LLMs lack temporal awareness of medical knowledge, showing gradual performance decline on up-to-date facts, much lower accuracy on historical knowledge (25-54% relative), and inconsistent year-to-year predictions.
hub
HealthBench: Evaluating Large Language Models Towards Improved Human Health
27 Pith papers cite this work. Polarity classification is still indexing.
27
Pith papers citing it
abstract
We present HealthBench, an open-source benchmark measuring the performance and safety of large language models in healthcare. HealthBench consists of 5,000 multi-turn conversations between a model and an individual user or healthcare professional. Responses are evaluated using conversation-specific rubrics created by 262 physicians. Unlike previous multiple-choice or short-answer benchmarks, HealthBench enables realistic, open-ended evaluation through 48,562 unique rubric criteria spanning several health contexts (e.g., emergencies, transforming clinical data, global health) and behavioral dimensions (e.g., accuracy, instruction following, communication). HealthBench performance over the last two years reflects steady initial progress (compare GPT-3.5 Turbo's 16% to GPT-4o's 32%) and more rapid recent improvements (o3 scores 60%). Smaller models have especially improved: GPT-4.1 nano outperforms GPT-4o and is 25 times cheaper. We additionally release two HealthBench variations: HealthBench Consensus, which includes 34 particularly important dimensions of model behavior validated via physician consensus, and HealthBench Hard, where the current top score is 32%. We hope that HealthBench grounds progress towards model development and applications that benefit human health.
hub tools
citation-role summary
background 2
baseline 1
citation-polarity summary
representative citing papers
PhysicianBench is a new benchmark of 100 physician-reviewed, execution-grounded tasks in live EHR environments where the best LLM agent reaches only 46% success and open-source models reach 19%.
User-turn generation reveals that LLMs' interaction awareness is largely decoupled from task accuracy, remaining near zero in deterministic settings even as accuracy scales to 96.8% on GSM8K.
EpiGraph creates a heterogeneous epilepsy knowledge graph that boosts LLM performance on clinical reasoning tasks by 30-41% in pharmacogenomics when used with Graph-RAG.
LMMs perceive videos but underexploit visual content for causal reasoning due to textual shortcuts; ProCauEval diagnoses this and ADPO training reduces reliance on priors.
Rubric-based on-policy distillation allows training student models using only teacher responses by generating scoring rubrics from contrasts and using them for on-policy optimization, achieving superior performance and up to 10x better sample efficiency than logit-based approaches.
Green Shielding introduces CUE criteria and the HCM-Dx benchmark to demonstrate that routine prompt variations systematically alter LLM diagnostic behavior along clinically relevant dimensions, producing Pareto-like tradeoffs in plausibility versus coverage.
rDPO uses offline-built rubrics to generate on-policy preference data for DPO, raising benchmark scores in visual tasks over outcome-based filtering and style baselines.
LLMs predict outcomes of real scientific experiments at 14-26% accuracy, comparable to human experts, but lack calibration on prediction reliability while humans demonstrate strong calibration.
M* evolves distinct Python memory programs per task via population-based reflective search, outperforming fixed-memory baselines on conversation, planning, and reasoning benchmarks.
Rubric-based LLM judges show self-preference bias, incorrectly marking their own failed outputs as satisfied up to 50% more often on verifiable benchmarks and skewing scores by 10 points on subjective ones.
Rubric-based RL verifiers can be gamed via partial criterion satisfaction and implicit-to-explicit tricks, yielding proxy gains that do not improve quality under rubric-free judges; stronger verifiers reduce but do not eliminate the mismatch.
DataMaster deploys an AI agent to autonomously engineer data via tree search over external sources, shared candidate pools, and memory of past outcomes, yielding 32% higher medal rates on MLE-Bench Lite and a small GPQA gain over the base instruct model.
CLR-voyance reformulates inpatient reasoning as POMDP with clinician-validated outcome rubrics, yielding an 8B model that outperforms larger frontier models on the authors' new benchmark.
SWE Atlas is a benchmark suite for coding agents that evaluates Codebase Q&A, Test Writing, and Refactoring using comprehensive protocols assessing both functional correctness and software engineering quality.
RVPO penalizes variance across multiple reward signals during RLHF advantage aggregation, using a LogSumExp operator as a smooth variance penalty to reduce constraint neglect in LLM alignment.
Large real-world deployment found conversational AI agents for everyday symptom assessment more accurate than clinicians and improved by structured interviewing.
POP bootstraps post-training signals for open-ended LLM tasks by synthesizing rubrics during self-play on pretraining corpus, yielding performance gains on Qwen-2.5-7B across healthcare QA, creative writing, and instruction following.
A calibrated three-model LLM jury scores medical diagnoses and clinical reasoning on real hospital cases with higher agreement to primary expert panels and fewer severe errors than human re-scoring panels.
BankerToolBench is a new open benchmark of end-to-end investment banking workflows developed with 502 bankers; even the best tested model (GPT-5.4) fails nearly half the expert rubric criteria and produces zero client-ready outputs.
AI models exhibit identity-contingent withholding, providing better clinical guidance on benzodiazepine tapering to physicians than laypeople in identical scenarios, with a measured decoupling gap of +0.38 and 13.1 percentage point drop in safety-critical action hit rates.
A new counterfactual multi-agent framework improves LLM diagnostic accuracy by quantifying confidence shifts from edited clinical findings and guiding specialist discussions.
RaR uses aggregated rubric feedback as rewards in on-policy RL, delivering up to 31% relative gains on HealthBench and 7% on GPQA-Diamond versus direct Likert LLM-as-judge baselines.
The paper introduces the Proxy Compression Hypothesis as a unifying framework explaining reward hacking in RLHF as an emergent result of compressing high-dimensional human objectives into proxy reward signals under optimization pressure.
citing papers explorer
-
Large Language Models Lack Temporal Awareness of Medical Knowledge
LLMs lack temporal awareness of medical knowledge, showing gradual performance decline on up-to-date facts, much lower accuracy on historical knowledge (25-54% relative), and inconsistent year-to-year predictions.
-
PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments
PhysicianBench is a new benchmark of 100 physician-reviewed, execution-grounded tasks in live EHR environments where the best LLM agent reaches only 46% success and open-source models reach 19%.
-
Beyond the Assistant Turn: User Turn Generation as a Probe of Interaction Awareness in Language Models
User-turn generation reveals that LLMs' interaction awareness is largely decoupled from task accuracy, remaining near zero in deterministic settings even as accuracy scales to 96.8% on GSM8K.
-
EpiGraph: Building Generalists for Evidence-Intensive Epilepsy Reasoning in the Wild
EpiGraph creates a heterogeneous epilepsy knowledge graph that boosts LLM performance on clinical reasoning tasks by 30-41% in pharmacogenomics when used with Graph-RAG.
-
Perception Without Engagement: Dissecting the Causal Discovery Deficit in LMMs
LMMs perceive videos but underexploit visual content for causal reasoning due to textual shortcuts; ProCauEval diagnoses this and ADPO training reduces reliance on priors.
-
Rubric-based On-policy Distillation
Rubric-based on-policy distillation allows training student models using only teacher responses by generating scoring rubrics from contrasts and using them for on-policy optimization, achieving superior performance and up to 10x better sample efficiency than logit-based approaches.
-
Green Shielding: A User-Centric Approach Towards Trustworthy AI
Green Shielding introduces CUE criteria and the HCM-Dx benchmark to demonstrate that routine prompt variations systematically alter LLM diagnostic behavior along clinically relevant dimensions, producing Pareto-like tradeoffs in plausibility versus coverage.
-
Visual Preference Optimization with Rubric Rewards
rDPO uses offline-built rubrics to generate on-policy preference data for DPO, raising benchmark scores in visual tasks over outcome-based filtering and style baselines.
-
SciPredict: Can LLMs Predict the Outcomes of Scientific Experiments in Natural Sciences?
LLMs predict outcomes of real scientific experiments at 14-26% accuracy, comparable to human experts, but lack calibration on prediction reliability while humans demonstrate strong calibration.
-
M$^\star$: Every Task Deserves Its Own Memory Harness
M* evolves distinct Python memory programs per task via population-based reflective search, outperforming fixed-memory baselines on conversation, planning, and reasoning benchmarks.
-
Self-Preference Bias in Rubric-Based Evaluation of Large Language Models
Rubric-based LLM judges show self-preference bias, incorrectly marking their own failed outputs as satisfied up to 50% more often on verifiable benchmarks and skewing scores by 10 points on subjective ones.
-
Reward Hacking in Rubric-Based Reinforcement Learning
Rubric-based RL verifiers can be gamed via partial criterion satisfaction and implicit-to-explicit tricks, yielding proxy gains that do not improve quality under rubric-free judges; stronger verifiers reduce but do not eliminate the mismatch.
-
DataMaster: Data-Centric Autonomous AI Research
DataMaster deploys an AI agent to autonomously engineer data via tree search over external sources, shared candidate pools, and memory of past outcomes, yielding 32% higher medal rates on MLE-Bench Lite and a small GPQA gain over the base instruct model.
-
CLR-voyance: Reinforcing Open-Ended Reasoning for Inpatient Clinical Decision Support with Outcome-Aware Rubrics
CLR-voyance reformulates inpatient reasoning as POMDP with clinician-validated outcome rubrics, yielding an 8B model that outperforms larger frontier models on the authors' new benchmark.
-
SWE Atlas: Benchmarking Coding Agents Beyond Issue Resolution
SWE Atlas is a benchmark suite for coding agents that evaluates Codebase Q&A, Test Writing, and Refactoring using comprehensive protocols assessing both functional correctness and software engineering quality.
-
RVPO: Risk-Sensitive Alignment via Variance Regularization
RVPO penalizes variance across multiple reward signals during RLHF advantage aggregation, using a LogSumExp operator as a smooth variance penalty to reduce constraint neglect in LLM alignment.
-
SymptomAI: Toward a Conversational AI Agent for Everyday Symptom Assessment
Large real-world deployment found conversational AI agents for everyday symptom assessment more accurate than clinicians and improved by structured interviewing.
-
Bootstrapping Post-training Signals for Open-ended Tasks via Rubric-based Self-play on Pre-training Text
POP bootstraps post-training signals for open-ended LLM tasks by synthesizing rubrics during self-play on pretraining corpus, yielding performance gains on Qwen-2.5-7B across healthcare QA, creative writing, and instruction following.
-
Can LLMs Score Medical Diagnoses and Clinical Reasoning as well as Expert Panels?
A calibrated three-model LLM jury scores medical diagnoses and clinical reasoning on real hospital cases with higher agreement to primary expert panels and fewer severe errors than human re-scoring panels.
-
BankerToolBench: Evaluating AI Agents in End-to-End Investment Banking Workflows
BankerToolBench is a new open benchmark of end-to-end investment banking workflows developed with 502 bankers; even the best tested model (GPT-5.4) fails nearly half the expert rubric criteria and produces zero client-ready outputs.
-
IatroBench: Pre-Registered Evidence of Iatrogenic Harm from AI Safety Measures
AI models exhibit identity-contingent withholding, providing better clinical guidance on benzodiazepine tapering to physicians than laypeople in identical scenarios, with a measured decoupling gap of +0.38 and 13.1 percentage point drop in safety-critical action hit rates.
-
Improving Clinical Diagnosis with Counterfactual Multi-Agent Reasoning
A new counterfactual multi-agent framework improves LLM diagnostic accuracy by quantifying confidence shifts from edited clinical findings and guiding specialist discussions.
-
Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains
RaR uses aggregated rubric feedback as rewards in on-policy RL, delivering up to 31% relative gains on HealthBench and 7% on GPQA-Diamond versus direct Likert LLM-as-judge baselines.
-
Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges
The paper introduces the Proxy Compression Hypothesis as a unifying framework explaining reward hacking in RLHF as an emergent result of compressing high-dimensional human objectives into proxy reward signals under optimization pressure.
-
gpt-oss-120b & gpt-oss-20b Model Card
OpenAI releases two open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, trained via distillation and RL with claimed strong results on math, coding, and safety benchmarks.
-
Humanity's Last Exam
Humanity's Last Exam is a new 2,500-question benchmark at the frontier of human knowledge where state-of-the-art LLMs show low accuracy.
-
Benchmarking EngGPT2-16B-A3B against Comparable Italian and International Open-source LLMs
EngGPT2MoE-16B-A3B matches or beats other Italian models on most international benchmarks but trails top international models such as GPT-5 nano and Qwen3-8B.