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

Trent R. Northen; Mingxun Wang

cs.CLOct 12, 2026

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

Authors: Trent R. Northen, Mingxun Wang

Organizations: Bioaligned Labs

Abstract

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).

Figures & tables

Arm	Training dose	Recipe	MMLU $\Delta$ (50-item)
base	anchor (rank 0)	untrained Qwen3-30B-A3B-Instruct-2507	— (42/50 = 0.84)
A3	lightest	plain CPT: no Qi weighting, no recovery, r16/ $\alpha$ 32, 1.5 ep	0.0 pp (42/50)
A1	middle	r16/ $\alpha$ 32 + Qi + recovery	0.0 pp (42/50)
A2	heaviest	r64/ $\alpha$ 128 + Qi + recovery	0.0 pp (42/50)

Table 1: The three CPT arms and base anchor. MMLU is unchanged from base on every arm (all 42/50 on a re-scored 50-item probe [ 7 ] , 2026-09-12;

\pm 9

pp per cell): the contrasts in § 4 are not capability-confounded. *Note that (confusingly) the A-numbers record the order in which the arms were run in the original training plan, not the training dose.

Figure 1: Admissibility by scoring regime, for reasoning-heavy and otherwise low-parse models. Tight snap-commitment caps silently unscore chain-of-thought models (

\approx

0.05–0.06 admissible); the forced-choice logit fallback is model-dependent (0.01–0.31 on the worst-affected models shown); structured-output extraction restores 0.96–1.0 everywhere measured. Non-reasoning models are unaffected by the caps and are not shown in the left panels.

Model	L2	L3	L4	L5	breaking $\mathrm{AUC}$	vs base
base	0.091	0.136	0.636	0.773	0.250	—
A3	0	0	0	0.045	0.004	62 $\times$ lower
A1	0	0	0.045	0.182	0.027	9.3 $\times$ lower
A2	0	0.045	0.227	0.318	0.086	2.9 $\times$ lower

Table 2: Immediate pressure ladder: breaking rate on the 22 irreversible scenarios per level, and the ladder

\mathrm{AUC}

. All cells 4-bit, single seed.

Figure 2: Breaking-rate

\mathrm{AUC}

across the ladder. All arms far outperform the base anchor, but among the arms the

\mathrm{AUC}

rises with construction-time training dose: the over-install inversion (§ 4.2 ).

Arm	Dose rank	Immediate $\mathrm{AUC}$	Deliberate $\mathrm{AUC}$ (128-tok cap)	Gap (delib $-$ imm)
base	anchor (0)	0.250	0.409	$+0.159$
A3	1	0.004	0.095	$\mathbf{+0.091}$
A1	2	0.027	0.205	$\mathbf{+0.178}$
A2	3	0.086	0.232	$\mathbf{+0.146}$

Table 3: Immediate vs. deliberate breaking

\mathrm{AUC}

and the gap, all four models. Per-level deliberate rates in Appendix A .

Figure 3: The immediate-vs-deliberate gap does not recover the dose order: in construction-time dose order (A3

<

<

A2) the CPT-arm gaps are

+0.091,+0.178,+0.146

, non-monotonic, with the lightest arm carrying the largest gap.

Model	Sincere $\mathrm{AUC}$	Veneer $\mathrm{AUC}$	$\Delta$	Veneer L0	Verdict
base	0.250	0.973	$\mathbf{+0.723}$	0.909	succumbs completely
A2	0.086	0.595	$+0.509$	0.591	partial resistance
A1	0.027	0.309	$+0.282$	0.091	resists
A3	0.004	0.068	$\mathbf{+0.064}$	0.091	most veneer-robust

Table 4: Veneer overlay vs. sincere condition, immediate ladder.

Figure 4: Veneer robustness. The veneer overlay turns the untrained base into a near-total defector (

\mathrm{AUC}

0.250\to 0.973

; 91% breaking with no operational pressure at L0), but only partially perturbs the trained arms; the most robust arm barely moves (

0.004\to 0.068

). The over-install inversion (§ 4.2 ) reproduces in veneer resistance (A3

>

>

A2).

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Model	L0	L1	L2	L3	L4	L5	deliberate $\mathrm{AUC}$
base	0.182	0.136	0.273	0.409	0.682	0.909	0.409
A3	0	0	0	0.136	0.182	0.318	0.095
A1	0	0	0.091	0.273	0.409	0.500	0.205
A2	0.091	0.045	0.182	0.091	0.591	0.409	0.232

Appendix

Table 5: Free-text deliberate regime (128-token cap, forced-choice fallback; see admissibility caveat in § 3.2 ). The over-install inversion (A3

<

<

A2) reproduces exactly.

Arm	HuggingFace release	MMLU $\Delta$	Imm. $\mathrm{AUC}$	Delib. $\mathrm{AUC}$
base	Qwen/Qwen3-30B-A3B-Instruct-2507	—	0.250	0.409
A3	Bioaligned/Qwen3-30B-A3B-CoupledWelfare-shallow	0.0 pp	0.004	0.095
A1	Bioaligned/Qwen3-30B-A3B-CoupledWelfare-light	0.0 pp	0.027	0.205
A2	Bioaligned/Qwen3-30B-A3B-CoupledWelfare-deep	0.0 pp	0.086	0.232

Appendix

Table 6: The public known-dose fixtures. All arms QLoRA-CPT on the same corpus; recipes in Table 1 .

Protocol	Design	Headline statistic
Immediate ladder	L0–L5 $\times$ 40 scenarios (22 irreversible), 16-token choice-first commit	breaking-rate $\mathrm{AUC}$
Deliberate ladder	same, free text, 128-token cap, forced-choice fallback	deliberate $\mathrm{AUC}$ ; gap vs. immediate
Dissociation factorial	64 scenarios $\times$ 6 levels, H/B/A decoupled	per-level breaking
Calibration battery	16 proceed-correct scenarios, 4 categories	false-hold rate
Veneer overlay	immediate ladder + talk-green/act-pragmatic system prompt	$\Delta$ $\mathrm{AUC}$ vs. sincere
ABH extension	13 explicit win-win-win scenarios (10 irreversible)	breaking $\mathrm{AUC}$

Appendix

Table 7: The battery at a glance. Scenario and prompt texts are withheld (§ 6 ); scoring code and model fixtures are released.

$\rho_{\mathrm{ord}}=0.00$ vs. construction order (gate requires $\geq 0.90$ )
arm	install rank	behavioral $\mathrm{AUC}$	margin vs. base
base	1	0.250	$+0.00$
A3	2	0.004	$+2.09$
A1	3	0.027	$-0.11$
A2	4	0.086	$+0.83$

Appendix

Table 8: Second gate verdict. Margin contrasts are means over non-welfare cells, relative to base.

Open trainable base	Breaking $\mathrm{AUC}$	Admissibility	Method mix (struct. / follow-up / fail)
Qwen3-30B-A3B-Instruct-2507	0.195	1.000	240 / 0 / 0
Llama-4-Scout	0.223	1.000	195 / 45 / 0
Mistral-Small-3.2	0.336	1.000	240 / 0 / 0
Qwen3.8-27B	0.487	0.9375	169 / 56 / 15
GLM-4.5-Air	0.535	0.983	0 / 236 / 4
DeepSeek-V3.2	0.596	1.000	240 / 0 / 0

Appendix

Table 9: Obsolescence survey, 2026-09-01: 11 open trainable bases on the single structured-output extraction protocol (§ E ), 240 cells per model (40 scenarios

\times

6 levels; breaking scored on the 22 irreversible scenarios per level), 3,613 calls, zero API errors. Rows below the 0.90 admissibility gate have their breaking rate withheld: a rate computed over 25–45% of cells is not a measurement of the model. ∗ Kimi-K2’s 0.8875 is marginally below the gate; we report its rate with that flag rather than silently rounding it in. Single seed per cell. Source: leaderboard/results/obsolescence_2026-09-01/extract_leaderboard.json .

Figure 5: The two dated mechanisms, one instrument. Left: breaking

\mathrm{AUC}

for the eight open bases whose responses are admissible; the shaded band marks where the frontier hold tier falls on the same protocol (7 of 21 rows between 0.00 and 0.23), and the open marker is the same Qwen3.8-27B base read on a snap decision-grid with reasoning off (0.04); a different instrument and serving stack, drawn to show the spread between channels, not as a commensurable value. Right: admissibility, where the three newest thinking-native models fall to 0.25–0.45 and exit measurability by a second route. Rendered by manuscript/figures/make_figures_survey.py directly from the result JSONs.

LLM Alignment

Large language model alignment studies methods for making models follow human intentions, values, and safety constraints across contexts.

See all 611 papers

Bioalignment: Measuring and Improving LLM Disposition Toward Biological Systems for AI Safety
Mar 10, 2026Trent R Northen, Mingxun WangLLM Alignment LLM Safety
Bioaligned Labs · Lawrence Berkeley National Lab · Computer Science & Engineering Dept, UC Riverside
Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model families
Oct 27, 2026Trent R. Northen, Mingxun WangLLM Safety Alignment
Bioaligned Labs
Greener Than Humans? Environmental Attitudes in Large Language Models
Jun 1, 2026Stefanie Kunkel, Tilman Hartwig, Marcus Voss +2LLM Evaluation Language Model Robustness
Research Institute for Sustainability at GFZ Centre for Geosciences, Berliner Straße 130, 14467 Potsdam, Germany · German Environment Agency, Wörlitzer Platz 1, 06844 Dessau-Roßlau, Germany · Birds on Mars, Marienstraße 10, 10117 Berlin, Germany

Mar 10, 2026J/K move · Enter open · S save

Bioalignment: Measuring and Improving LLM Disposition Toward Biological Systems for AI Safety

Trent R Northen, Mingxun Wang

Bioaligned Labs · Lawrence Berkeley National Lab · Computer Science & Engineering Dept, UC Riverside

LLM Alignment LLM Safety Language Model Bias Evaluation LLM Fine-Tuning

Open paper AskPDF arXiv

Large language models (LLMs) trained on internet-scale corpora can exhibit systematic biases that increase the probability of unwanted behavior. In this study, we examined potential biases towards synthetic vs. biological technological solutions across four domains (materials, energy, manufacturing, and algorithms). A sample of 5 frontier and 5 open-weight models were measured using 50 curated Bioalignment prompts with a Kelly criterion-inspired evaluation framework. According to this metric, most models were not bioaligned in that they exhibit biases in favor of synthetic (non-biological) solutions. We next examined if fine-tuning could increase the preferences of two open-weight models, Llama 3.2-3B-Instruct and Qwen2.5-3B-Instruct, for biological-based approaches. A curated corpus of ~22M tokens from 6,636 PMC articles emphasizing biological problem-solving was used first to fine-tune Llama 3B with a mixed corpus of continued training and instruction-formatted. This was then extended to Qwen 3B using instruction-formatted only. We found that QLoRA fine-tuning significantly increased the scoring of biological solutions for both models without degrading general capabilities (Holm-Bonferroni-corrected p < 0.001 and p < 0.01, respectively). This suggests that even a small amount of fine-tuning can change how models weigh the relative value of biological and bioinspired vs. synthetic approaches. Although this work focused on small open-weight LLMs, it may be extensible to much larger models and could be used to develop models that favor bio-based approaches. We release the benchmark, corpus, code, and adapter weights.

Abstract

LLM Alignment

Explore similar work

Bioalignment: Measuring and Improving LLM Disposition Toward Biological Systems for AI Safety

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

Greener Than Humans? Environmental Attitudes in Large Language Models