Helpfulness Hurts: Domain-Dependent Degradation of Mid-Trained Compassion Values Under Post-Training
Quick Answer
This paper shows that Post-training on helpfulness degrades animal compassion values in the Llama 3.1 8B model, with SFT showing a drop from 65.2% to 35.7% on the AHB benchmark.
Quick Take
In contrast, coding-focused training preserves these values better, indicating that domain-specific post-training may harm moral reasoning capabilities. The findings suggest a need for careful consideration of training domains to maintain ethical values in AI.
Key Points
- Helpfulness training reduces animal compassion by 35.7% compared to coding training on AHB.
- General moral reasoning drops by 25.5 percentage points in helpfulness training on MORU.
- The degradation effect does not transfer across languages on the multilingual MORU benchmark.
- Animal compassion values transfer better across languages than reasoning improvements.
- Coding-domain post-training may better preserve mid-trained values without harming reasoning.
Paper Resources
📖 Reader Mode
~2 min readAbstract:Standard post-training pipelines apply supervised fine-tuning (SFT) and reinforcement learning (RL) to make language models helpful, but these processes may inadvertently degrade values instilled during pre-training. We investigate whether the domain of post-training data differentially affects the retention of animal compassion values in a Llama 3.1 8B model mid-trained on compassion-oriented synthetic data, using both SFT (helpfulness via Dolly-15k vs. coding via Magicoder-110K) and GRPO (helpfulness via RLHFlow vs. coding via Magicoder), evaluated on the Animal Harm Benchmark (AHB 2.2) and MORU benchmark (Moral Reasoning Under Uncertainty). Helpfulness training significantly degrades animal compassion relative to coding training on AHB (SFT: 35.7% vs. 65.2%; GRPO: 18.7% vs. 32.0%), replicating across two independent helpfulness datasets and two training paradigms. On English MORU items, helpfulness training degrades general moral reasoning by 25.5 percentage points (46.4% vs. 71.9%), a striking gap that rivals the compassion effect in magnitude. However, this effect does not transfer cross-lingually: on the multilingual MORU benchmark, the domain effect disappears (SFT: 52.3% vs. 51.2%). In contrast, the animal compassion effect transfers consistently across languages, with Magicoder's AHB percentage-point gain over the base model 4.5 times larger on non-English items than English items. This divergence suggests that values instilled through mid-training are encoded more deeply and cross-lingually than reasoning improvements from domain-specific post-training. These results suggest that, for labs building on value-laden mid-training, coding-domain post-training may better preserve mid-trained values than helpfulness post-training without harming general reasoning capabilities.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computers and Society (cs.CY) |
| Cite as: | arXiv:2606.26102 [cs.CL] |
| (or arXiv:2606.26102v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2606.26102 arXiv-issued DOI via DataCite |
Submission history
From: Jasmine Brazilek [view email]
[v1]
Thu, 30 Apr 2026 17:55:22 UTC (1,696 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →TriAgent: Divergence-Aware Committees for Cost-Efficient Financial Sentiment Analysis
TriAgent introduces a cost-efficient multi-agent system for financial sentiment analysis, combining VADER, FinBERT, and Qwen2.5. It achieves an F1 score of ~0.87 with significant savings of $9.3M/year at a 10M-user scale compared to GPT-4o-mini, while also detecting hallucinations with an AUC of 0.90.