The business of intelligence is moving faster than the technology itself. We track every major round, series, and acquisition to keep founders and investors ahead of the curve.
AiPX Ventures has officially launched in Australia as an AI-native venture studio aimed at collaborating with early-stage founders to create innovative software platforms. This initiative is spearheaded by Sally Tobin, a seasoned entrepreneur with extensive experience in commerce marketing and technology product innovation. Unlike traditional venture capital funds, AiPX operates as a venture studio, actively engaging with founders to identify market opportunities, design products, and develop minimum viable products (MVPs) for market entry. The studio's focus is on sectors that still rely on fragmented technology stacks, with the goal of simplifying workflows and enhancing operational decision-making through AI-native solutions. The leadership team includes Gary Head, who joins as Chief Product Officer, bringing over 15 years of experience in advertising and product design. His role will involve overseeing product architecture and platform development within the studio's portfolio. AiPX plans to develop a pipeline of four AI-native platforms centered around workflow systems and integrated partner marketplaces, targeting industries such as trades, construction, and marketing operations. The first product is expected to launch mid-year, although specific funding arrangements and technology details remain undisclosed. The venture studio model reflects a growing trend towards applied AI solutions tailored to specific industry needs, positioning AiPX as a significant player in the evolving landscape of AI-driven business solutions.
The paper presents a significant advancement in the field of vision-language models (VLMs), which are designed to integrate and reason with both visual and textual inputs. The authors identify a critical challenge in the current architecture of VLMs: the reliance on visual inputs only at the onset of generation, which leads to a predominance of textual reasoning that can result in compounded errors from initial visual grounding. This issue is exacerbated by the coarse and noisy nature of existing guidance mechanisms for visual grounding during inference, making it difficult to maintain accuracy in reasoning over extended textual outputs. To tackle these challenges, the authors propose the Saliency-Aware Principle (SAP) selection method. This innovative approach emphasizes high-level reasoning principles rather than focusing on token-level trajectories, which allows for more stable control over the generation process, even in the presence of noisy feedback. Notably, SAP enables the model to revisit visual evidence during later reasoning steps, thereby enhancing the accuracy of visual grounding when necessary. Furthermore, SAP facilitates multi-route inference, allowing for parallel exploration of various reasoning behaviors, which can lead to richer and more nuanced outputs. The authors highlight that SAP is model-agnostic and does not require additional training, making it a versatile solution applicable across different VLM architectures. Empirical evaluations demonstrate that SAP significantly reduces the phenomenon of object hallucination—where the model generates incorrect or fabricated visual information—while maintaining competitive performance within the same token-generation budgets. Additionally, the implementation of SAP results in more stable reasoning processes and reduced response latency compared to traditional chain-of-thought (CoT) reasoning methods. This research not only addresses pressing issues in VLMs but also sets the stage for future developments in multimodal AI systems, emphasizing the importance of effective visual grounding in enhancing the overall reasoning capabilities of these models.
The research paper addresses the critical challenges in the interpretability of large language models (LLMs), particularly focusing on the limitations of existing interpretability studies. The authors argue that while significant insights have been gained regarding model behavior, persistent issues such as non-generalizable findings and causal interpretations that exceed the available evidence remain prevalent. The paper emphasizes the importance of causal inference as a framework for establishing valid mappings from model activations to high-level structures that are invariant across different contexts. By employing Judea Pearl's causal hierarchy, the authors delineate the boundaries of what interpretability studies can substantiate.
The paper outlines the distinction between mere observations, which identify associations between model behavior and internal components, and more rigorous interventions, such as ablations or activation patching, that can demonstrate how specific modifications influence behavioral metrics like changes in token probabilities across various prompts. However, the authors highlight a significant gap in the ability to make counterfactual claims—predictions about model outputs under hypothetical unobserved interventions—without the presence of controlled supervision. This limitation underscores the need for a more robust methodological approach to causal inference in the context of LLMs.
To address these challenges, the authors propose the concept of causal representation learning (CRL), which operationalizes the causal hierarchy and clarifies which variables can be inferred from model activations and under what assumptions. This framework not only aids in understanding the causal relationships within LLMs but also provides a diagnostic tool for practitioners. It helps in selecting appropriate methods and evaluations that align claims with empirical evidence, thereby enhancing the generalizability of findings. Overall, this research contributes to the ongoing discourse on model interpretability by advocating for a more structured and evidence-based approach to causal inference, ultimately aiming to improve the reliability and applicability of interpretability studies in the field of artificial intelligence.
In the realm of single-cell RNA sequencing (scRNA-seq), the inherent statistical structure of the data has spurred the development of sophisticated computational models, notably the TranscriptFormer, which employs transformer-based architectures to generate representations of gene expression. These representations, or embeddings, have demonstrated remarkable efficacy in various downstream applications, including cell-type classification, disease-state prediction, and cross-species learning. However, the reliance on deep learning methods raises questions about the necessity of such complex models when simpler approaches may yield comparable results. This study investigates the performance of straightforward, interpretable methodologies that leverage careful normalization and linear techniques against the backdrop of established benchmarks for single-cell foundation models. The authors report that their simplified pipelines achieve state-of-the-art (SOTA) or near SOTA performance across multiple evaluation metrics, even outperforming existing foundation models in scenarios involving out-of-distribution tasks with novel cell types and organisms not included in the training datasets. These findings underscore the importance of rigorous benchmarking in the field and suggest that the biological nuances of cell identity can be effectively captured through linear representations of single-cell gene expression data. This research not only challenges the prevailing notion that deep learning is indispensable for high-performance outcomes in scRNA-seq analysis but also emphasizes the potential for more interpretable and computationally efficient methods to contribute meaningfully to our understanding of cellular biology.
Ripple has made significant strides in expanding its operational capabilities through a series of acquisitions totaling nearly $3 billion since 2023. These strategic moves are aimed at establishing Ripple as a comprehensive financial infrastructure provider, enhancing the utility of its cryptocurrency, XRP. The acquisitions span various sectors, including custody, brokerage, treasury management, and stablecoin infrastructure, all designed to facilitate institutional adoption of XRP. Notably, Ripple's acquisition of Metaco, a Swiss custody provider, and Hidden Road, a prime brokerage firm, positions the company to cater to large financial institutions, thereby increasing the potential demand for XRP. However, while these acquisitions strengthen Ripple's ecosystem, the immediate impact on XRP's price is expected to be limited as the integration of these services will take time to materialize. The article outlines a phased outlook for XRP's price, suggesting that significant price movements may not occur until 2026, contingent upon the successful rollout of these new services and the extent to which institutions begin holding XRP rather than merely using it as a transactional bridge. The bullish case predicts XRP could reach between $3.50 and $5.00 if institutions adopt XRP for long-term holding, while a bearish scenario could see prices drop to between $1.20 and $1.85 if the acquisitions fail to drive sustained demand. Overall, Ripple's investments represent a long-term vision for creating a robust financial ecosystem that could eventually lead to increased XRP usage and price appreciation, but the path to achieving this remains uncertain and dependent on market dynamics.
Compositional generalization is a crucial cognitive ability that allows humans to understand and reason about new combinations of familiar concepts. This capability poses a significant challenge in the field of machine learning, particularly within the domain of visual question answering (VQA). The study presented in this paper addresses this challenge by investigating the effectiveness of object-centric (OC) representations in enhancing compositional generalization in visually rich environments. The authors introduce a benchmark designed to evaluate how well various vision encoders, both with and without object-centric biases, can generalize to unseen combinations of object properties across three distinct visual worlds: CLEVRTex, Super-CLEVR, and MOVi-C.
To ensure a robust and fair evaluation, the authors meticulously controlled for several factors, including training data diversity, sample size, representation size, downstream model capacity, and computational resources. They utilized two prominent vision encoders—DINOv2 and SigLIP2—as the foundational models, alongside their object-centric counterparts. The results of their experiments yield several key insights: firstly, OC approaches demonstrate superior performance in more challenging compositional generalization scenarios, suggesting that these representations are particularly adept at handling complex visual reasoning tasks. Conversely, traditional dense representations tend to excel only in simpler settings, often requiring significantly more computational resources to achieve comparable results. Secondly, the study highlights the sample efficiency of OC models, which can achieve robust generalization with fewer training images compared to dense encoders. The latter only begin to match or exceed the performance of OC models when provided with ample data and diversity in training samples. Overall, the findings underscore the potential of object-centric representations to facilitate stronger compositional generalization, particularly when constraints are placed on dataset size, training diversity, or computational capacity. This research contributes valuable insights into the ongoing discourse surrounding the development of more effective machine learning models capable of mimicking human-like reasoning abilities in complex visual environments.
The paper presents a significant advancement in the field of audio language modeling, addressing the limitations of current models that primarily focus on text-first approaches. These traditional models either extend pre-trained text-based large language models (LLMs) or utilize semantic-only audio tokens, which restricts their ability to perform comprehensive audio modeling. The authors propose a novel framework for native audio foundation models that utilize next-token prediction directly on audio data at scale. This approach aims to jointly model semantic content, acoustic details, and textual information, thereby enhancing both general audio generation and cross-modal functionalities.
The research is structured around a systematic empirical study that provides extensive insights into the design and training of these audio models. The authors begin by investigating critical design choices, including the selection of data sources, the ratios of text mixtures, and the composition of tokens. This investigation culminates in a validated training recipe that serves as a foundation for building effective audio models.
A pivotal aspect of the study is the introduction of the first scaling law analysis for discrete audio models, conducted through IsoFLOP analysis across 64 models with computational requirements ranging from $3{ imes}10^{18}$ to $3{ imes}10^{20}$ FLOPs. The findings reveal that optimal data size increases at a rate of 1.6 times faster than the optimal model size, providing crucial insights for future model scaling and training efficiency.
Building on these empirical findings, the authors introduce SODA (Scaling Open Discrete Audio), a suite of models ranging from 135 million to 4 billion parameters, trained on a dataset comprising 500 billion tokens. The performance of SODA is evaluated against the scaling predictions established earlier in the study, as well as against existing models in the field. The versatility of SODA is highlighted through its application in various audio and text tasks, including a fine-tuning example for voice-preserving speech-to-speech translation, which showcases the model's capability to maintain a unified architecture across different modalities.
The significance of this research lies in its potential to redefine audio modeling paradigms by providing a robust framework that integrates audio generation with text and semantic understanding. The insights gained from this study not only advance the state of the art in audio language models but also pave the way for future explorations in cross-modal AI applications, enhancing the interaction between audio and textual data in innovative ways.
The paper introduces SAW-Bench (Situated Awareness in the Real World), a novel benchmark designed to evaluate egocentric situated awareness in multimodal foundation models (MFMs). Situated awareness is defined as the ability to relate oneself to the surrounding environment and to reason about possible actions based on that context. Traditional benchmarks have primarily focused on environment-centric spatial relations, which assess relationships among objects within a scene, neglecting the crucial observer-centric relationships that depend on the agent's viewpoint, pose, and motion. This oversight presents a significant gap in the evaluation of models intended to understand human-like perception and interaction with the environment. To address this issue, the authors developed SAW-Bench, which consists of 786 self-recorded videos captured using Ray-Ban Meta (Gen 2) smart glasses, showcasing a variety of indoor and outdoor environments. Accompanying these videos are over 2,071 human-annotated question-answer pairs that are structured to probe a model's observer-centric understanding through six distinct awareness tasks. The comprehensive evaluation conducted reveals a substantial performance gap of 37.66% between human participants and the best-performing MFM, Gemini 3 Flash. This gap underscores the limitations of current models in achieving human-like situational awareness. Further analysis indicates that while these models can leverage partial geometric cues present in egocentric videos, they frequently struggle to infer coherent camera geometry, resulting in systematic errors in spatial reasoning. The authors argue that SAW-Bench serves as a critical benchmark for assessing situated spatial intelligence, emphasizing the need for models to progress beyond mere passive observation to a more profound understanding of physically grounded, observer-centric dynamics. This research not only highlights the deficiencies in existing models but also sets the stage for future advancements in the field of artificial intelligence, particularly in enhancing the situational awareness capabilities of MFMs.
The paper presents VETime, a novel framework for Time-Series Anomaly Detection (TSAD) that addresses the limitations of existing models by integrating both temporal and visual modalities. Traditional TSAD approaches often grapple with a trade-off between the granularity of pointwise anomaly localization and the broader contextual understanding necessary for effective anomaly detection. Specifically, 1D temporal models excel in pinpointing immediate anomalies but fail to capture the global context, while 2D vision-based models can identify overarching patterns yet struggle with precise temporal alignment and pointwise detection. VETime seeks to bridge this gap through innovative methodologies that enhance the detection capabilities across both dimensions.
The core of VETime lies in its Reversible Image Conversion and Patch-Level Temporal Alignment modules, which work in tandem to create a unified visual-temporal timeline. This integration allows the model to retain essential discriminative details while ensuring sensitivity to temporal variations, a crucial aspect for accurately identifying anomalies in time-series data. By establishing a shared timeline, VETime enables the model to leverage the strengths of both modalities effectively.
In addition to the alignment modules, the framework incorporates an Anomaly Window Contrastive Learning mechanism. This mechanism is designed to improve the model's ability to differentiate between normal and anomalous patterns by contrasting them within defined temporal windows. This contrastive approach enhances the model's learning process, allowing it to adaptively focus on the most relevant features for anomaly detection.
Moreover, VETime employs a Task-Adaptive Multi-Modal Fusion strategy, which dynamically integrates the complementary strengths of the temporal and visual data. This adaptability is particularly beneficial in zero-shot scenarios, where the model must generalize to detect anomalies in unseen data without prior training on those specific instances.
The experimental results presented in the paper demonstrate VETime's superior performance compared to state-of-the-art models, particularly in terms of localization precision and computational efficiency. The framework not only achieves higher accuracy in identifying anomalies but does so with reduced computational overhead, making it a promising solution for real-world applications where resources may be limited.
Overall, VETime represents a significant advancement in the field of TSAD, providing a robust framework that effectively combines temporal and visual information to enhance anomaly detection capabilities. The implications of this research extend beyond academic interest, offering practical solutions for industries reliant on time-series data analysis, such as finance, healthcare, and IoT systems. The availability of the code on GitHub further facilitates the adoption and adaptation of this framework by researchers and practitioners alike.
The challenge of automated unit test generation for C programming stems from the inherent semantic gap between the high-level intent of a programmer and the low-level syntactic requirements imposed by C's pointer arithmetic and manual memory management. This paper introduces SPARC, a neuro-symbolic framework designed to address these challenges by enhancing the capabilities of Large Language Models (LLMs) in generating meaningful unit tests. The traditional approach of intent-to-code synthesis often leads to issues such as non-compilable tests, hallucinated function signatures, and low coverage metrics due to the leap-to-code failure mode, where LLMs prematurely generate code without adequate grounding in the underlying program structure and semantics.
SPARC operates through a four-stage process:
1. **Control Flow Graph (CFG) Analysis**: This initial stage involves analyzing the program's control flow to understand the logical structure and paths that can be taken during execution.
2. **Operation Map**: This component grounds the reasoning of LLMs in validated utility helpers, ensuring that the generated tests are relevant and applicable to the program's context.
3. **Path-targeted Test Synthesis**: In this stage, the framework synthesizes tests that are specifically targeted at identified paths within the CFG, enhancing the likelihood of meaningful test coverage.
4. **Iterative Self-correction Validation Loop**: Finally, SPARC employs a validation loop that utilizes feedback from both the compiler and runtime to iteratively refine the generated tests, correcting any issues that may arise during initial test synthesis.
The evaluation of SPARC was conducted on 59 real-world and algorithmic subjects, demonstrating significant improvements over traditional prompt generation baselines. Specifically, SPARC achieved a 31.36% increase in line coverage, a 26.01% increase in branch coverage, and a 20.78% improvement in mutation score compared to baseline methods. Notably, SPARC's performance was comparable to or exceeded that of the established symbolic execution tool KLEE, particularly on more complex subjects. Furthermore, the framework retained 94.3% of the tests through its iterative repair process, indicating a robust approach to test generation. The generated code also received higher ratings for readability and maintainability from developers, suggesting that SPARC not only improves test coverage but also enhances the quality of the generated tests.
In conclusion, SPARC represents a significant advancement in the field of automated unit testing for legacy C codebases. By effectively aligning LLM reasoning with the program structure, SPARC provides a scalable solution that addresses the pressing need for improved testing methodologies in the software development industry, particularly for legacy systems that continue to be critical in various applications.
The 2026 AI Funding Landscape: A Deep Dive into High-Stakes Venture
The Era of "Realized ROI"
As we progress through 2026, the AI startup funding market has entered its second complete cycle. In 2023-2024, the hype was foundational—investors poured billions into infrastructure and large language models (LLMs). Today, the capital is moving "up the stack." Venture firms are no longer satisfied with token-per-second metrics or model benchmarks; they are hunting for companies with high retention, defensible data moats, and vertical-specific workflows.
The term "AI-first" has evolved from a differentiator into a baseline. Startups that are successfully raising Series B and C rounds in 2026 are those that have replaced expensive human-in-the-loop processes with autonomous agentic systems. This shift in AI startup news highlights a fundamental change in how startups are valued: valuation multiples are now increasingly tied to productivity gains per employee.
Vertical Superiority
General purpose AI assistants are struggling to find sustainable moats. In contrast, AI investment news is dominated by "Vertical Sovereignty"—where a startup builds a highly specialized model for a specific industry like civil engineering, advanced pharmacology, or luxury logistics.
The Talent War
A significant portion of AI startup funding is being earmarked for specialized talent acquisitions. With the median salary for a Senior AI Research Scientist exceeding $800k in major hubs, funding rounds are becoming necessary wars of attrition to secure the intellectual property contained within human experts.
Acquisition as the New "Exit"
The public markets (IPOs) have remained selective, leading to a surge in strategic acquisitions by Big Tech. In our AI startup news feed, we've noted a pattern of "Acqui-hires" moving away from individual experts to "Acqui-stacks"—where a conglomerate buys a startup solely for its proprietary data pipeline or its specialized reinforcement learning framework.
For founders, the goal in 2026 is often creating a "Target" rather than an "Empire." A startup that can solve context-window management 10% more efficiently than the baseline becomes an immediate acquisition target for the hyperscalers.
What VCs are Looking for in H2 2026
We've interviewed lead partners at seven top-tier AI firms to distill the criteria for a successful raise in late 2026. The consensus? Focus on these three metrics:
01
Cost-to-Reason Ratio
How much compute are you spending to reach a valuable decision? The most profitable startups are those optimizing for "Sparse Intelligence."
02
Data Recirculation
Does your system get smarter every time a user interacts with it without manual retraining? Flywheel effects are mandatory for Series A rounds.
03
Agency and Autonomy
Can your AI complete a 10-step workflow without human intervention? The era of "Copilots" is fading; the era of "Delegates" is here.