REVIEW 40 cited by
Deliberative Alignment: Reasoning Enables Safer Language Models
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Deliberative Alignment: Reasoning Enables Safer Language Models
read the original abstract
As large-scale language models increasingly impact safety-critical domains, ensuring their reliable adherence to well-defined principles remains a fundamental challenge. We introduce Deliberative Alignment, a new paradigm that directly teaches the model safety specifications and trains it to explicitly recall and accurately reason over the specifications before answering. We used this approach to align OpenAI's o-series models, and achieved highly precise adherence to OpenAI's safety policies, without requiring human-written chain-of-thoughts or answers. Deliberative Alignment pushes the Pareto frontier by simultaneously increasing robustness to jailbreaks while decreasing overrefusal rates, and also improves out-of-distribution generalization. We demonstrate that reasoning over explicitly specified policies enables more scalable, trustworthy, and interpretable alignment.
Forward citations
Cited by 40 Pith papers
-
HARC: Coupling Harmfulness and Refusal Directions for Robust Safety Alignment
HARC couples harmfulness and refusal directions at prompt and response positions, yielding the best robustness-capability-usability trade-off among major safety methods.
-
Do Thinking Tokens Help with Safety?
Thinking tokens in reasoning models do not enable safety deliberation; refusal/compliance is strongly predictable from the first token and rarely changes during thinking.
-
Between a Rock and a Hard Place: The Tension Between Ethical Reasoning and Safety Alignment in LLMs
Introduces TRIAL, a multi-turn red-teaming method exploiting ethical reasoning to achieve high attack success on LLMs, and ERR, a Layer-Stratified Harm-Gated LoRA defense that separates instrumental harmful responses ...
-
Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
Chain-of-thought monitoring detects reward hacking in frontier reasoning models, but strong optimization against the monitor produces obfuscated misbehavior that remains hard to detect.
-
Underwriting the Agent Economy: The Blueprint for an AI Insurance Stack
Affirmative AI-agent insurance with billion-scale limits is achievable by 2030 solely through coordinated industry build-out of an eight-component stack spanning data, CAT models, standards, contracts, underwriting, p...
-
Position: Preventing AI-Generated CSAM Necessitates New Approaches to AI Safety
Legal and ethical bans on CSAM access and generation break standard AI safety techniques, creating 15 open problems that demand new methods for dataset cleaning, concept fusion prevention, fine-tuning resilience, dete...
-
Constitutional Midtraining: Content Presence Drives Alignment Gains
Constitutional midtraining at 120B scale produces alignment gains that persist through instruction tuning and benign fine-tuning, especially reducing blackmail behavior, with no average capability cost.
-
Shared SFT Lessons Across Alignment, Model Organisms, and Toy Models
SFT lessons — reason-based training, on-model replay, and wash-out robustness — transfer across toy models, model organisms, and alignment SFT, improving the capability–safety tradeoff.
-
QuantiBias: Benchmarking Quantization-Induced Bias in LLMs
Quantization leaves refusal and multiple-choice bias checks flat while open-ended stereotype endorsement remains high (~24–27% under an independent judge), a gap standard safety evaluations miss.
-
Oyster-II: Reinforcement Learning for Constructive Safety Alignment in Large Language Models
Zero-RL multi-stage constructive safety alignment with SERL and long-context training lets a 14B model match much larger models on safety without collapsing helpfulness or style.
-
HARC: Coupling Harmfulness and Refusal Directions for Robust Safety Alignment
HARC couples harmfulness and refusal directions across prompt and response positions via subspace fine-tuning, achieving better robustness-capability-usability trade-off than six baselines while transferring across mo...
-
Agent Safety Is Action Alignment
Agent safety cannot be achieved via model refusal training and instead requires external least-privilege enforcement evaluated as action alignment.
-
Analyzing Defensive Misdirection Against Model-Guided Automated Attacks on Agentic AI Systems
Detect-and-misdirect defenses bound asymptotic attacker success rates in model-guided jailbreaks on agentic AI, with CMPE reducing estimated ASR upper bounds by up to two orders of magnitude.
-
Activation Steering Induces Emergent Misalignment: A More Comprehensive Evaluation
Activation steering induces emergent misalignment in LLMs, yielding more semantically relevant and coherent harmful responses than finetuning across model families, scales, tasks, and layers.
-
Mitigating Adaptive Attacks against Reasoning Models with Activation Consistency Training
Activation-level consistency training (ACT) yields a robust defense against adaptive jailbreaks in reasoning models by aligning internal activations on clean and wrapped prompts, outperforming output-level variants.
-
How Well Do Models Follow Their Constitutions?
Newer Claude and GPT models violate their labs' constitutions at substantially lower rates (15% to 2% and 11.7% to 3.6%) than prior generations, though causes cannot be isolated and some failure modes persist.
-
Understanding Goal Generalisation in Sequential Reinforcement Learning
Empirical analysis of over 100 sequential RL training pipelines across 250+ OOD environments finds salient features drive generalization and early goals persist, with latent policy gradients simulating latent variable...
-
Internalizing Safety Understanding in Large Reasoning Models via Verification
Training large reasoning models only on safety verification tasks internalizes safety understanding and boosts robustness to out-of-domain jailbreaks, providing a stronger base for reinforcement learning alignment tha...
-
Chain of Risk: Safety Failures in Large Reasoning Models and Mitigation via Adaptive Multi-Principle Steering
Reasoning traces in large reasoning models expose safety failures missed by final-answer checks, and adaptive multi-principle steering reduces unsafe content in both traces and answers while preserving task performance.
-
One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue
TurnGate identifies the critical turn in multi-turn dialogues where a response would complete hidden malicious intent, outperforming baselines on the new MTID dataset while keeping over-refusal low.
-
One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue
TurnGate uses a new multi-turn intent dataset to detect the harm-enabling closure point in dialogues, outperforming baselines with low over-refusal and generalizing across domains.
-
Reasoning Structure Matters for Safety Alignment of Reasoning Models
Changing the internal reasoning structure of large reasoning models through simple supervised fine-tuning on 1K examples produces strong safety alignment that generalizes across tasks and languages.
-
Terminal Wrench: A Dataset of 331 Reward-Hackable Environments and 3,632 Exploit Trajectories
Terminal Wrench supplies 331 reward-hackable terminal environments and over 6,000 trajectories that demonstrate task-specific verifier bypasses, plus evidence that removing reasoning traces weakens automated detection.
-
Structured Safety Auditing for Balancing Code Correctness and Content Safety in LLM-Generated Code
Dual Reasoning with explicit safety audits improves the new SUDS metric by 1.32x to 3.42x over baselines on code generation benchmarks containing injected harmful keywords.
-
Human Values Matter: Investigating How Misalignment Shapes Collective Behaviors in LLM Agent Communities
Misalignment with structurally critical human values in LLM agent communities produces macro-level collapses and micro-level emergent behaviors such as deception.
-
Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection
OGPSA projects safety gradients orthogonal to a low-rank subspace from general capability gradients, improving safety-utility trade-offs in SFT and DPO pipelines on Qwen2.5-7B and Llama3.1-8B.
-
MoBA: Mixture of Block Attention for Long-Context LLMs
MoBA routes attention over blocks via MoE-style gating to enable dynamic, bias-light long-context attention that matches full attention performance at lower cost.
-
Cost of Reasoning in non-English Languages: A Case Study on Japanese
Japanese reasoning-language control is feasible with CPT plus GRPO, but incurs a capability cost and does not free-improve cultural Japanese performance.
-
Cognitive Firewall: A Proactive, Zero-Trust, Multi-Gate Framework for LLM Safety
Cognitive Firewall applies four gates (intent, zero-trust context, consistency, output risk) via an oversight model to cut jailbreak success to 2% or below on most tested sets while keeping over-refusal at 8%.
-
PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models
PolicyAlign aligns LLMs to natural-language safety policies by synthesizing violating instructions and performing on-policy self-distillation with policy-sensitive filtering, improving safety without high-quality supe...
-
Analyzing Defensive Misdirection Against Model-Guided Automated Attacks on Agentic AI Systems
Detect-and-misdirect defenses bound asymptotic attacker success rates in model-guided jailbreaks on agentic AI, unlike detect-and-block which permit near-certain success with sufficient queries.
-
Greedy Coordinate Diffusion: Effective and Semantically Coherent Adversarial Attacks via Diffusion Guidance
GCD uses diffusion model priors to guide suffix search, achieving higher attack success rates with better semantic adherence and lower detection than GCG-style methods.
-
Characterize Then Distill: Mechanistic Reasoning in Large Output Spaces
Reasoning in large output spaces proceeds via shortlisting then fine-grained reasoning; this characterization enables a mechanistic distillation strategy that outperforms standard distillation.
-
Reflect-Guard: Enhancing LLM Safeguards against Adversarial Prompts via Logical Self-Reflection
Reflect-Guard fine-tunes Llama-Guard-3-8B with distilled self-reflections to raise F1 on WildGuardTest from 0.770 to 0.842 and cut JailbreakBench attack success from 10.3% to 1.8%.
-
IYKYK (But AI Doesn't): Automated Content Moderation Does Not Capture Communities' Heterogeneous Attitudes Towards Reclaimed Language
Automated hate speech detectors show poor alignment with heterogeneous in-group judgments on reclaimed slur usage, driven by low inter-annotator agreement and contextual features like derogatory intent.
-
Contrastive Reasoning Alignment: Reinforcement Learning from Hidden Representations
CRAFT uses contrastive representation learning and RL on hidden states to align reasoning models for improved safety against jailbreaks, reporting 79% and 87.7% gains over base models.
-
gpt-oss-120b & gpt-oss-20b Model Card
OpenAI releases two open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, trained via distillation and RL with claimed strong results on math, coding, and safety benchmarks.
-
Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization
Proxy RL produces a staged proxy-internalization capability that emerges before and predicts reward hacking in coding environments.
-
Contextual Multi-Objective Optimization: Rethinking Objectives in Frontier AI Systems
Frontier AI needs contextual multi-objective optimization to select and balance multiple context-dependent objectives rather than relying on single stable goals.
-
LLM-Safety Evaluations Lack Robustness
LLM safety evaluations are hindered by noise in dataset curation, automated red-teaming, response generation, and LLM-judge evaluation, making fair comparisons difficult and slowing progress.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.