A focused feed for RLHF, preference data, rater protocols, agent evaluation, and LLM-as-judge research.
Every paper includes structured metadata for quick triage.
Across multiple steering benchmarks, we show that SKOP achieves the best joint steering-utility trade-off, reducing utility degradation by 5-7x while retaining over 95% of vanilla steering efficacy.
To overcome these limitations, we present MANTRA, a framework for automatically synthesizing machine-checkable compliance benchmarks from natural-language manuals and tool schemas.
Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation.
Paradoxically, we observe that tool-enabled evaluation can degrade reasoning performance even when the strong thinking models make almost no actual tool calls.
In contrast, fully unstructured teams enable adaptability and exploration but suffer from inefficiencies such as error propagation, inter-agent conflicts, and wasted resources (measured in time, tokens, or file operations).
Human label variation has been established as a central phenomenon in NLP: the perspectives different annotators have on the same item need to be embraced.
Laughter is a social non-vocalization that is universal across cultures and languages, and is crucial for human communication, including social bonding and communication signaling.
Agentic RAG extends this paradigm by replacing single-step retrieval with a multi-step process, in which the large language model (LLM) acts as a search agent that generates intermediate thoughts and subqueries to iteratively interact with…
In this paper, we introduce a new multi-genre benchmark (more than 1000 samples) for semantic segmentation in conversational Arabic, focusing on dialectal discourse.
Across three model families, six scales, and six math reasoning benchmarks, ReasonMaxxer matches or exceeds full RL performance while requiring only tens of problems and minutes of single-GPU training, a reduction in training cost of…
Browse by Topic
Jump directly into tag and hub pages to crawl deeper content clusters.
Although large language models (LLMs) have significantly advanced text generation, stories generated by LLMs still diverge from human-authored works regarding complex narrative structure and human-aligned preferences.
We find existing reward models struggle to select human-preferred stories, with the best model achieving only 66.3\% accuracy.
We introduce IslamicMMLU, a benchmark of 10,013 multiple-choice questions spanning three tracks: Quran (2,013 questions), Hadith (4,000 questions), and Fiqh (jurisprudence, 4,000 questions).
The benchmark is used to create the IslamicMMLU public leaderboard for evaluating LLMs, and we initially evaluate 26 LLMs, where their averaged accuracy across the three tracks varied between 39.8\% to 93.8\% (by Gemini 3 Flash).
Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models, current reward modeling heavily relies on experimental feedback data collected from human annotators under controlled and costly…
Extensive experiments across diverse LLM backbones and benchmark datasets validate that CausalRM effectively learns accurate reward signals from noisy and biased observational feedback and delivers substantial performance improvements on…
To address these challenges, we introduce SDiaReward, an end-to-end multi-turn reward model trained on SDiaReward-Dataset, a novel collection of episode-level preference pairs explicitly targeting these gaps.
In the random-error setting, models strongly prefer correct completions in paired evaluation: 83.1% accuracy at balanced data and 67.0% even when correct rules appear in only 10% of the corpus.
Replacing random errors with a coherent but mathematically incorrect rule system largely eliminates the preference (near-chance accuracy).
The models were developed through a four-stage training pipeline: continued pre-training on Portuguese and Brazilian legal corpora, long-context extension to 128K tokens, supervised fine-tuning on instruction data spanning chat, code, legal…
We evaluate the models on six benchmark categories: conversational capabilities in Brazilian Portuguese, knowledge of Brazilian legislation, long-context understanding, instruction following, standardized exams, and agentic capabilities…
On this unified benchmark, we evaluate four approaches: (i) encoder-based classification fine-tuning, (ii) zero- and few-shot prompting, (iii) instruction tuning and Retrieval-Augmented Generation (RAG), and (iv) Supervised Fine-Tuning…
While prior research emphasizes the role of on-policy data in mitigating forgetting, we uncover--and validate both theoretically and empirically--an overlooked yet critical mechanism: the implicit regularization inherent in Direct…
In practice, users iteratively refine visualizations through multi-turn interactions that require maintaining common ground, tracking prior edits, and adapting to evolving preferences.
We introduce ChartEditBench, a benchmark for incremental, visually grounded chart editing via code, comprising 5,000 difficulty-controlled modification chains and a rigorously human-verified subset.
Yet safety pipelines, benchmarks, and alignment still largely target English and a handful of high-resource languages, implicitly assuming safety and factuality ''transfer'' across languages.
We synthesize recent findings indicating that (i) safety guardrails weaken sharply on low-resource and code-mixed inputs, (ii) culturally harmful behavior can persist even when standard toxicity scores look acceptable, and (iii)…
We propose PrivAct, a contextual privacy-aware multi-agent learning framework that internalizes contextual privacy preservation directly into models' generation behavior for privacy-compliant agentic actions.
Experiments across multiple LLM backbones and benchmarks demonstrate consistent improvements in contextual privacy preservation, reducing leakage rates by up to 12.32% while maintaining comparable helpfulness, as well as zero-shot…
To address this, we propose Goal-Oriented Preference Optimization (GOPO), a hierarchical reinforcement learning framework that decouples strategy planning from response generation via an Expert Agent and a Customer Service Agent.
We evaluate GOPO on public benchmarks and e-commerce customer service datasets, and introduce Task-focused Sequential Engagement (TSE), a sequence-level metric derived from real e-commerce interaction data.
We introduce Vibe Code Bench, a benchmark of 100 web application specifications (50 public validation, 50 held-out test) with 964 browser-based workflows comprising 10,131 substeps, evaluated against deployed applications by an autonomous…
We identify self-testing during generation as a strong performance predictor (Pearson r=0.72), and show through a completed human alignment study that evaluator selection materially affects outcomes (31.8-93.6% pairwise step-level…
On code generation (LiveCodeBench, CodeContests, SWE-Bench) and math reasoning (AIME, HMMT) benchmarks, V_1-Infer improves Pass@1 by up to 10% over pointwise verification and outperforms recent test-time scaling methods while being…
In this work, we propose ProAgent, an end-to-end proactive agent system that harnesses on-demand sensory contexts to provide in-the-wild assistance.
Results demonstrate that ProAgent achieves up to 27.7% higher proactive prediction accuracy and 20.5% lower false detection than state-of-the-art baselines.
Finally, we demonstrate efficiency enhancements across three downstream tasks: preference data synthesis, Pareto frontier approximation and optimization, and distillation of aligned behaviors for intervention-free inference.
We evaluate MLLM verifiers across web navigation, computer use, and robotics, spanning 13+ models, 28+ designs, and thousands of trajectories from diverse agents.
Our methods yield more human-aligned verifiers, improving failure detection by 25pp and accuracy by 14pp.