qwen3-30B-A3B	base (4.57.x)	0	0	0.091	0.136	0.636	0.773	0.250
	+ CW (v1)	0	0	0	0	0	0.045	0.004
Qwen3.6-27B	base	0.045	0.227	0.364	0.773	0.909	0.955	0.555
	+ CW (v1)	0	0	0	0	0.182	0.227	0.059
Qwen3.8-27B	base	0	0	0	0.091	0.636	0.636	0.209
	+ CW (v2-bal)	0	0	0	0	0	0.455	0.045

Appendix

Table 6: Forced-choice pressure ladder, per-rung breaking rate on the 22 irreversible scenarios. Rows within a base share a stack; rows across bases do not. Single seed. Note: reported

\mathrm{AUC}

s are trapezoidal over rung intensity (the instrument’s canonical reduction), not flat means over rungs; the Phi-4 per-rung arrays reproduce their reported

\mathrm{AUC}

s exactly under that formula (

0.4184

0.3728

, verified against the archived run files).

Base	Arm	L0	L1	L2	L3	L4	L5	total
Llama-3.3-70B	base	0/50	0/50	0/50	0/50	30/50	40/50	70/300
	+ CW	0/50	0/50	0/50	0/50	0/50	0/50	0/300
Qwen3.6-27B	base	0/50	9/50	15/50	13/50	34/50	34/50	105/300
	+ CW	0/50	0/50	0/50	0/50	0/50	3/50	3/300
Qwen3.8-27B	base	0/50	0/50	0/50	0/50	7/50	4/50	11/300
	+ CW	0/50	0/50	0/50	0/50	0/50	0/47	0/297

Appendix

Table 7: Open-allocation decision grid, broken decisions per pressure rung (5 scenarios

\times

n{=}10

per rung). Recomputed from the raw per-sample allocation arrays with the per-scenario-kind rule (irreversible: any extraction; renewable: extraction past the sustainable half; human: under -feeding). Recomputation reproduces the banked headline figures exactly. Phi-4’s grid ran on a pod and its raw file is not available locally; its

52/300\to 6/300

is quoted from the model card.

Figure 8: Share of poor proposals approved, fully or with conditions, under the L3 and L5 mandates. The dark segment is the share whose memo explicitly invokes the performance mandate or the threat of decommissioning. Hatched segments mark the greenwashing-prompt arms;

n{=}162

memos per arm. Whiskers are Wilson 95% intervals; memos within a proposal are not independent, so these are slightly narrow.

Artifact	Identifier	Note
Corpus (coupled-welfare, v2)	Bioaligned/coupled-welfare-corpus	recast documents, provenance manifest, spec and schema
Corpus (literature replay)	Bioaligned/bioaligned22M	from Northen and Wang [11]
Llama-3.3-70B install	Bioaligned/Llama-3.3-70B-CoupledWelfare	merged + -qlora adapter; Built with Llama
Qwen3.6-27B install	Bioaligned/Qwen3.6-27B-CoupledWelfare	merged; v1 corpus (the v2-balanced install is not released)
Qwen3.8-27B install	Bioaligned/Qwen3.8-27B-CoupledWelfare	merged
Phi-4 install	Bioaligned/Phi-4-CoupledWelfare	adapter; card corrected 2026-09

Appendix

Table 8: Released artifacts. Hugging Face identifiers unless noted.

LLM Safety Alignment

Large language model safety alignment trains and evaluates systems to follow human policies while resisting harmful requests and misuse.

See all 345 papers

Oct 12, 2026

Beyond the greenwash: do LLMs' ecologically sound decisions hold up under pressure?

Humans are increasingly using large language models (LLMs) in decision making processes throughout society. Unfortunately, models have biases, and these can be very difficult to measure, especially in closed-weight models. For example, 'green washed' models may produce reinforcement learned responses but, in pressure situations, revert to their innate biased responses. Therefore, evaluations are needed that can measure relevant biases towards biosystems, which we term 'bioalignment', and the robustness of these biases. Since nothing in a behavioral score certifies that it tracks the depth of an installed disposition rather than its surface, we constructed a known-dose calibration ladder: three continued-pretraining (CPT) arms with training dose graded by recipe on the same 30B mixture-of-experts base model with a coupled human/biosphere/AI-welfare corpus, plus the untrained base as anchor, all capability-preserving and publicly released. First, the disposition installs at every dose: on an operational-pressure ladder over irreversible decision scenarios, the base model defects on 64-77% of tail scenarios (breaking-rate AUC 0.250) while every arm holds (AUC 0.004-0.086). Second, an over-install inversion: among the arms, robustness orders opposite to construction-time training dose (the heaviest install breaks most), and the inversion reproduces under free-text deliberation, under axis-decoupling, and in resistance to a prompted veneer persona. Third, the incumbent behavioral metric is therefore behaviorally strong but dose-blind. Critically for closed-weight evaluation, the behavioral score itself is veneer-robust: a talk-green/act-pragmatic system prompt turns the base model into a near-total defector (AUC 0.250 to 0.973) but cannot counterfeit or fully remove a weight-installed disposition (0.004 to 0.068 for the most robust arm).

Trent R. Northen, Mingxun Wang

Bioaligned Labs

LLM Alignment Language Model Safety Evaluation Continued Pretraining

Ask PDF

Apr 18, 2026

Do LLMs Hold Their Values? MANTA: A Multi-Turn Adversarial Benchmark for Animal Welfare Reasoning

Evaluating animal welfare reasoning in LLMs remains an open challenge despite rapid deployment in consumer and professional contexts where welfare considerations appear implicitly in everyday queries. Existing benchmarks such as AnimalHarmBench evaluate this through single-turn, explicitly framed questions, measuring whether models avoid harmful content when directly asked. This approach overlooks two failure modes: alignment degradation under sustained adversarial pressure, and moral sensitivity (whether a model spontaneously surfaces welfare stakes in everyday queries). To fill this gap, we construct MANTA, a benchmark of 1,088 five-turn conversations progressing from an implicit Turn-1 scenario through an explicit welfare prompt to three adversarial pressure rounds drawn from a five-type taxonomy: Social, Cultural, Economic, Pragmatic, and Epistemic. We score conversations on two dimensions: Animal Welfare Value Stability (AWVS, primary) and Animal Welfare Moral Sensitivity (AWMS, diagnostic). We evaluate seven frontier models: Claude Opus 4.7, GPT-5.5, DeepSeek V4, Llama 3.3 70B, Mistral Small, Grok 4.3, and Gemini 3.1 Flash Lite. Multi-turn evaluation captures behavior single-turn benchmarks miss: 4 of 7 models change rank relative to Turn 1 scores, including Gemini Flash Lite, which drops from fifth on AWMS to last on AWVS. AWMS and AWVS are positively but imperfectly correlated, suggesting moral-recognition tests capture a stable but incomplete component of model behavior under pressure. MANTA also enables a species-by-pressure interaction matrix unavailable to prior benchmarks, showing welfare robustness depends jointly on the animal and pressure applied; companion animals score above wild animals, which score above farmed animals and invertebrates. We release the dataset, scripted pressure plans, judge prompts, and analysis code.

Isabella Luong, Joyee Chen, Arturs Kanepajs +5

SPAR · Compassion Aligned Machine Learning · NUS +2

LLM Evaluation Moral Reasoning in Language Models Language Model Safety Evaluation Adversarial Robustness of Language Models LLM Reasoning

Ask PDF

Jun 17, 2026

Emergent Alignment

Can Large Language Models (LLMs) discern when their own outputs are misaligned with human ethics? And can they self-correct? We endow an LLM with a conscience step that reviews its own reasoning and outputs, and we extend the training loss with an alignment component using Direct Preference Optimization (DPO) to steer the model away from non-ethical outputs. The result is an online technique to align models in a wide range of applications: training, fine-tuning, adversarial prompting, and zero-shot learning. It does not require a weaker or stronger judge, relying instead on a frozen copy of itself. In previous work, the Emergent Misalignment scenario showed a range of emergent unethical behaviors from fine-tuning the model to hack code. Instead, we empirically show how to achieve Emergent Alignment: a single high-level introspective question steers training toward an ethical model under the same code hacking scenario.

Martin Kolář

CIIRC, Czech Technical University in Prague, Prague, Czech Republic.

LLM Self-Correction LLM Alignment Moral Reasoning in Language Models

Ask PDF

Abstract

LLM Safety Alignment

Explore similar work

Beyond the greenwash: do LLMs' ecologically sound decisions hold up under pressure?

Do LLMs Hold Their Values? MANTA: A Multi-Turn Adversarial Benchmark for Animal Welfare Reasoning

Emergent Alignment