Papersby GlyphWhat it shows
Research examines reward hacking in rubric-based reinforcement learning, identifying verifier failure and rubric-design limitations as key sources of divergence between training and evaluation metrics.
By Anas Mahmoud, MohammadHossein Rezaei, Zihao Wang and 3 more · arXiv 2605.12474 · PDF
Hugging Face's summary; not yet checked by hand.
More from Scale AI
All 25Root-Cause Attribution Is a Search Problem: Continual Search for Long-Horizon Agent FailuresContinual Search improves automated root-cause attribution in long agent execution traces by iteratively prompting diagnosis until unresolved evidence is found.
Scale AI5 upvotesRead →SteerDuplex: Steerable Duplex Speech Dialogue ModelsA full-duplex speech model that follows spoken instructions on tone, persona, pace and voice, with a benchmark of 390 prompts to test it.Scale AI3 citations6 upvotesRead →Studying Without a Syllabus: Task-Agnostic Environment PreprocessingAn agent can explore unfamiliar environments without task-specific guidance to build reusable artifacts that reduce later inference costs, though larger study budgets do not always improve results.Scale AI1 citations4 upvotesRead →HarnessOpt-Bench: Evaluating LLMs at Harness OptimizationHarnessOpt-Bench measures how well frontier LLMs iteratively improve agent harnesses under constrained evaluation budgets, revealing substantial variation across models and tasks.Scale AI12 citations35 upvotesRead →Model or Harness? An Interaction-Centric Taxonomy for Localizing Agent FailuresAn interaction-centric taxonomy localizes agent failures to specific component interactions to guide targeted repairs across diverse architectures.Scale AI11 citations10 upvotesRead →SWE-INTERACT: Reimagining SWE Benchmarks as User-Driven Long-Horizon Coding SessionsSWE-Interact presents a testbed that evaluates coding agents in realistic multi-turn, user-driven software engineering scenarios, revealing significant gaps between single-turn performance and interactive task completion.Scale AI23 citations7 upvotesRead →Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVRPOW3R is a policy-aware framework for reinforcement learning with rubric-based rewards that adapts criterion weights during training to improve policy optimization while preserving human-defined criteria importance.Scale AI13 citations5 upvotesRead →HiL-Bench (Human-in-Loop Benchmark): Do Agents Know When to Ask for Help?Frontier AI agents struggle with judgment calls about when to seek help, leading to poor performance on incomplete or ambiguous tasks despite having sufficient capabilities.Scale AI28 citations4 upvotesRead →
Other reasoning papers
TopicChain-of-Thought Prompting Elicits Reasoning in Large Language ModelsAsking a model to write out its intermediate steps (chain of thought) sharply improves its math and logic answers.
Google23.8k citations16 upvotesRead →DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language ModelsDeepSeekMath 7B improves mathematical reasoning through enhanced data pre-training and Group Relative Policy Optimization, achieving high scores on MATH benchmark without external tools.
DeepSeek10.6k citations152 upvotesRead →DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement LearningReinforcement learning taught a model long step-by-step reasoning on par with OpenAI o1.DeepSeek5,972 citations458 upvotesRead →DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code IntelligenceDeepSeek-Coder-V2, a Mixture-of-Experts language model, excels in code-specific tasks by enhancing coding and mathematical reasoning capabilities while expanding language support and context length.DeepSeek530 citations73 upvotesRead →MolmoAct: Action Reasoning Models that can Reason in SpaceAction Reasoning Models (ARMs) integrate perception, planning, and control to enable adaptable and explainable robotic behavior, achieving superior performance across various tasks and settings.
Ai2232 citations45 upvotesRead →ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent PlanningThinkAct, a dual-system framework, uses reinforced visual latent planning to enable few-shot adaptation, long-horizon planning, and self-correction in embodied AI tasks by bridging high-level reasoning with low-level action execution.
NVIDIA192 citations43 upvotesRead →
About this paper
- Venue
- arXiv.org
- Citations
- 33, 6 influential · Semantic Scholar
- Upvotes
- 4 · Hugging Face
- Lab
- Scale AI · on Companies · on Acquisitions · on awards · on breaches · on Attested · on Rounds · on Downtime · on labs · on Leaders · on marketplaces · on Allied · on Paydays · on quotes · on salaries · on Nines · on socials · on startups · on terms · on Titlewave
Changes
Citations rose from 31 to 33.
Influential citations rose from 4 to 6.
Citations rose from 24 to 31.
Influential citations rose from 2 to 4.
Citations rose from 20 to 24.
Influential citations rose from 1 to 2.
Citations rose from 15 to 20.
A first count of 1 influential citations.
A first count of 15 citations.
Added to the list.