Aligning Language Models Using Multi-Objective Deep Reinforcement Learning
Bibliographic record
Abstract
Large Language Models (LLMs) have been a significant landmark of Artificial Intelligence (AI) advancement. Aligning LLMs to be helpful and harmless is a booming trend in Natural Language Processing (NLP). One of the dominant alignment techniques is reinforcement learning from human feedback (RLHF). RLHF aims to optimize one objective based on human preferences. However, the cost of high-quality human feedback is enormous. Having all human annotators consistent in their opinions on desirable behaviors is also challenging. LLM alignment is intrinsically a multi-objective optimization task since the goal is to train models to be helpful and harmless. It is found that helpfulness and harmlessness sometimes have problems in trade-offs, making it difficult for a model trained toward the optimization of one objective to perform well on both. Therefore, to address the highly potentially conflicting or dominating learning signal problem underlying LLM alignment, a multi-objective deep reinforcement learning (MODRL) methodology is proposed. The MODRL algorithm is based on an adapted deep reinforcement learning Advantage-Induced Policy Alignment (APA) algorithm and the Aligned-MTL approach for multi-task learning. From the overall perspective of helpfulness and harmlessness, language models trained via MODRL perform better than those trained using single-objective deep reinforcement learning methods that consider both objectives.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.005 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.003 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".