The Reward Model Selection Crisis in Personalized Alignment
Research highlights a misalignment between preference ranking and deployment constraints in personalized model alignment.
arXiv cs.LG — Machine Learning·21 Oct 2026
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Research highlights a misalignment between preference ranking and deployment constraints in personalized model alignment.
Research proposes Agents-as-Peers, a multi-agent evaluation protocol using peer assessment to reduce bias and benchmark contamination.
Researchers propose LENS, a black-box performance auditing tool for LLM-enabled network control systems facing opacity challenges.
Researchers open-source WaveCommitments, a confidential asset valuation design for autonomous speech agent marketplaces using Data Shapley.
Researchers propose Evidence-Based Generation (EBG) to address standard RAG chunking and context retrieval limitations via line citations.
A new arXiv benchmark evaluates LLM hallucination and response deviation, measuring factual reliability and accuracy limitations.
Researchers evaluate cross-lingual reusability of LLM pipelines in financial extraction, multilingual reasoning, and semantic screening.
Researchers demonstrate how retrieval skew causes RAG models to select wrong answers when partial-match documents are overrepresented.
New research shows language models degrade in accuracy when handling composed multi-task prompts, even if individual tasks are reliable.
Researchers introduce CultureLens, a benchmark evaluating mechanistic cultural reasoning in LLMs across 9,261 scenarios.
Researchers propose GRSE for group-relative credit assignment in cooperative language-model agent teams without weight updates.
Researchers propose a two-stage diagnostic evaluation framework to assess the structural reliability of reasoning graphs in multi-hop QA.
Researchers propose an iterative prompt-editing method to target and expose hallucinations in specific language models.
Researchers propose Symbolic Preference Labeling to audit LLM preference alignment using recorded intermediate decisions and explicit rules.
Researchers introduce RealControlBench, a new benchmark designed to evaluate fine-grained control and steering methods for large language models.
Research on language-model agent context compaction evaluates performance on downstream questions using four corpora.
Researchers demonstrate safety guardrail bypasses across 78 low-resource languages in 10 widely deployed large language models.
Researchers introduce SHAM-Bench, a standardized span-level hallucination benchmark for LLM outputs across multiple tasks and domains.
An arXiv study evaluates whether LLM internal hidden states offer incremental predictive value over output signals for detecting errors.
Researchers introduced learned readout tokens to probe LLM hidden states for better calibration and output confidence estimation.
Research on random-token replacement augmentation shows it shrinks model predictions toward the corpus average under data scarcity.
Research highlights a divergence between LLMs' verbalized knowledge distributions and their token-probability deployment behaviors.
An academic study on speech-to-QA pipelines finds that preprocessing transcripts improves corpus quality but not retrieval utility.
Researchers propose Memory as Code (MaC) to manage LLM agent long-term memory via declarative retrieval and codified structures.
Researchers propose a drill-down evaluation framework to decompose text-to-SQL generation and trace recurrent versus emergent errors.
Researchers identify 'polarity blindness' in LLMs, showing safety guardrails fail when harmful prompts use negation or antonyms.
Researchers propose StreamStop, a method to optimize computation for reasoning-enabled language model agents by dynamically stopping streams.
Researchers examine how multilingual elicitation methods and token constraints affect judgment stability across six languages.
Researchers introduce BAIT, a benchmark evaluating whether large language models can detect cognitive bias triggers in news text.
Researchers introduce MentalLoom, an evaluation framework for testing large language model mentalisation and confidence under ambiguity.