Paper Radar
New AI papers from arXiv, each summarized by a small model on a laptop and compared with recent papers. Part of the AI Experiments. Click a paper for its summary and the closest earlier papers.
How this works, and its limits
Summaries are written by a 0.5B model and not edited by hand. Each one is checked that its numbers and names appear in the paper. Summaries that fail are withheld, and a summary can still be wrong when it passes. The model only reads the first few pages of each paper (about 10,000 characters), so results and conclusions are often missing from its summary. More on this gap.
Similar papers are found only among cs.AI, cs.LG and cs.CL papers submitted since Nov 2, 2026, so older related work won't show up and a paper can look newer than it is. The similarity score (0 to 1) measures overlap in topic and wording, not whether the idea is the same.
20 papers · 1 summary withheld
AI Isn't Getting Dumber, We're Getting Pickier: Rising Standards in Human Evaluation of Large Language ModelsCompares the evaluation of large language models (LLMs) over time, showing that the ratings of unchanged models decrease over time.
The paper compares the evaluation of large language models (LLMs) over time, showing that the ratings of unchanged models decrease over time. The ratings of fixed models, such as dated snapshots and open-weight models, show no change over time. The rise in the number of votes saying both models are bad over time, indicating a decline in the quality of the ratings. The findings suggest that the evaluation of LLMs is influenced by the ratings themselves and the composition of the raters, and that the system's state, such as prompts and content, does not fully account for the decline. The findings are consistent across models and across different periods, indicating that the evaluation of LLMs does not always follow the same pattern.
An Energy-Based Perspective on Machine LearningAddresses the problem of learning relationships among variables and using these relationships to infer unknown quantities, proposing a unified perspective on energy-based models (EBMs) that represent the compatibility of configurations through scalar energy functions.
The paper addresses the problem of learning relationships among variables and using these relationships to infer unknown quantities, proposing a unified perspective on energy-based models (EBMs) that represent the compatibility of configurations through scalar energy functions. It shows that any EBM is specified by three choices: the energy, the inference procedure, and the learning method. The learning methods share one contrastive structure, and a unified view through contrastive learning connects learned models at inference time. The paper also discusses the roles of representation, inference, and composition in the problem of learning relationships among variables and using these relationships to infer unknown quantities.
Bits or permutations? Spending a fixed inference budget on weight precision or on option-order averaging in small local language modelsInvestigates whether low-bit round-to-nearest quantization, when paired with option-order averaging, can improve accuracy and slot skew in small language models, particularly in cost-matched pairs.
The paper investigates whether low-bit round-to-nearest quantization, when paired with option-order averaging, can improve accuracy and slot skew in small language models, particularly in cost-matched pairs. It finds that higher bit loss leads to increased accuracy but decreases slot skew, suggesting that rounding can be more advantageous than option order averaging. The study demonstrates that the choice of bit budget and permutation strategy can significantly impact performance, with the choice of more bits and option permutations leading to higher accuracy but decreased slot skew.
Blurring the Block Boundaries: Overlapping Block Denoising for Efficient Diffusion Language ModelsAddresses the problem of overlapping block denoising in diffusion language models (DLMs), where tokens are not equally easy to predict and the number of tokens accepted per forward pass drops near the block boundaries.
The paper addresses the problem of overlapping block denoising in diffusion language models (DLMs), where tokens are not equally easy to predict and the number of tokens accepted per forward pass drops near the block boundaries. It proposes Bluring the Block Boundaries (B3), a method that refines two blocks in parallel within a shared forward pass, ensuring that the next block is drafted in a shared forward pass. The method is hyperparameter-free and requires no training or hyperparameters tuning. Evaluated on various DLMs, B3 achieves a significant speedup in generation quality without requiring training or hyperparameters tuning.
Causal Concept Explanations for Deep Neural ModelsAddresses the issue of interpretability in autonomous systems, where concepts are not always interpretable due to correlational relationships.
The paper addresses the issue of interpretability in autonomous systems, where concepts are not always interpretable due to correlational relationships. It proposes Causal Concept Wrapper Network (CCW-Net), which uses interventional mediation analysis from fields such as economics and medicine to shape model concepts representations such that they are both necessary and sufficient for outputs. CCW-Net quantifies the share of a model's output causally attributable to each concept, demonstrating that such explanations are not merely interpretable but quantifiably causal.
CLADE: Category-Level Answering from Deficient Evidence in Retrieval-Augmented Language ModelsAddresses the issue of resolvable insufficiency in retrieval-augmented language models, where coarse passages are insufficient but can be resolved by a domain ontology.
The paper addresses the issue of resolvable insufficiency in retrieval-augmented language models, where coarse passages are insufficient but can be resolved by a domain ontology. It proposes and measures the phenomenon of resolvable insufficiency, where a model's abstention is triggered by a coarse passage, and a domain ontology can turn this into the exact fine-grained answer. The findings show that domain ontologies can help resolve this issue, and a counterfactual test confirms this.
Cognitive Mapping of Learning by Thinking Mechanisms in Large Language ModelsProposes a Cognitive LbT Framework to map four mechanisms (Reflective Reasoning, Synthetic Memory Formation, Internal Simulation, and Conceptual Self-Organization) onto computational processes in Large Language Models (LLMs).
The paper proposes a Cognitive LbT Framework to map four mechanisms (Reflective Reasoning, Synthetic Memory Formation, Internal Simulation, and Conceptual Self-Organization) onto computational processes in Large Language Models (LLMs). The framework integrates existing findings into a unified taxonomy for methods studied in isolation. It argues that selected methods can functionally approximate aspects of these mechanisms and improve performance on some tasks without additional labeled examples, although this does not always constitute durable learning in model parameters. The framework offers testable propositions about when LbT in LLMs should succeed or fail and offers a structured basis for designing more reliable autonomous learning systems.
Comparative Analysis of Agent Evaluation Frameworks: Stability, Detection, and Discovery in Enterprise Analytics AgentsEvaluates three evaluation frameworks for agentic systems, finding that trace-based LLM judging, text-only LLM judging with red-teaming, and deterministic heuristics are distinct and complementary approaches.
The paper evaluates three evaluation frameworks for agentic systems, finding that trace-based LLM judging, text-only LLM judging with red-teaming, and deterministic heuristics are distinct and complementary approaches. Trace-based LLM judging, where the evaluation harness captures the full execution trace, is shown to be more stable and reliable than text-only LLM judgers. The text-only LLM judgers, which rely on red-teaming, are found to be more effective in detecting known failures. The study demonstrates that these frameworks are complementary rather than substitutable.
Contrastive Discovery: Open-Ended Scientific Discovery over Competing ExplanationsIntroduces Contrastive Discovery, a method for open-ended scientific discovery that uses contrast sets to represent and score hypotheses.
The paper introduces Contrastive Discovery, a method for open-ended scientific discovery that uses contrast sets to represent and score hypotheses. It shows that contrast sets change the unit an agent proposes and scores, not its architecture. The method is evaluated across six domains, demonstrating the effectiveness of contrast sets in resolving scientific questions.
EngineAgent: Agentizing On-Device LLM Inference Engines with Inference Stack Hierarchy Awareness and Hardware AbilitiesEngineAgent, an agentic system for automating on-device LLM model enablement, separates reusable engine and hardware knowledge from model-specific implementation.
EngineAgent, an agentic system for automating on-device LLM model enablement, separates reusable engine and hardware knowledge from model-specific implementation. It constructs a hierarchy-aware engine index that captures engine abstraction layers and cross-layer dependencies, enabling developers to coordinate cross-layer changes. For new models, a builder uses this knowledge to coordinate cross-layer changes and checks implementation completeness through device-side validation. The system reduces average per-task time and model-API cost compared to baselines, demonstrating its effectiveness in supporting multiple models.
Evaluating LLM-Generated World Models for Neuro-Symbolic ReasoningEvaluates five Large Language Models (LLMs) for generating world models, focusing on three key aspects: mechanical validity, structural fidelity, and semantic criteria.
The paper evaluates five Large Language Models (LLMs) for generating world models, focusing on three key aspects: mechanical validity, structural fidelity, and semantic criteria. It finds that while LLMs can produce models that are partially observable stochastic games (POSGs), they do not necessarily represent the intended decision problem. The study also shows that the generated world models are not necessarily faithful, as some assumptions are omitted or unsupported.
Forward-Value Discretization: Faster Stochastic Diffusion Sampling Without High-Order ApproximationAddresses the problem of stochastic sampling in diffusion models, where higher-order discretization accelerates deterministic sampling but its benefits are unclear.
The paper addresses the problem of stochastic sampling in diffusion models, where higher-order discretization accelerates deterministic sampling but its benefits are unclear. It proposes Forward-Value Discretization (F-DDPM), a training-free framework that evaluates the model at a lookahead estimate constructed from previous predictions. The model improves sample quality at low NFE budgets without increasing the convergence order. The framework is shown to be first-order convergent, even with exact score functions, and it can reuse previous model predictions and injected noise.
Medical Image Segmentation: Foundation Model Adaptation versus Task-Specific TrainingCompares the performance of foundation models trained on large natural-image collections versus task-specific models trained from random initialisation.
The paper compares the performance of foundation models trained on large natural-image collections versus task-specific models trained from random initialisation. It finds that foundation models, when adapted, generally perform better than task-specific models, even when the task-specific models are not trained from random initialisation. The study also finds that the task-specific models remain competitive when training data are abundant, and delineate boundaries more precisely. However, only a small fraction of the foundation models require adaptation, and the capacity needed is determined by the imaging domain rather than the adaptation method.
One Policy for 100 Household Tasks: Our Solution for the 2026 BEHAVIOR ChallengeDescribes a solution for the 2026 BEHAVIOR Challenge, which involves a single vision-language-action (VLA) model built on the 1st-place 2025 solution.
The paper describes a solution for the 2026 BEHAVIOR Challenge, which involves a single vision-language-action (VLA) model built on the 1st-place 2025 solution. The model is trained for 100 tasks, each with a specific task and stage embedding. The model's performance is evaluated on 50 tasks, which are shared with the 2025 solution. The model's performance is found to be undertrained, indicating that the model's training time is too long and the model's performance is not consistent.
Precision on Demand: Escalating Weight Bits at Uncertain Decoding Steps in a Small Local ModelSummary withheld: it failed the check.
Summary withheld. It had a number or name the check couldn't find in the paper, so it isn't shown. Read the abstract on arXiv instead. Why summaries can be withheld.
Quantifying the Sampling Frequency Invariance of Time-Series Foundation Models for Sensor Data ApplicationsEvaluates four prominent Time-Series Foundation Models (TSFMs) including MOMENT, Chronos-2, Moirai-MoE, and TimesFM, assessing their invariance to changes in sampling frequency (SF).
The paper evaluates four prominent Time-Series Foundation Models (TSFMs) including MOMENT, Chronos-2, Moirai-MoE, and TimesFM, assessing their invariance to changes in sampling frequency (SF). The authors introduce novel metrics such as Frequency Sensitivity Index (FSI), Forecast Degradation Index (FDI), and Classification Degradation Index (CDI) to quantify how these models behave under varying SF conditions. Key findings include the observation that the models' embeddings begin to drift immediately with SF reductions, but their forecasting and downstream performance degrade more gradually. The study reveals that the extent of SF-invariance varies based on the gap between the native SF and the bandwidth required for the downstream tasks, rather than the choice of the TSFM itself. The results provide insights for practitioners on when to expect robust performance from TSFMs under varying SF conditions.
Re-Search: Dissecting LLM Ideation Through a Two-Player Questioning GameIntroduces RE-SEARCH, a two-player questioning game designed to isolate the research ideation process from execution, making every step of ideation observable and measurable.
The paper introduces RE-SEARCH, a two-player questioning game designed to isolate the research ideation process from execution, making every step of ideation observable and measurable. Six recent LLMs, including GPT-5.4, demonstrate that the bottleneck is generation, not judgment, leading to slow progress. The findings suggest that models recognize good ideas or questions when offered, but fail to generate them on their own, indicating that external candidates are needed to improve autoresearch systems.
ReTrace: Distilling Tool-Using Agents into Planner-Free WorkflowsReTrace addresses the problem of automating business processes by reusing past experiences, which are stored in agent traces.
ReTrace addresses the problem of automating business processes by reusing past experiences, which are stored in agent traces. It combines agent reasoning with tool calls and input sources to extract action conditions, input sources, and alternatives. The generated workflow is then compiled into a reusable workflow that can be deployed without requiring a large language model agent. The paper provides a method to retrace the workflow of an agent, which is then compiled into executable workflows. The results show that ReTrace reduces costs and improves performance, especially when executed on production platforms.
Toward Foundation Detection Models via Example-Based Grounding in EM-DETRAddresses the limitations of supervised object detectors, proposing a multi-domain foundation model that integrates diverse medical image datasets annotated for both anatomical structures and abnormalities.
The paper addresses the limitations of supervised object detectors, proposing a multi-domain foundation model that integrates diverse medical image datasets annotated for both anatomical structures and abnormalities. It integrates diverse medical image datasets annotated for both anatomical structures and abnormalities to enrich supervision during baseline training. The model achieves state-of-the-art results in CXRs and mammography, demonstrating state-of-the-art performance on five key findings in CXRs and mammography.
Two Communities, One Problem: Branch-and-Bound vs. A* for Resource-Constrained Project SchedulingInvestigates the performance of Branch-and-Bound (B&B) and A* methods for Resource-Constrained Project Scheduling (RCPSP) problems, focusing on their differences in search mechanisms.
The paper investigates the performance of Branch-and-Bound (B&B) and A* methods for Resource-Constrained Project Scheduling (RCPSP) problems, focusing on their differences in search mechanisms. B&B, being a heuristic method, is based on lower bounds and state evaluations, while A* is based on state evaluations and lower bounds. The study finds that A* outperforms B&B under certain conditions, demonstrating the importance of duplicate-state detection and pruning.