NeurIPS 2026 reviews are up: reviews, scores, rebuttals and meta-reviews for 7,877 papers. See the highest rated
acceptodds: the papers of ICLR 2027
2,349 results for “keyword:"evaluation"” in ICLR 2027. 571 withdrawn or desk rejected listed apart. clearsort:relevanceratingnewestmost read
[PDF]A World Model Can Predict the Future Without Knowing Its Cause: Interventional Evaluation of Action-Conditioned World Models▲3.0k2.33
[PDF]How Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model Circuits▲5103.00
[PDF]Benign Safety Is Not a Robustness Certificate: Evidence of a Structural Evaluation Gap Across Attack Conditions▲11.00
[PDF]TCSALGBENCH: BENCHMARKING AUTOMATED PROVING FOR RESEARCH-LEVEL THEORETICAL COMPUTER SCIENCE▲5.4k3.00
[PDF]Risk Under Pressure: Compute-Aware Evaluation of Adversarial Robustness in Language Models▲3.0k2.33
[PDF]When Does Cross-Examination Beat Self-Assessment? A Text-Surface Visibility Law for Training-Free Agent Trustworthiness▼1361.33
[PDF]NovelAudit: A Benchmark for Open-Ended Narrative Consistency Auditing in Full-Length Novels▼4.3k2.00
[PDF]Balanced Perturbation and Subspace Erasure Do Not Identify What an Activation Monitor Reads▼1.2k1.67
[PDF]Is This Agent Evaluable? Certifying Verdicts About LLM Agents Under Nondeterminism, Oracle and Coverage Obstructions▼1.2k1.67
[PDF]WorkCraft-Arena: How Can Models Turn Heterogeneous Workspaces to Well-Designed Work Artifacts?▲3.0k2.33
[PDF]Interactive Visual Gym: Evaluating visual assistants with simulated human users in interactive worlds▲143.33
[PDF]Learning from What Goes Wrong: Failure-Aware World Models for Faithful Robot Policy Evaluation▼4.3k2.00
[PDF]Does Probe Conditioning Isolate Composition? An Information-Control Audit of Multi-Hop Evaluation▼1361.33
[PDF]Trust Labels Should Calibrate, Not Decide: Diagnosing and Reducing Source-Label Shortcuts in RL-Trained Fact-Checkers▲3.0k2.33
[PDF]Where Documentation Ends: Source-Derived Behavioral Specifications for LLM Tool Simulation▼3.7k2.25
[PDF]Disagreement in, Disagreement Out? Uneven Downstream Effects of Harm Classifier Differences▼5.8k2.33
[PDF]Expert-Grounded Synthetic Data for Cognitive Biases: Controlled Generation and Blind Evaluation of Construct Fidelity▼71.25
[PDF]Rethinking Link Prediction Evaluation: Full-Space Reference and Model-Screened Candidates▼1.2k1.67
Withdrawn or desk rejected (571)
[PDF]Machine Love: Emergent Bonding Behavior in Large Language Models and Its Implications for AI SafetyDesk rejected
[PDF]What a 3D Reconstruction Benchmark Can and Cannot Tell About Geometric Foundation ModelsWithdrawn2.00
[PDF]From Association to Decision Value: Testing Paired Instruction Dependence for VLA Corpus SelectionWithdrawn1.40