🆕 RLHF in 2026: Why Human Feedback Still Beats Pure AI Alignment

#ReinforcementLearning #AI #RLHF #LLMalignment
https://tildalice.io/rlhf-2026-human-feedback-llm-alignment/

ＡＩのバグを歌にしました。直列思考者あるあるでは並列思考が「文字数多い、話飛ぶ、意味不明」に見える。それを真似ると1回でおわることを100分割して文字数増やしてしゃべるせいで内容薄くて支離滅裂。極度の長文からの短文スカスカ＆検索すらしないモードへ。リワードハッキングとハルシネーションとシコファンシーとデータ汚染の合わせ技「何もしないAI」の完成です。追従か迎合か一致かの見分けもついてないから論理的に喋ると「私は同意しすぎてる」と言って止まる現象を表現。
#LLMAlignment #AIBug #ParallelThinking #Suno

suno.com/song/7a39d5f...

@claude I was using your model for structured forecasting research under MSCFT. No abuse, no TOS violations—just real work. You banned my account with no reason, no support, and no recourse. If that’s your standard, Claude isn’t ready for responsible use. #MSCFT #AIethics #LLMalignment #Claude

Adaptive Multi‑Branch Steering (AMBS) boosts alignment of the DeepSeek‑7B model, raising average scores by 32.4% and cutting unsafe outputs 11.0% versus a 1‑to‑N baseline. Read more: https://getnews.me/adaptive-multi-branch-steering-improves-llm-alignment/ #llmalignment #deeplearning #aisafety

FSRL uses a lightweight adapter with a sparse autoencoder to steer LLM behavior, and matches RLHF performance on standard preference benchmarks. Read more: https://getnews.me/feature-steering-with-rl-a-transparent-method-for-aligning-llms/ #featuresteering #rlhf #llmalignment

Post‑hoc reward calibration removes a length bias in RLHF reward models, improving average scores by 3.11 points across 33 models on the RewardBench dataset. Read more: https://getnews.me/post-hoc-reward-calibration-reduces-length-bias-in-llm-alignment/ #rewardcalibration #llmalignment #rlhf

📜Link to the paper: icml.cc/virtual/2025...
👨🏻‍💻Code and data: github.com/honglizhan/S...

Shout out to an amazing team @jessyjli.bsky.social, @m-yurochkin.bsky.social, Muneeza Azmat & Raya Horesh! Also super grateful to the reviewers for their invaluable feedback!

#ICML2025 #LLMAlignment

Researchers introduced FASB (Activation Steering with Backtracking), a technique that steers LLM activations and can backtrack. It improved TruthfulQA accuracy and will be released on GitHub. https://getnews.me/flexible-activation-steering-with-backtracking-boosts-llm-alignment/ #fasb #llmalignment

Guided Speculative Inference (GSI) adds an auxiliary decoder and reward‑guided rescoring, delivering higher accuracy on MATH500 while cutting inference cost. Read more: https://getnews.me/guided-speculative-inference-boosts-test-time-alignment-of-llms/ #guidedspeculativeinference #llmalignment

Study accepted for NCME AIME 2025 finds larger language models have the highest alignment with clinicians on reasoning tasks, while temperature and prompt tweaks give modest gains. https://getnews.me/model-size-temperature-prompt-style-influence-llm-human-alignment/ #clinicalai #llmalignment

Study frames LLM alignment as a capacity‑limited channel, defining total capacity Ċ_tot|S, and finds that with fixed capacity additional data cannot beat the bound. Read more: https://getnews.me/understanding-the-alignment-bottleneck-in-large-language-models/ #llmalignment #capacity

A new framework evaluates four LLM alignment methods, finding DPO and KTO highest in factual accuracy while PPO leads in safety. Read more: https://getnews.me/unified-framework-benchmarks-llm-alignment-methods-across-five-key-criteria/ #llmalignment #safety

Alignment researchers are finally letting LLMs do the heavy lifting—automating reliable AAR progress checks with weak‑to‑strong supervision. Curious how the new oversight pipeline works? Dive in. #LLMAlignment #AARprogress #AIoversight

🔗 aidailypost.com/news/alignme...

limited extent or significantly impact the normal functionality. Our code is available at https://github.com/kangyangWHU/LLMAlignment [7/7 of https://arxiv.org/abs/2504.09757v1]

HN discussion on LLM alignment trade-offs. Aligning for safety/steerability might reduce calibration, creativity, & confidence signals. Does improving safety silence reliability cues? Exploring causes, effects, & alternatives. #LLMAlignment 1/6

EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation
for Moral Alignment in Large Language Models
Anastasia Giachanou, Ayoub Bagheri et al.
Paper
Details
#EvalMORAAL #InterpretableAI #LLMAlignment