#Inference

ConflictScore introduces a new metric for evaluating language models' handling of conflicting evidence, measuring both the prevalence and balance of claims. It decomposes responses into claims, using ConflictScore-Count and ConflictScore-Ratio to quantify conflicts. The accompanying ConflictBench benchmark assesses various conflict types, demonstrating effective detection of overconfident claims and improving truthfulness on TruthfulQA.

The introduction of ConflictScore as a metric for evaluating language models' handling of conflicting evidence is significant for builders and PMs as it provides a standardized way to measure and improve model reliability. This can enhance user trust and application effectiveness, making it a critical consideration for investors looking at AI technologies focused on truthfulness and accuracy.

CRISP introduces a novel evaluation paradigm for visual spatial intelligence, revealing a disconnect between perception and reasoning in proprietary and open-source models. While proprietary models show strong latent reasoning, they struggle with metric estimation, whereas open-source models lack multi-hop reasoning capabilities. This framework shifts focus from simple guessing to genuine perception and reasoning.

The introduction of the CRISP evaluation paradigm highlights critical gaps in visual spatial intelligence among AI models, particularly the disparity in reasoning capabilities. Builders and PMs should consider these insights when developing applications that require accurate perception and reasoning, while investors may need to reassess the potential of both proprietary and open-source models based on their performance in real-world tasks.

This study compares few-shot prompting with Llama 4 Maverick and fine-tuned BERT (deepset/gbert-large) for classifying German climate news as threat or solution-oriented. BERT achieved an F1 score of 0.83, outperforming the LLM's 0.78, highlighting the effectiveness of contextual sentence input in classification tasks.

The study demonstrates that fine-tuned BERT can outperform few-shot prompting with Llama 4 Maverick in classifying climate news, achieving a higher F1 score. This indicates that for builders and PMs focused on NLP applications, leveraging specialized models like BERT may yield better performance in specific classification tasks, which is crucial for effective decision-making and strategy formulation.

A new multi-agent framework integrates 466,525 Reddit posts and 60,782 WebMD reviews with FDA records, achieving F1 scores of 0.969 for medications. This approach highlights the independent safety signals from patient-generated data, particularly for sertraline, where adverse events were reported much earlier than FDA records.

The development of a multi-agent framework that integrates patient-generated data with FDA records for mental health medications is significant as it demonstrates the potential for early detection of adverse events. Builders and PMs can leverage this approach to enhance drug safety monitoring systems, while investors may see opportunities in AI-driven healthcare solutions that prioritize patient insights.

ForeAgent is a novel forensics framework for AI-generated image detection, achieving 82.18% accuracy on the Chameleon benchmark, outperforming AIDE by 16.41%. It employs a Perception-Verdict architecture and a Hindsight-Driven Self-Refining strategy for continual self-improvement, demonstrating superior reasoning consistency compared to GPT-5.

The development of ForeAgent, a forensics framework for AI-generated image detection with 82.18% accuracy, highlights the growing need for reliable tools to combat misinformation and deepfakes. Builders and PMs should consider integrating such technologies to enhance content verification, while investors may find opportunities in companies focusing on AI ethics and security solutions.

This paper explores layer-specific prompt fusion in Vision Transformers (ViTs) using differentiable architecture search, proposing new fusion methods like affine transformation and cross-attention. Experiments on 34 datasets demonstrate improved performance over traditional prompt-tuning methods, highlighting the importance of fusion schemes in visual prompt tuning.

The development of layer-specific prompt fusion methods in Vision Transformers (ViTs) can significantly enhance the performance of visual models across various datasets. For builders and PMs, this means more effective tuning strategies for AI applications, while investors should note the potential for improved model capabilities that can lead to competitive advantages in the market.

The proposed self-supervised framework learns implicit 3D physics from video signals using a Volumetric Latent Space, achieving high structural stability and physical plausibility on benchmarks like CLEVERER and PhysInOne, without relying on traditional physics engines.

The development of Neural Voxel Dynamics introduces a self-supervised framework that learns 3D physics from video signals, which could significantly reduce reliance on traditional physics engines in game development and simulations. This innovation offers builders and PMs a more efficient way to create realistic environments, while investors may see potential for cost savings and enhanced product offerings in the gaming and simulation markets.

As coding agents evolve, verifying solutions becomes more challenging than generating them, necessitating a focus on scalable, faithful, and robust verification methods. The study reveals that no fixed reward function can sustain effectiveness as model capabilities advance, emphasizing the need for verification to evolve alongside solution generation.

The study highlights that as coding agents improve, the challenge of verifying their outputs will grow, indicating a need for builders and PMs to invest in scalable verification methods. For investors, this signals an opportunity to support innovations that focus on robust verification frameworks, which are essential for maintaining trust in automated solutions.

AlgoEvolve leverages Large Language Models to evolve algorithmic trading strategies, demonstrating superior performance over human-designed methods. The framework adapts trading rules autonomously and utilizes a meta-evolutionary approach to enhance prompt generation, significantly reducing zero-trade failures.

The development of AlgoEvolve, which uses LLMs for the autonomous evolution of trading strategies, signals a shift towards more adaptive and efficient trading systems. Builders and PMs can leverage this technology to enhance algorithmic trading solutions, while investors may see reduced risk and improved returns due to the framework's ability to minimize zero-trade failures.

This study introduces an iterative data generation pipeline for isolating cascading linear features to detect and control sycophancy in language models. By moving beyond binary sample pairs, the method enhances interpretability and performance, outperforming existing baselines in detection and steering with lower computational costs. The findings suggest that sycophancy features form linearly separable subspaces, improving model activation selection.

The introduction of an iterative data generation pipeline to detect and control sycophancy in language models enhances interpretability and performance while reducing computational costs. This development is significant for builders and PMs as it allows for more effective model tuning and deployment, ultimately leading to better user interactions and trust in AI systems.

APRTrack introduces a hierarchical perturbation and retrieval framework for RGB-Event visual object tracking, enhancing robustness against partial target loss and modal degradation. The model utilizes adversarial perturbation to simulate real-world signal corruption and employs Footprint-guided Channel-calibrated Hopfield Retrieval for effective historical information compensation. Extensive experiments on multiple datasets demonstrate its effectiveness in challenging tracking scenarios.

The introduction of APRTrack, which uses adversarial perturbation for RGB-Event visual object tracking, signifies a significant advancement in tracking robustness. For builders and PMs, this means improved performance in real-world applications, while investors should note the potential for enhanced product offerings in sectors reliant on reliable visual tracking technology.

A multimodal deep learning model with a feature-guided methane enhancement mechanism achieves superior methane plume segmentation on the MPDataset, improving MIoU by +0.92, MPrecision by +0.87, and Recall by +1.01, while maintaining lower computational costs compared to existing architectures.

The development of a multimodal deep learning model for methane plume segmentation, which improves MIoU, Precision, and Recall while reducing computational costs, signals a significant advancement in environmental monitoring technology. Builders and PMs can leverage this model for more efficient emissions tracking, while investors may see opportunities in sustainable tech solutions addressing climate change.

ContextForge enhances long-horizon reasoning in large language models (LLMs) by recycling context through structured query generation and external memory retrieval. In a 15-turn conversational benchmark, it shows improved consistency and reduced token usage compared to baseline models, maintaining response accuracy. This approach allows LLMs to extend their capabilities without larger context windows or retraining.

The development of ContextForge for long-horizon reasoning in LLMs enables builders and PMs to create applications that maintain conversational context over extended interactions without increasing computational costs. This innovation reduces the need for larger context windows, allowing for more efficient use of resources while improving user experience, which is crucial for investors looking for scalable AI solutions.

The evaluation of Multimodal Large Language Models (MLLMs) is lagging behind their rapid advancements, with existing benchmarks failing to assess cross-modal integration. Key gaps include temporal-spatial coherence and multimodal consistency, which are essential for accurately measuring multimodal intelligence progress.

The identification of gaps in evaluating Multimodal Large Language Models (MLLMs) highlights the need for improved benchmarks that assess cross-modal integration. Builders and PMs should prioritize developing metrics that ensure MLLMs are effectively measuring multimodal intelligence, while investors should consider funding projects that address these critical evaluation challenges to enhance product reliability and market competitiveness.

A hybrid approach combining image processing and CNNs predicts fruit freshness with over 90% accuracy, using logistic regression to streamline real-time classification without high computational demands. This method addresses agricultural spoilage issues effectively, though it requires fruits to be isolated on specific backgrounds.

The development of a hybrid machine learning and image processing approach for predicting fruit quality with over 90% accuracy is significant for builders and PMs in the agricultural tech space, as it enables efficient real-time classification of produce, potentially reducing spoilage and increasing supply chain efficiency. Investors should note the scalability of this technology, which addresses a critical need in food preservation.

The Geometry-Aware Monte Carlo Tree Search (MCTS) framework significantly improves solutions for extremal problems in combinatorial geometry, reducing constraint checking complexity from O(n^3) to O(n^2). This framework achieved new best-known results for five out of six problems, including configurations of size approximately 1.8n for Max-N3IL and 0.95n for the Smallest Complete Set problem.

The development of the Geometry-Aware Monte Carlo Tree Search (MCTS) framework, which reduces constraint checking complexity from O(n^3) to O(n^2), is significant for builders and PMs as it enables more efficient algorithms for solving complex combinatorial problems. This could lead to faster and more scalable solutions in applications such as optimization and AI planning, attracting potential investors interested in advanced computational techniques.

This study demonstrates that integrating EEG signals with eye-tracking data significantly enhances automatic keyphrase extraction (AKE) from microblogs. Using the ZuCo corpus, the research shows that EEG features provide the most substantial performance improvements, indicating their potential as valuable cognitive evidence for AKE models.

The integration of EEG signals with eye-tracking data for automatic keyphrase extraction (AKE) represents a significant advancement in natural language processing. This development suggests that incorporating cognitive signals can enhance AKE models, offering builders and PMs new avenues for improving content analysis tools, while investors may see potential for innovative applications in AI-driven marketing and information retrieval systems.

The study introduces the Turbid Underwater Baseline (TUB) dataset with 1,320 images and over 16,000 segmentation masks to quantify information loss in turbid underwater scenes. A new metric, PCD, shows a strong correlation with instance segmentation model performance, outperforming traditional metrics in assessing real-world turbidity effects.

The introduction of the Turbid Underwater Baseline (TUB) dataset and the new PCD metric provides builders and PMs with a reliable tool to evaluate and improve instance segmentation models in challenging underwater environments. This advancement can lead to better performance in applications like underwater robotics and environmental monitoring, making it a valuable consideration for investors in AI and robotics sectors.

AnySimLite is a lightweight similarity encoder designed for on-device speech-adjacent classification, achieving state-of-the-art performance in few-shot settings while using less than 1/250th the model size of the qLLaMA_LoRA-7B baseline. It effectively combines word-level and character-level channels to minimize memory footprint and maintain low inference latency on edge devices.

The development of AnySimLite, a lightweight few-shot similarity encoder for on-device speech-adjacent classification, is significant as it allows builders and PMs to implement advanced AI capabilities on edge devices without heavy resource requirements. This opens up new opportunities for investors in the AI space, particularly in applications requiring efficient processing and low latency.

CORE-Bench Hard reveals that after accuracy saturation, evaluating agent performance on dimensions like efficiency and reliability provides deeper insights. The introduction of CORE-Bench v1.1 and CORE-Bench OOD enhances measurement capabilities, showing significant performance uplift from human-agent collaboration, with speed improvements around twofold.

The introduction of CORE-Bench v1.1 and CORE-Bench OOD provides a new framework for evaluating AI agents beyond accuracy, emphasizing efficiency and reliability. This shift allows builders and PMs to better understand the practical performance of their systems, while investors can identify more nuanced metrics for assessing AI solutions, potentially leading to more informed funding decisions.

The study identifies compositional behavioral leakage (CBL) in prompt-composed systems, where editing one module affects others without direct dependencies. Testing on Claude Sonnet 4.6 revealed significant interference through content changes, highlighting the need for cross-module interference measurement in agent evaluation.

The identification of compositional behavioral leakage (CBL) in prompt-composed systems, as seen in Claude Sonnet 4.6, underscores the importance of measuring cross-module interference in AI agents. For builders and PMs, this signals a need to refine evaluation methods to ensure module independence, while investors should recognize potential risks in system reliability and performance.

Larger models like Qwen3-32B and GPT-OSS-120B outperform their smaller counterparts by 6.43% and 7.38% respectively on reasoning benchmarks. The AdvCluster framework reveals that these models excel in Constraint-Guided Reasoning, effectively identifying and organizing constraints to enhance reasoning accuracy.

The performance improvements of larger models like Qwen3-32B and GPT-OSS-120B in Constraint-Guided Reasoning highlight the importance of model size in enhancing reasoning accuracy. Builders and PMs should consider leveraging these advanced models for applications requiring complex decision-making, while investors may see potential for higher returns in companies adopting these technologies.

The Frame Forgetting Network (FFN) introduces a novel approach to Test Time Training (TTT) for long videos, optimizing computational efficiency by processing only three frames at a time. This method reduces unnecessary computations and adapts to new information effectively, demonstrating significant performance improvements on dense-segmentation and video classification tasks using a new dataset of up to 3-hour long videos.

The introduction of the Frame Forgetting Network (FFN) for Test Time Training (TTT) optimizes video processing by focusing on three frames at a time, which enhances computational efficiency and adaptability. This development is crucial for builders and PMs in video analytics and AI applications, as it enables more effective handling of long video content with reduced resource consumption.

TaskTok introduces a framework for Task-Driven Image Restoration (TDIR) that selectively refines task-relevant tokens, improving computational efficiency and performance in image classification, semantic segmentation, and object detection. By focusing on unevenly distributed visual information, TaskTok enhances task performance significantly while minimizing unnecessary updates to latent tokens.

TaskTok's framework for Task-Driven Image Restoration (TDIR) enhances computational efficiency and performance in key computer vision tasks like image classification and object detection. This development signals a shift towards more efficient AI models, which can lead to reduced operational costs and faster deployment for builders, PMs, and investors in the AI space.

PhyEditBench introduces a benchmark for evaluating physics-aware image editing models, featuring 238 real-world instances and 35 synthetic cases. The study reveals significant limitations in current state-of-the-art methods, while the proposed PhyWorld baseline demonstrates superior performance through innovative reasoning mechanisms.

The introduction of PhyEditBench provides a comprehensive benchmark for physics-aware image editing, highlighting the limitations of current models and the effectiveness of the PhyWorld baseline. This signals to builders and PMs the need to invest in innovative reasoning mechanisms to enhance image editing capabilities, while investors should note the potential for improved performance in a growing market.

Dynamic-dLLM introduces a training-free framework that enhances inference efficiency of diffusion LLMs like LLaDA-8B-Instruct by over 3 times. It employs Dynamic Cache Updating and Adaptive Parallel Decoding to optimize performance on benchmarks such as MMLU and GSM8K, outperforming existing acceleration methods. This solution allows for efficient deployment without sacrificing model performance.

The introduction of Dynamic-dLLM, which enhances the inference efficiency of diffusion LLMs by over 3 times without requiring training, is significant for builders and PMs as it enables faster deployment and cost-effective scaling of AI applications. For investors, this advancement signals a competitive edge in the rapidly evolving AI landscape, potentially leading to higher returns on investment.

LogicIR introduces a novel Logic Gate Network for image restoration, achieving strong performance with reduced computational costs. This UNet-inspired architecture utilizes logic gates and includes a differentiable bit decoding layer, enhancing information propagation. Experimental results show its effectiveness across multiple benchmarks, making it a promising alternative in the field.

The introduction of LogicIR's Logic Gate Network for image restoration highlights a significant advancement in computational efficiency and performance. Builders and PMs can leverage this architecture to reduce costs while improving image processing capabilities, making it a competitive alternative in AI-driven imaging solutions, which could attract investor interest in scalable applications.

The Narration-of-Thought (NoT) system prompt significantly enhances ethical reasoning in large language models, reducing stakeholder collapse from 31% to under 1% and uncertainty suppression from 72% to 1-24% across four model generators. This method requires no additional training and achieves a consensus increase from 6% to 95% in multi-stakeholder debates, providing a robust framework for ethical decision-making.

The development of the Narration-of-Thought (NoT) system significantly improves ethical reasoning in large language models, reducing stakeholder collapse from 31% to under 1%. This advancement allows builders and PMs to implement more reliable AI systems for decision-making, while investors can recognize the potential for increased trust and adoption in AI applications.

This paper evaluates confidence interval methods for classifier performance metrics in text classification, highlighting that traditional methods like the Wald interval are often inaccurate. It proposes improved techniques such as Agresti-Coull and a novel pseudo-count regularized bootstrap, particularly for small datasets and nested data scenarios, enhancing transparency in machine learning applications.

The paper introduces improved confidence interval methods for classifier performance metrics, particularly in small datasets and nested data scenarios. For builders and PMs, adopting these techniques can enhance model reliability and transparency, leading to better decision-making; investors should note that robust performance evaluation can increase the attractiveness of AI products in the market.