Papersby GlyphHard2Verify: A Step-Level Verification Benchmark for Open-Ended Frontier Math research paper by Salesforce, 2025
Salesforce · Agents and evaluation · Oct 15, 2025 · 23 citations · 6 upvotes
What it shows
Hard2Verify, a human-annotated benchmark, evaluates step-level verifiers for LLM-based mathematical reasoning systems, highlighting the challenges and performance gaps between open-source and closed-source models.
By Shrey Pandit, Austin Xu, Xuan-Phi Nguyen and 3 more · arXiv 2510.13744 · PDF · Code
Hugging Face's summary; not yet checked by hand.
More from Salesforce
All 34Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior ForesightSelf-Retrospection Distillation improves agent performance by using hindsight from completed trajectories to supervise foresight predictions during training.
Salesforce1 citations197 upvotesRead →Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM AgentsKeeps an agent's raw past runs and writes a task-specific memory only when a new task arrives, rather than deciding up front what to keep.Salesforce2 citations56 upvotesRead →Flattening Every Memory Peak in Long-Context Mixture-of-Experts TrainingBounds the four memory peaks that crash long-context mixture-of-experts training, from expert dispatch to optimizer state, with fixed-size GPU schedules.Salesforce19 upvotesRead →RISE: Recursive Improvement via Self-Extrapolating Policy DistillationRISE improves language model post-training by recursively generating dense token-level supervision from the model's own reinforcement learning trajectory via self-extrapolation, avoiding external teachers.Salesforce5 citations17 upvotesRead →EVOHARNESSBENCH: Can Your Agents Keep Pace with an Evolving Harness?The study introduces a benchmark to evaluate LLM agents under evolving tool, skill, and agent harnesses, revealing persistent gaps in retention, adaptation, and harness-induced forgetting.Salesforce7 citations32 upvotesRead →Random Attention: Rethinking KV Cache Eviction for Efficient ReasoningRandom eviction of reasoning tokens matches selective KV cache compression because reasoning traces are self-protecting through redundancy, making scoring unnecessary once prompts are preserved.Salesforce5 citations187 upvotesRead →DarwinX: Evolving Agent Harnesses Through Natural SelectionDarwinX evolves agent harnesses via population selection with frozen models, improving verified performance across benchmarks without benchmark-specific patches.Salesforce16 citations116 upvotesRead →StateAct: Program State, before Pixels, for Long-Horizon Computer-Use AgentsStateAct improves computer-use agents by grounding actions, verification, and memory in direct program state rather than screenshots, boosting success rates while reducing cost.Salesforce6 citations62 upvotesRead →
Other agents and evaluation papers
TopicGemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic CapabilitiesGemini 2.X model family, including Gemini 2.5 Pro and Flash, offers superior coding, reasoning, and multimodal understanding capabilities across a range of computational efficiencies.
Google4,594 citations67 upvotesRead →DeepSeek-V3.2: Pushing the Frontier of Open Large Language ModelsDeepSeek-V3.2 introduces DeepSeek Sparse Attention and a scalable reinforcement learning framework, achieving superior reasoning and performance compared to GPT-5 and Gemini-3.0-Pro in complex reasoning tasks.
DeepSeek887 citations275 upvotesRead →WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks?WorkArena and BrowserGym evaluate large language model-based agents' ability to perform enterprise software tasks, revealing gaps in current agent capabilities and differences between open and closed-source LLMs.
ServiceNow443 citations2 upvotesRead →Aya Model: An Instruction Finetuned Open-Access Multilingual Language ModelAya, a multilingual generative language model supporting over 50% lower-resourced languages, excels in both generative and discriminative tasks across 99 languages, and provides extensive evaluations and open-source resources.
Cohere438 citations48 upvotesRead →Rubrics as Rewards: Reinforcement Learning Beyond Verifiable DomainsRubrics as Rewards (RaR) framework uses structured rubrics as interpretable reward signals for on-policy training, improving performance in real-world reinforcement learning tasks with subjective criteria.
Scale AI431 citations6 upvotesRead →SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?SWE-Bench Pro is a challenging benchmark for coding models, featuring complex, enterprise-level problems that require substantial code modifications, with performance evaluations showing significant limitations in current models.Scale AI367 citations21 upvotesRead →
About this paper
- Venue
- Annual Meeting of the Association for Computational Linguistics
- Citations
- 23, 3 influential · Semantic Scholar
- Upvotes
- 6 · Hugging Face
- Lab
- Salesforce · on Companies · on Acquisitions · on Quarterly · on Paydays · on TechConf
Changes
Citations rose from 22 to 23.
Influential citations rose from 2 to 3.
Citations rose from 20 to 22.
Citations rose from 15 to 20.
A first count of 2 influential citations.
A first count of 15 citations.
Added to the list.