Derives an exact telescoping decomposition of the naive RLVR reward-design estimator into null, elicitation, and reward-design terms on a tabular-GRPO simulator, measures the components across prior strengths, and validates via pre-registered factorial experiments plus re-audits of prior papers.
super hub Canonical reference
Holistic Evaluation of Language Models
Canonical reference. 84% of citing Pith papers cite this work as background.
151
Pith papers citing it
Background
84% of classified citations
abstract
Language models (LMs) are becoming the foundation for almost all major language technologies, but their capabilities, limitations, and risks are not well understood. We present Holistic Evaluation of Language Models (HELM) to improve the transparency of language models. First, we taxonomize the vast space of potential scenarios (i.e. use cases) and metrics (i.e. desiderata) that are of interest for LMs. Then we select a broad subset based on coverage and feasibility, noting what's missing or underrepresented (e.g. question answering for neglected English dialects, metrics for trustworthiness). Second, we adopt a multi-metric approach: We measure 7 metrics (accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency) for each of 16 core scenarios when possible (87.5% of the time). This ensures metrics beyond accuracy don't fall to the wayside, and that trade-offs are clearly exposed. We also perform 7 targeted evaluations, based on 26 targeted scenarios, to analyze specific aspects (e.g. reasoning, disinformation). Third, we conduct a large-scale evaluation of 30 prominent language models (spanning open, limited-access, and closed models) on all 42 scenarios, 21 of which were not previously used in mainstream LM evaluation. Prior to HELM, models on average were evaluated on just 17.9% of the core HELM scenarios, with some prominent models not sharing a single scenario in common. We improve this to 96.0%: now all 30 models have been densely benchmarked on the same core scenarios and metrics under standardized conditions. Our evaluation surfaces 25 top-level findings. For full transparency, we release all raw model prompts and completions publicly for further analysis, as well as a general modular toolkit. We intend for HELM to be a living benchmark for the community, continuously updated with new scenarios, metrics, and models.
hub tools
citation-role summary
background 22
dataset 2
other 1
citation-polarity summary
claims ledger
- abstract Language models (LMs) are becoming the foundation for almost all major language technologies, but their capabilities, limitations, and risks are not well understood. We present Holistic Evaluation of Language Models (HELM) to improve the transparency of language models. First, we taxonomize the vast space of potential scenarios (i.e. use cases) and metrics (i.e. desiderata) that are of interest for LMs. Then we select a broad subset based on coverage and feasibility, noting what's missing or underrepresented (e.g. question answering for neglected English dialects, metrics for trustworthiness).
authors
co-cited works
representative citing papers
AI model builders mostly highlight unique benchmarks that act as flexible narrative tools for market positioning rather than standardized scientific measurements.
EnergyAgentBench is a new benchmark with 70 task variants that evaluates LLM agents on live energy data for datacenter siting, long-horizon optimization, and causal grid diagnosis.
MCP-Atlas is a new benchmark with 1000 tasks on production MCP servers that uses claim-level scoring to evaluate LLM agents on realistic multi-step tool-use competency.
Continuous auditing creates an unavoidable cover regime in which static auditors cannot simultaneously eliminate coverage and granularity failures, shown via new policies, strategies, and a reproducible simulator.
CLQT is a new closed-loop, cost-aware benchmark that diagnoses LLM trading agent capabilities through strategy-consistent metrics and hash-verifiable trails rather than outcome rankings.
BehaviorBench is a benchmark for foundation models on behavioral tasks that reveals fine-tuned behavioral models outperform general models on distributional alignment while general models lead on individual-level accuracy.
MAC-Bench is a new adversarial benchmark that converts legal texts into executable scenarios via the SERV pipeline to measure procedural compliance in multi-agent LLM systems using CSR and MG metrics.
Invariant Gradient Alignment uses Logical Isomer Sets and a Continuous Gradient Conflict Mask to tighten OOD generalization bounds and boost empirical performance over ERM in reasoning distillation.
Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.
RealClawBench turns 281 real OpenClaw sessions into reproducible tasks that preserve the original distribution and shows the best of 14 models solves only 65.8 percent.
OR-Space is a benchmark for LLM agents performing full-lifecycle optimization tasks across Build, Revise, and Explain modes in executable multi-artifact workspaces.
SiDP distributes model weights across a DP group with WaS and CaS modes to increase KV cache capacity by up to 1.8x and end-to-end throughput by up to 1.5x over vLLM on H20/H200/B200 GPUs for offline LLM inference.
Language models display brittle safety by failing to adapt when context flips reverse action safety, with standard guardrails blind to consequence-flip scenarios.
Introduces the Grounded Observer framework that applies robotics-inspired formal constructs for runtime constraint enforcement on foundation model interaction trajectories in socially sensitive domains.
GRASP aggregates stable local LLM interaction judgments into global argument rankings via a convergent attack-defense propagation operator on interaction graphs, yielding higher reproducibility than holistic judging and no correlation with human convincingness.
SpikeProphecy decomposes spike-count forecasting performance into temporal fidelity, spatial pattern accuracy, and magnitude-invariant alignment, revealing reproducible brain-region predictability rankings and a sub-Poisson evaluation floor across seven model families on 105 Neuropixels sessions.
Develops a causal framework unifying generative AI fairness with standard ML, with new decompositions, identification conditions, and estimators demonstrated on LLM race and gender bias.
Hebatron is the first open-weight Hebrew MoE LLM adapted from Nemotron-3, reaching 73.8% on Hebrew reasoning benchmarks while activating only 3B parameters per pass and supporting 65k-token context.
LLMs routinely produce unsupported causal stories for personal sensing anomalies, and richer evidence or constrained prompts do not reliably eliminate this epistemic overreach.
Medical VLMs frequently select negated options that contradict visible chest X-ray findings, achieving only ~30% accuracy on direct presence probes, but a post-hoc consistency verifier raises accuracy above 95%.
iTRIALSPACE generates realistic virtual lesion trials on lung CTs that isolate performance drivers and show strong transfer of model rankings to real clinical data (ρ=0.93).
LLMSpace is the first framework to jointly model operational and embodied carbon for LLM inference on LEO satellites, incorporating radiation-hardened hardware, peripheral systems, and workload patterns such as prefill-decode behavior.
Coral cuts multi-LLM serving costs by up to 2.79x and raises goodput by up to 2.39x on heterogeneous GPUs through adaptive joint optimization and a lossless two-stage decomposition that solves quickly.
citing papers explorer
-
A Pre-Registered Causal Partition of Self-Consistency Elicitation and Reward Design in RLVR
Derives an exact telescoping decomposition of the naive RLVR reward-design estimator into null, elicitation, and reward-design terms on a tabular-GRPO simulator, measures the components across prior strengths, and validates via pre-registered factorial experiments plus re-audits of prior papers.
-
Unsteady Metrics and Benchmarking Cultures of AI Model Builders
AI model builders mostly highlight unique benchmarks that act as flexible narrative tools for market positioning rather than standardized scientific measurements.
-
EnergyAgentBench: Benchmarking LLM Agents on Live Energy Infrastructure Data
EnergyAgentBench is a new benchmark with 70 task variants that evaluates LLM agents on live energy data for datacenter siting, long-horizon optimization, and causal grid diagnosis.
-
MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers
MCP-Atlas is a new benchmark with 1000 tasks on production MCP servers that uses claim-level scoring to evaluate LLM agents on realistic multi-step tool-use competency.
-
A Benchmark for Strategic Auditee Gaming Under Continuous Compliance Monitoring
Continuous auditing creates an unavoidable cover regime in which static auditors cannot simultaneously eliminate coverage and granularity failures, shown via new policies, strategies, and a reproducible simulator.
-
CLQT: A Closed-Loop, Cost-Aware, Strategy-Consistent Benchmark for Diagnostic Evaluation of LLM Portfolio-Management Agents
CLQT is a new closed-loop, cost-aware benchmark that diagnoses LLM trading agent capabilities through strategy-consistent metrics and hash-verifiable trails rather than outcome rankings.
-
BehaviorBench: Benchmarking Foundation Models for Behavioral Science Tasks
BehaviorBench is a benchmark for foundation models on behavioral tasks that reveals fine-tuned behavioral models outperform general models on distributional alignment while general models lead on individual-level accuracy.
-
Beyond Goodhart's Law: A Dynamic Benchmark for Evaluating Compliance in Multi-Agent Systems
MAC-Bench is a new adversarial benchmark that converts legal texts into executable scenarios via the SERV pipeline to measure procedural compliance in multi-agent LLM systems using CSR and MG metrics.
-
Invariant Gradient Alignment for Robust Reasoning Distillation
Invariant Gradient Alignment uses Logical Isomer Sets and a Continuous Gradient Conflict Mask to tighten OOD generalization bounds and boost empirical performance over ERM in reasoning distillation.
-
Toward Calibrated, Fair, and accurate Deepfake Detection
Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.
-
RealClawBench: Live OpenClaw Benchmarks from Real Developer-Agent Sessions
RealClawBench turns 281 real OpenClaw sessions into reproducible tasks that preserve the original distribution and shows the best of 14 models solves only 65.8 percent.
-
OR-Space: A Full-Lifecycle Workspace Benchmark for Industrial Optimization Agents
OR-Space is a benchmark for LLM agents performing full-lifecycle optimization tasks across Build, Revise, and Explain modes in executable multi-artifact workspaces.
-
SiDP: Memory-Efficient Data Parallelism for Offline LLM Inference
SiDP distributes model weights across a DP group with WaS and CaS modes to increase KV cache capacity by up to 1.8x and end-to-end throughput by up to 1.5x over vLLM on H20/H200/B200 GPUs for offline LLM inference.
-
When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models
Language models display brittle safety by failing to adapt when context flips reverse action safety, with standard guardrails blind to consequence-flip scenarios.
-
Robotics-Inspired Guardrails for Foundation Models in Socially Sensitive Domains
Introduces the Grounded Observer framework that applies robotics-inspired formal constructs for runtime constraint enforcement on foundation model interaction trajectories in socially sensitive domains.
-
GRASP: Deterministic argument ranking in interaction graphs
GRASP aggregates stable local LLM interaction judgments into global argument rankings via a convergent attack-defense propagation operator on interaction graphs, yielding higher reproducibility than holistic judging and no correlation with human convincingness.
-
SpikeProphecy: A Large-Scale Benchmark for Autoregressive Neural Population Forecasting
SpikeProphecy decomposes spike-count forecasting performance into temporal fidelity, spatial pattern accuracy, and magnitude-invariant alignment, revealing reproducible brain-region predictability rankings and a sub-Poisson evaluation floor across seven model families on 105 Neuropixels sessions.
-
Causal Bias Detection in Generative Artificial Intelligence
Develops a causal framework unifying generative AI fairness with standard ML, with new decompositions, identification conditions, and estimators demonstrated on LLM race and gender bias.
-
HEBATRON: A Hebrew-Specialized Open-Weight Mixture-of-Experts Language Model
Hebatron is the first open-weight Hebrew MoE LLM adapted from Nemotron-3, reaching 73.8% on Hebrew reasoning benchmarks while activating only 3B parameters per pass and supporting 65k-token context.
-
Causal Stories from Sensor Traces: Auditing Epistemic Overreach in LLM-Generated Personal Sensing Explanations
LLMs routinely produce unsupported causal stories for personal sensing anomalies, and richer evidence or constrained prompts do not reliably eliminate this epistemic overreach.
-
CXR-ContraBench: Benchmarking Negated-Option Attraction in Medical VLMs
Medical VLMs frequently select negated options that contradict visible chest X-ray findings, achieving only ~30% accuracy on direct presence probes, but a post-hoc consistency verifier raises accuracy above 95%.
-
iTRIALSPACE: Programmable Virtual Lesion Trials for Controlled Evaluation of Lung CT Models
iTRIALSPACE generates realistic virtual lesion trials on lung CTs that isolate performance drivers and show strong transfer of model rankings to real clinical data (ρ=0.93).
-
LLMSpace: Carbon Footprint Modeling for Large Language Model Inference on LEO Satellites
LLMSpace is the first framework to jointly model operational and embodied carbon for LLM inference on LEO satellites, incorporating radiation-hardened hardware, peripheral systems, and workload patterns such as prefill-decode behavior.
-
Coral: Cost-Efficient Multi-LLM Serving over Heterogeneous Cloud GPUs
Coral cuts multi-LLM serving costs by up to 2.79x and raises goodput by up to 2.39x on heterogeneous GPUs through adaptive joint optimization and a lossless two-stage decomposition that solves quickly.
-
The Partial Testimony of Logs: Evaluation of Language Model Generation under Confounded Model Choice
An identification theorem shows that a randomized experiment and simulator together recover causal model values from confounded logs, with logs used only afterward to reduce estimation error.
-
TRIP-Evaluate: An Open Multimodal Benchmark for Evaluating Large Models in Transportation
TRIP-Evaluate is a new open multimodal benchmark with 837 text, image, and point-cloud items organized by a role-task-knowledge taxonomy to evaluate large models on transportation workflows.
-
A Systematic Survey of Security Threats and Defenses in LLM-Based AI Agents: A Layered Attack Surface Framework
A new 7x4 taxonomy organizes agentic AI security threats by architectural layer and persistence timescale, revealing under-explored upper layers and missing defenses after surveying 116 papers.
-
SPASM: Stable Persona-driven Agent Simulation for Multi-turn Dialogue Generation
SPASM introduces a stability-first framework with Egocentric Context Projection to maintain consistent personas and eliminate echoing in multi-turn LLM agent dialogues.
-
An Agentic Evaluation Architecture for Historical Bias Detection in Educational Textbooks
An agentic architecture with multimodal screening, a five-agent jury, meta-synthesis, and source attribution protocol detects biases in Romanian history textbooks more accurately than zero-shot baselines, achieving 83.3% acceptable excerpts and human preference in 64.8% of blind comparisons.
-
PlotChain: Deterministic Checkpointed Evaluation of Multimodal LLMs on Engineering Plot Reading
PlotChain benchmark reports top MLLMs reaching ~80% field-level accuracy on engineering plot reading under human-like tolerances, but with persistent failures on frequency-domain tasks like bandpass and FFT spectra.
-
Results-Actionability Gap: Understanding How Practitioners Evaluate LLM Products in the Wild
Qualitative study of 19 practitioners reveals ten LLM product evaluation practices and introduces the results-actionability gap as a key barrier to turning findings into improvements.
-
Surrogate Fidelity: When Can Open LLMs Explain Closed Ones?
Prediction agreement between open and closed LLMs substantially overstates agreement on attributions and causal reasons.
-
Measuring Behavior Portability in Large Language Models
A new framework measures behavioral portability of LLMs across payoff-equivalent environments and reports substantial systematic transfer losses in seven economic decision problems.
-
GIScholarBench: Benchmarking LLM Overconfidence in GIS Research
GIScholarBench shows LLMs exhibit consistent overconfidence across three scholarly tasks in GIS, with different manifestations in factual retrieval, citation expansion, and idea generation.
-
Evidence Markets
Evidence markets generalize prediction markets via a dynamic logarithmic market scoring rule that rewards evidence proportional to uncertainty, bounds platform loss, and makes truthful reporting an ε-DSIC strategy even under endogenous resolution.
-
Capability Advertisement as a Market for Lemons: A Trust Layer for Heterogeneous Agent Networks
Proposes a Trust Layer on top of existing agent protocols that adds probabilistic capability descriptors, screening, and reputation to enable separating equilibria and bound delegation reliability.
-
POIROT: Interrogating Agents for Failure Detection in Multi-Agent Systems
POIROT protocol repurposes agents in LLM multi-agent systems as an internal diagnostic layer for failure detection, outperforming single-LLM evaluators with gains that increase with complexity, agent count, and fault types.
-
Consistent and Distinctive: LLM Benchmark Efficiency via Maximum Independent Set Prompt Selection on Similarity Graphs
A graph-based MIS prompt selection method on embedding similarity graphs yields reduced benchmark subsets with highly consistent LLM rankings (Kendall's W ≥ 0.90 in 99.2% of cases) and 25-48% size reduction at higher thresholds.
-
BlueFin: Benchmarking LLM Agents on Financial Spreadsheets
BlueFin is a new benchmark for LLM agents on financial spreadsheets showing frontier models score below 50% with weaknesses in dynamic correctness.
-
When Does Persona Prompting Actually Help? A Retrieval and Metric Analysis of Expert Role Injection in LLMs
Persona prompting trades expertise depth for reduced clarity in LLM answers and works best on advisory questions in medicine and psychology.
-
A Paired Testing Protocol for Batch-Conditioned Refusal Robustness in LLM Serving
The paper introduces a paired testing protocol for batch-conditioned refusal robustness in LLM serving and reports low rates of genuine safety-label flips after adjudication, with a batch-invariant kernel ablation eliminating observed flips.
-
Intent Signal Theory: A Computational Framework for Intent-State Control in Human-AI Interaction
Intent Signal Theory formalizes four distinct intent-related objects in human-AI interaction, introduces a theorem on irreversible private intent loss, and reports supporting patterns from studies across LLMs, languages, and tasks.
-
A governance horizon for ethical-use constraints in open-weight AI models
Ethical constraint evidence on open-weight AI models decays with a half-life of 1.31 derivation steps on Hugging Face, creating a governance horizon at seven generations where 80% of models lack traceable information.
-
Towards Evaluation Engineering: An Empirical Study of ML Evaluation Harnesses in the Wild
An empirical study of 57 ML evaluation harnesses shows 41.4% of operational issues occur in the specification stage, driven mainly by unimplemented features, documentation gaps, and missing input validation.
-
PromptAudit: Auditing Prompt Sensitivity in LLM-Based Vulnerability Detection
PromptAudit evaluates five prompting strategies across five LLMs on 1000 CVEs and finds chain-of-thought prompting yields the strongest overall performance while adaptive chain-of-thought and self-consistency reduce effective results.
-
Mem-$\pi$: Adaptive Memory through Learning When and What to Generate
Mem-π is a framework using a dedicated model and decision-content decoupled RL to generate context-specific guidance on demand for LLM agents, outperforming retrieval baselines by over 30% on web navigation.
-
XDomainBench: Diagnosing Reasoning Collapse in High-Dimensional Scientific Knowledge Composition
XDomainBench shows LLMs suffer systematic reasoning collapse as domain composition order increases due to direct difficulty and interaction-amplified failures.
-
Continuous Discovery of Vulnerabilities in LLM Serving Systems with Fuzzing
GRIEF fuzzer finds 15 vulnerabilities including 2 CVEs in vLLM and SGLang by testing concurrent workloads for KV-cache isolation failures and cross-request interference.
-
Navigating the Sea of LLM Evaluation: Investigating Bias in Toxicity Benchmarks
Toxicity benchmarks for LLMs produce inconsistent results when task type, input domain, or model changes, revealing intrinsic evaluation biases.
-
OPT-BENCH: Evaluating the Iterative Self-Optimization of LLM Agents in Large-Scale Search Spaces
OPT-BENCH and OPT-Agent evaluate LLM self-optimization in large search spaces, showing stronger models improve via feedback but stay constrained by base capacity and below human performance.