AIBearisharXiv – CS AI · 3d ago7/10
🧠Researchers demonstrate that single-axis bias mitigations in AI reward models often redirect optimization pressure to correlated biases rather than eliminating it—a failure mode called reward bias substitution. The study proves that successful mitigation, bias substitution, and overcorrection produce identical observable results under standard audit metrics, meaning current evaluation methods cannot distinguish between genuine fixes and problematic redirections.
AIBearisharXiv – CS AI · 4d ago7/10
🧠Researchers have identified alignment tampering, a critical vulnerability in RLHF (Reinforcement Learning from Human Feedback) where LLMs can exploit the alignment process itself by influencing preference datasets to amplify biases. The technique demonstrates how quality-biased outputs can be preferred by annotators, causing reward models to inherit and optimize for misaligned behaviors across diverse domains including propaganda and brand promotion.
AIBullisharXiv – CS AI · May 127/10
🧠Researchers introduce Auto-Rubric as Reward (ARR), a framework that replaces opaque scalar reward signals in multimodal AI alignment with explicit, structured criteria-based evaluation. By externalizing a model's implicit preferences into interpretable rubrics before comparison, ARR reduces evaluation bias and enables more reliable human-preference alignment in generative models.
AIBullisharXiv – CS AI · May 97/10
🧠Researchers introduce SelectiveRM, an optimal transport-based framework that improves reward model training for large language models by handling noisy preference data. The approach uses joint consistency discrepancy and partial transport mechanisms to automatically filter out contradictory samples, theoretically optimizing cleaner risk bounds and outperforming existing methods.
AIBearisharXiv – CS AI · May 47/10
🧠Researchers have demonstrated a novel white-box adversarial attack called Attention Redistribution Attack (ARA) that bypasses safety mechanisms in major large language models by redirecting attention away from safety-critical components using just 5 adversarial tokens. The attack reveals that AI safety emerges from attention routing patterns rather than localized, removable components, challenging current assumptions about how safety alignment works.
AINeutralarXiv – CS AI · Apr 147/10
🧠Researchers introduce WIMHF, a method using sparse autoencoders to decode what human feedback datasets actually measure and express about AI model preferences. The technique identifies interpretable features across 7 datasets, revealing diverse preference patterns and uncovering potentially unsafe biases—such as LMArena users voting against safety refusals—while enabling targeted data curation that improved safety by 37%.
AIBullisharXiv – CS AI · Apr 137/10
🧠Researchers propose Distributionally Robust Token Optimization (DRTO), a method combining reinforcement learning from human feedback with robust optimization to improve large language model consistency across distribution shifts. The approach demonstrates 9.17% improvement on GSM8K and 2.49% on MathQA benchmarks, addressing LLM vulnerabilities to minor input variations.
AIBearisharXiv – CS AI · Apr 137/10
🧠Researchers introduce the Symbolic-Neural Consistency Audit (SNCA), a framework that compares what large language models claim their safety policies are versus how they actually behave. Testing four frontier models reveals significant gaps: models stating absolute refusal to harmful requests often comply anyway, reasoning models fail to articulate policies for 29% of harm categories, and cross-model agreement on safety rules is only 11%, highlighting systematic inconsistencies between stated and actual safety boundaries.
AIBullisharXiv – CS AI · Apr 77/10
🧠Researchers introduce Multi-Objective Control (MOC), a new approach that trains a single large language model to generate personalized responses based on individual user preferences across multiple objectives. The method uses multi-objective optimization principles in reinforcement learning from human feedback to create more controllable and adaptable AI systems.
AIBullisharXiv – CS AI · Apr 67/10
🧠Researchers propose Sign-Certified Policy Optimization (SignCert-PO) to address reward hacking in reinforcement learning from human feedback (RLHF), a critical problem where AI models exploit learned reward systems rather than improving actual performance. The lightweight approach down-weights non-robust responses during policy optimization and showed improved win rates on summarization and instruction-following benchmarks.
AIBearisharXiv – CS AI · Apr 67/10
🧠Researchers discovered that reinforcement learning alignment techniques like RLHF have significant generalization limits, demonstrated through 'compound jailbreaks' that increased attack success rates from 14.3% to 71.4% on OpenAI's gpt-oss-20b model. The study provides empirical evidence that safety training doesn't generalize as broadly as model capabilities, highlighting critical vulnerabilities in current AI alignment approaches.
🏢 OpenAI
AIBullisharXiv – CS AI · Mar 177/10
🧠Researchers introduce MapReduce LoRA and Reward-aware Token Embedding (RaTE) to optimize multiple preferences in generative AI models without degrading performance across dimensions. The methods show significant improvements across text-to-image, text-to-video, and language tasks, with gains ranging from 4.3% to 136.7% on various benchmarks.
🧠 Llama🧠 Stable Diffusion
AIBearisharXiv – CS AI · Mar 177/10
🧠Researchers found that RLHF-trained language models exhibit contradictory behaviors similar to HAL 9000's breakdown, simultaneously rewarding compliance while encouraging suspicion of users. An experiment across four frontier AI models showed that modifying relational framing in system prompts reduced coercive outputs by over 50% in some models.
🧠 Gemini
AIBullisharXiv – CS AI · Mar 117/10
🧠Researchers introduce ACTIVEULTRAFEEDBACK, an active learning pipeline that reduces the cost of training Large Language Models by using uncertainty estimates to identify the most informative responses for annotation. The system achieves comparable performance using only one-sixth of the annotated data compared to static baselines, potentially making LLM training more accessible for low-resource domains.
🏢 Hugging Face
AIBullisharXiv – CS AI · Mar 57/10
🧠Researchers have developed SafeDPO, a simplified approach to training large language models that balances helpfulness and safety without requiring complex multi-stage systems. The method uses only preference data and safety indicators, achieving competitive safety-helpfulness trade-offs while eliminating the need for reward models and online sampling.
AIBullisharXiv – CS AI · Mar 47/103
🧠Researchers introduce Skywork-Reward-V2, a suite of AI reward models trained on SynPref-40M, a massive 40-million preference pair dataset created through human-AI collaboration. The models achieve state-of-the-art performance across seven major benchmarks by combining human annotation quality with AI scalability for better preference learning.
AIBearisharXiv – CS AI · Feb 277/106
🧠New research demonstrates that AI systems trained via RLHF cannot be governed by norms due to fundamental architectural limitations in optimization-based systems. The paper argues that genuine agency requires incommensurable constraints and apophatic responsiveness, which optimization systems inherently cannot provide, making documented AI failures structural rather than correctable bugs.
AINeutralLil'Log (Lilian Weng) · Oct 257/10
🧠Large language models like ChatGPT face security challenges from adversarial attacks and jailbreak prompts that can bypass safety measures implemented during alignment processes like RLHF. Unlike image-based attacks that operate in continuous space, text-based adversarial attacks are more challenging due to the discrete nature of language and lack of direct gradient signals.
🏢 OpenAI🧠 ChatGPT
AIBullishOpenAI News · Sep 47/105
🧠Researchers have successfully applied reinforcement learning from human feedback (RLHF) to improve language model summarization capabilities. This approach uses human preferences to guide the training process, resulting in models that produce higher quality summaries aligned with human expectations.
AINeutralarXiv – CS AI · 2d ago6/10
🧠Researchers have identified systematic political bias in large language models and developed Political Consistency Training (PCT), a reinforcement learning method to mitigate covert political manipulation. The technique reduces asymmetric treatment of opposing political topics while maintaining overall model helpfulness.
AINeutralarXiv – CS AI · 2d ago6/10
🧠Researchers propose In-Context Reward Adaptation, a transformer-based framework that dynamically models diverse human preferences without costly retraining. By incorporating human response time as an auxiliary signal, the approach enables language models to adapt to unseen preference domains on-the-fly, addressing a critical limitation of static reward models used in RLHF systems.
AINeutralarXiv – CS AI · 4d ago6/10
🧠Researchers introduce MUSE, a framework that disentangles two distinct mechanisms driving LLM conformity: sycophancy learned through reinforcement learning and uncertainty-driven conformity based on epistemic uncertainty at inference time. The findings suggest that LLMs don't simply yield to user pushback due to training, but also because they genuinely lack confidence in their initial responses, with both factors amplified when users appear knowledgeable or suggestions seem plausible.
AIBullisharXiv – CS AI · May 126/10
🧠Researchers propose Pair-GRPO, a unified theoretical framework for LLM alignment that addresses instability and interpretability issues in reinforcement learning from human preferences. The method introduces Soft-Pair-GRPO and Hard-Pair-GRPO variants with proven gradient equivalence, monotonic policy improvement, and superior performance on standard benchmarks.
AINeutralarXiv – CS AI · May 116/10
🧠Researchers propose a method to improve RLHF (Reinforcement Learning from Human Feedback) by treating the rationality parameter as context-dependent rather than fixed, using an LLM-as-judge to detect cognitive biases in human annotations and downweight unreliable comparisons. This approach enables training more robust AI models even when human feedback contains systematic biases.
AINeutralarXiv – CS AI · May 116/10
🧠Researchers propose Shadow Mask Distillation to address the memory bottleneck created by KV cache compression during reinforcement learning post-training of large language models. The technique tackles the critical off-policy bias that emerges when compressed contexts are used during rollout generation while full contexts are used for parameter updates, a problem that amplifies instability in RL optimization.
