AI research papers

Recent AI research papers, abstracts, figures and original publication sources.

AI & MACHINE LEARNING

Latest AI research

Lukáš Jochymek, Markéta Vašinková, Petr Gajdoš

Published implementations of the Poincaré convolution in Hyperbolic Neural Networks++ (HNN++) map every overlapping image patch to the Poincaré ball before evaluating its score. We show that the clipped score depends on each patch only through two statistics, its squared norm and its inner products with the filters, which a native convolution and a single-channel box filter compute directly. The factorized layer therefore evaluates the specified score exactly, without constructing ball-valued patches, and its single-precision outputs and gradients agree with double-precision evaluations at least as closely as those of the explicit reference. On an RTX A6000 under two versions of JAX, complete training updates of a Poincaré U-Net and a Poincaré ResNet-20 are 1.5 to 1.8 times faster than the faster of two explicit implementations that map patches to the ball (1.2 to 1.4 times against an explicit implementation that skips this mapping); on a V100, the gain over the gather-based one persists as the ResNet-20 is widened eightfold. In 15 paired training runs per task, updates are 2.0 and 1.8 times faster than a reference that follows the published execution order, with no consistent change in final segmentation or classification quality, and peak memory falls by 38 to 55% in a U-Net scaling study. Replacing the convolutions of the official PyTorch Hyperbolic U-Net makes its training step 1.4 times faster with half the memory. Evaluating an existing layer from its patch statistics thus makes Poincaré convolutional networks cheaper to train in the tested models and software stacks. Abstract

Published implementations of the Poincaré convolution in Hyperbolic Neural Networks++ (HNN++) map every overlapping image patch to the Poincaré ball before evaluating its score. We show that the clipped score depends on each patch only through two statistics, its squared norm and its inner products with the filters, which a native convolution and a single-channel box filter compute directly. The factorized layer therefore evaluates the specified score exactly, without constructing ball-valued patches, and its single-precision outputs and gradients agree with double-precision evaluations at least as closely as those of the explicit reference. On an RTX A6000 under two versions of JAX, complete training updates of a Poincaré U-Net and a Poincaré ResNet-20 are 1.5 to 1.8 times faster than the faster of two explicit implementations that map patches to the ball (1.2 to 1.4 times against an explicit implementation that skips this mapping); on a V100, the gain over the gather-based one persists as the ResNet-20 is widened eightfold. In 15 paired training runs per task, updates are 2.0 and 1.8 times faster than a reference that follows the published execution order, with no consistent change in final segmentation or classification quality, and peak memory falls by 38 to 55% in a U-Net scaling study. Replacing the convolutions of the official PyTorch Hyperbolic U-Net makes its training step 1.4 times faster with half the memory. Evaluating an existing layer from its patch statistics thus makes Poincaré convolutional networks cheaper to train in the tested models and software stacks.

PDF ↗

Ziyuan Li, Uwe Jaekel, Babette Dellen

Understanding nonlinear feature interactions is crucial in science and engineering, yet standard multilayer perceptrons (MLPs) often capture such interactions only implicitly, leading to entangled representations that can impair robustness and interpretability. We investigate complex-valued product-unit residual networks (CV-PURe) that integrate multiplicative product units with residual connections to explicitly model cross-feature couplings while stabilizing optimization. We conduct a systematic evaluation on an interaction-driven synthetic benchmark and real-world datasets and evaluate prediction together with input-level interaction recovery under a reproducible protocol. Across five data splits and five training seeds, CV-PURe achieves the lowest mean RMSE on four of six scientific benchmarks, reducing Friedman error by 53.9% relative to MLP. Its Friedman interaction maps assign 60.7% of total strength to the known interacting pair, compared with 31.0% for MLP. On a synthetic task containing third- and fourth-order products and 20 distractors, CV-PURe reduces RMSE from 0.1376 to 0.0797 relative to a complex-valued MLP and recovers all nine true pairwise projections in every training/sampling repetition. Bike Sharing experiments connect improved prediction to the contribution of temporal, calendar, and seasonal context through paired ablations of both PURe variants. Parkinsons Telemonitoring extends the context analysis to repeated measurements. Noise and learning-curve evaluations demonstrate sustained predictive gains under perturbed inputs and limited supervision, while inference measurements quantify computational cost. This article extends our ICCS 2026 paper, Modeling Nonlinear Feature Interactions with Product-Unit Residual Networks (doi:10.1007/978-3-032-29921-5_6). Abstract

Understanding nonlinear feature interactions is crucial in science and engineering, yet standard multilayer perceptrons (MLPs) often capture such interactions only implicitly, leading to entangled representations that can impair robustness and interpretability. We investigate complex-valued product-unit residual networks (CV-PURe) that integrate multiplicative product units with residual connections to explicitly model cross-feature couplings while stabilizing optimization. We conduct a systematic evaluation on an interaction-driven synthetic benchmark and real-world datasets and evaluate prediction together with input-level interaction recovery under a reproducible protocol. Across five data splits and five training seeds, CV-PURe achieves the lowest mean RMSE on four of six scientific benchmarks, reducing Friedman error by 53.9% relative to MLP. Its Friedman interaction maps assign 60.7% of total strength to the known interacting pair, compared with 31.0% for MLP. On a synthetic task containing third- and fourth-order products and 20 distractors, CV-PURe reduces RMSE from 0.1376 to 0.0797 relative to a complex-valued MLP and recovers all nine true pairwise projections in every training/sampling repetition. Bike Sharing experiments connect improved prediction to the contribution of temporal, calendar, and seasonal context through paired ablations of both PURe variants. Parkinsons Telemonitoring extends the context analysis to repeated measurements. Noise and learning-curve evaluations demonstrate sustained predictive gains under perturbed inputs and limited supervision, while inference measurements quantify computational cost. This article extends our ICCS 2026 paper, Modeling Nonlinear Feature Interactions with Product-Unit Residual Networks (doi:10.1007/978-3-032-29921-5_6).

PDF ↗

Matheus Pereira, Jefersson Santos, Laurent Najman

Pixelwise classification of hyperspectral images (HSI) is a fundamental task in remote sensing, but its deployment in real-world applications is often limited by the scarcity of labeled data and the presence of unknown classes at inference time. In this work, we propose Few-shot Open-set hyperspectral recognition from Self-Supervised Learning training (FOSSeL), a unified framework for few-shot, open-set hyperspectral pixelwise classification that does not require supervised training. The proposed approach first leverages self-supervised learning to construct a discriminative feature space from unlabeled data. Then, given a small set of labeled seed pixels, class predictions are obtained through spatially-aware graph-based label propagation over superpixel representations, enabling the joint exploitation of spectral similarity and spatial consistency. To address the open-set scenario, we introduce simple yet effective strategies to identify pixels that do not belong to the known classes. We evaluate FOSSeL on three benchmark hyperspectral datasets under multiple few-shot settings. The results show that the proposed framework achieves competitive or superior performance compared to fully supervised methods while maintaining robustness in the presence of unknown classes. These findings highlight the potential of combining self-supervised feature learning with graph-based inference for practical HSI analysis in low-annotation and open-world scenarios. Abstract

Pixelwise classification of hyperspectral images (HSI) is a fundamental task in remote sensing, but its deployment in real-world applications is often limited by the scarcity of labeled data and the presence of unknown classes at inference time. In this work, we propose Few-shot Open-set hyperspectral recognition from Self-Supervised Learning training (FOSSeL), a unified framework for few-shot, open-set hyperspectral pixelwise classification that does not require supervised training. The proposed approach first leverages self-supervised learning to construct a discriminative feature space from unlabeled data. Then, given a small set of labeled seed pixels, class predictions are obtained through spatially-aware graph-based label propagation over superpixel representations, enabling the joint exploitation of spectral similarity and spatial consistency. To address the open-set scenario, we introduce simple yet effective strategies to identify pixels that do not belong to the known classes. We evaluate FOSSeL on three benchmark hyperspectral datasets under multiple few-shot settings. The results show that the proposed framework achieves competitive or superior performance compared to fully supervised methods while maintaining robustness in the presence of unknown classes. These findings highlight the potential of combining self-supervised feature learning with graph-based inference for practical HSI analysis in low-annotation and open-world scenarios.

PDF ↗

arXivcs.AI · cs.AR · cs.CY · cs.ET · cs.NE

The Transformer Paradigm: Dynamic Processing and a New Principle of Inference

Marco Giunti, Fabrizia Giulia Garavaglia

We argue that the Transformer revolution is not adequately understood as a merely quantitative expansion of deep learning, but as the emergence of a new paradigm based on a different principle of inference. We interpret the Transformer as a system in which token representations are transformed by parameterized transformations, or transforming concepts. During inference, some of these transformations are dynamically constructed from the current input: previously acquired transformations are first adapted and then applied. At the mechanical-structural level, this dynamic processing is realized through output-weight interconnections, whereby outputs of some networks determine the weights of others. The resulting mode of processing, which we call Sequence-level Interactive Dynamic Parallel Processing (SIDPP), provides a framework for explaining strong prompt sensitivity and opens research directions concerning interpretability, control, alternative AI architectures, neuromorphic systems, and possible biological realizations of dynamic processing. Abstract

We argue that the Transformer revolution is not adequately understood as a merely quantitative expansion of deep learning, but as the emergence of a new paradigm based on a different principle of inference. We interpret the Transformer as a system in which token representations are transformed by parameterized transformations, or transforming concepts. During inference, some of these transformations are dynamically constructed from the current input: previously acquired transformations are first adapted and then applied. At the mechanical-structural level, this dynamic processing is realized through output-weight interconnections, whereby outputs of some networks determine the weights of others. The resulting mode of processing, which we call Sequence-level Interactive Dynamic Parallel Processing (SIDPP), provides a framework for explaining strong prompt sensitivity and opens research directions concerning interpretability, control, alternative AI architectures, neuromorphic systems, and possible biological realizations of dynamic processing.

PDF ↗

Zelin He, Siyu Zhu, Jacob Helwig, Yujian Liu, Li An, Rohit Jain, et al.

Agentic reinforcement learning (RL) increasingly relies on a harness, the scaffold around the model that decides what it sees of its own interaction and what guidance it receives. Most existing approaches train under a hand-written harness that stays fixed throughout. However, an effective harness depends on the domain, the data and the model, and keeps changing as the model trains. We introduce ModHarness-RL, which recasts the harness as programs that are learned from data and evolve as the model trains with RL. ModHarness-RL records the interaction as structured State, a set of typed blocks, and two core components of the harness are programs over these blocks, a Compression program that assembles the accumulated history into the model's context and a Steering program that supplies guidance when behavior deviates. During RL training, candidate programs are proposed from recent trajectories, verified by execution, and replayed over past trajectories, so the harness tracks how the model's behavior shifts. We study CodeSearch in depth and further analyze ALFWorld and BIRD SQL, with models from 1.5B to 8B. The proposed Compression and Steering mechanisms improve held-out performance over the fixed-harness baseline in most settings. These results suggest that a harness for agentic RL can be learned from the task's data and evolved with the model. Abstract

Agentic reinforcement learning (RL) increasingly relies on a harness, the scaffold around the model that decides what it sees of its own interaction and what guidance it receives. Most existing approaches train under a hand-written harness that stays fixed throughout. However, an effective harness depends on the domain, the data and the model, and keeps changing as the model trains. We introduce ModHarness-RL, which recasts the harness as programs that are learned from data and evolve as the model trains with RL. ModHarness-RL records the interaction as structured State, a set of typed blocks, and two core components of the harness are programs over these blocks, a Compression program that assembles the accumulated history into the model's context and a Steering program that supplies guidance when behavior deviates. During RL training, candidate programs are proposed from recent trajectories, verified by execution, and replayed over past trajectories, so the harness tracks how the model's behavior shifts. We study CodeSearch in depth and further analyze ALFWorld and BIRD SQL, with models from 1.5B to 8B. The proposed Compression and Steering mechanisms improve held-out performance over the fixed-harness baseline in most settings. These results suggest that a harness for agentic RL can be learned from the task's data and evolved with the model.

PDF ↗

Dayou Du, Congjie He, Andrew Boutros, Luo Mai, Vaughn Betz, Jianyi Cheng

Wafer-scale accelerators offer massive compute capacity through hundreds of thousands of processing elements (PEs) interconnected across a single wafer, but realizing this potential requires efficient application mapping. Their scale makes programming these systems and mapping applications onto them challenging. Existing approaches rely on template-based mappings to maintain productivity, but these templates restrict the mapping design space. In particular, they impose fixed geometric layouts over the PE grid, often leaving compute and memory resources underutilized. In this work, we present WaferTetris, a system that removes these geometric restrictions through freeform mapping. Its Tetris-like strategy flexibly packs computation across the wafer, improving PE utilization and usable memory capacity while preserving PE-task code. We develop a recurrence-based throughput model that captures branch-and-join dependencies to guide spatial parallelism, and an architecture model with per-color gates that enable existing FPGA place-and-route algorithms to enforce WSE routing constraints. We evaluate WaferTetris on Cerebras WSE-3 using LLM decoding and scientific stencil benchmarks. For LLM decoding, it achieves up to 4.42×4.42\times the system throughput of a template-based automatic mapping baseline on the same fabric size. Compared with the manually optimized WaferLLM, WaferTetris supports 1.8×1.8\times the context length for LLaMA3.1-8B (80k versus 44k), while delivering equal or higher per-user throughput at matched context lengths, up to 1.10×1.10\times. Abstract

Wafer-scale accelerators offer massive compute capacity through hundreds of thousands of processing elements (PEs) interconnected across a single wafer, but realizing this potential requires efficient application mapping. Their scale makes programming these systems and mapping applications onto them challenging. Existing approaches rely on template-based mappings to maintain productivity, but these templates restrict the mapping design space. In particular, they impose fixed geometric layouts over the PE grid, often leaving compute and memory resources underutilized. In this work, we present WaferTetris, a system that removes these geometric restrictions through freeform mapping. Its Tetris-like strategy flexibly packs computation across the wafer, improving PE utilization and usable memory capacity while preserving PE-task code. We develop a recurrence-based throughput model that captures branch-and-join dependencies to guide spatial parallelism, and an architecture model with per-color gates that enable existing FPGA place-and-route algorithms to enforce WSE routing constraints. We evaluate WaferTetris on Cerebras WSE-3 using LLM decoding and scientific stencil benchmarks. For LLM decoding, it achieves up to 4.42×4.42\times the system throughput of a template-based automatic mapping baseline on the same fabric size. Compared with the manually optimized WaferLLM, WaferTetris supports 1.8×1.8\times the context length for LLaMA3.1-8B (80k versus 44k), while delivering equal or higher per-user throughput at matched context lengths, up to 1.10×1.10\times.

PDF ↗

Mohamed Aghzal, Mor Geva, Gregory J. Stein, Ziyu Yao

Mechanistic Interpretability (MI) has made significant progress in explaining the internal computations of Large Language Models (LLMs); however, these advances have yet to demonstrate comparable success in explaining the decision-making processes of LLM-based agents. This paper examines the challenges MI tools face when applied to realistic agentic failures. We focus on one such failure: web agents hallucinating actions that are not available on a website. We apply a series of mechanistic analyzes to Qwen2.5-7B-Instruct as it interacts with live websites. We find that a linear probe trained on the last prompt token can predict hallucinations before generation begins (0.80 AUC) and steering along a probe direction corrects 85% of hallucinations, but disrupts 7% of valid actions. However, these results only provide a partial explanation of why the agent hallucinates. The probe direction is sensitive to prompt formatting, losing much of its steering effect after a minor change to how actions are represented. Additionally, circuit identification can be difficult as the computation is distributed across multiple components and positions, and the information the agent relies on varies across examples. We discuss how these observations challenge the assumptions underlying standard MI methods for circuit discovery and feature identification, and outline future directions for making these methods better suited to explain failures in LLM-based agents. Abstract

Mechanistic Interpretability (MI) has made significant progress in explaining the internal computations of Large Language Models (LLMs); however, these advances have yet to demonstrate comparable success in explaining the decision-making processes of LLM-based agents. This paper examines the challenges MI tools face when applied to realistic agentic failures. We focus on one such failure: web agents hallucinating actions that are not available on a website. We apply a series of mechanistic analyzes to Qwen2.5-7B-Instruct as it interacts with live websites. We find that a linear probe trained on the last prompt token can predict hallucinations before generation begins (0.80 AUC) and steering along a probe direction corrects 85% of hallucinations, but disrupts 7% of valid actions. However, these results only provide a partial explanation of why the agent hallucinates. The probe direction is sensitive to prompt formatting, losing much of its steering effect after a minor change to how actions are represented. Additionally, circuit identification can be difficult as the computation is distributed across multiple components and positions, and the information the agent relies on varies across examples. We discuss how these observations challenge the assumptions underlying standard MI methods for circuit discovery and feature identification, and outline future directions for making these methods better suited to explain failures in LLM-based agents.

PDF ↗

Mohammed Efaz, Udo Bub, Itilekha Podder

This paper introduces MeReader, a privacy-preserving, offline and progress-aware AI assistant designed to function as a contextual companion for narrative reading on consumer hardware. Unlike conventional Large Language Model (LLM)-based systems that process documents/books online in an all-texts-at-once type of way, thereby risking narrative spoilers, MeReader confines both retrieval and generation to the reader's distinct position in the text. The system integrates a hybrid retrieval architecture, combining vector similarity, BM25, and summary embeddings, to ensure high fidelity to the reader's current context. We evaluate the system across classical English books using both quantitative comprehension metrics and qualitative LLM-based independent judging. The results identify a quality-latency frontier, showing that mid-sized local models (2b-4b) can achieve comprehension parity with larger counterparts (5b+) while maintaining acceptable response quality. Importantly, the proposed progress-aware boundary mechanism is shown to prevent future content leakage without compromising retrieval quality and correctness. These findings suggest that a localized, privacy-centric approach can augment, rather than replace, the digital reading experience, offering a good alternative to cloud-dependent services. Therein, there remains a lack of amalgamate systems which combine progress-aware, privacy-preserving, and narrative eBook reading with AI support. It is in this very gap that the development of MeReader came to be; synthesizing said strands into an offline, book-constrained assistant focused on enhancing comprehension without compromising narrative continuity or user privacy. Abstract

This paper introduces MeReader, a privacy-preserving, offline and progress-aware AI assistant designed to function as a contextual companion for narrative reading on consumer hardware. Unlike conventional Large Language Model (LLM)-based systems that process documents/books online in an all-texts-at-once type of way, thereby risking narrative spoilers, MeReader confines both retrieval and generation to the reader's distinct position in the text. The system integrates a hybrid retrieval architecture, combining vector similarity, BM25, and summary embeddings, to ensure high fidelity to the reader's current context. We evaluate the system across classical English books using both quantitative comprehension metrics and qualitative LLM-based independent judging. The results identify a quality-latency frontier, showing that mid-sized local models (2b-4b) can achieve comprehension parity with larger counterparts (5b+) while maintaining acceptable response quality. Importantly, the proposed progress-aware boundary mechanism is shown to prevent future content leakage without compromising retrieval quality and correctness. These findings suggest that a localized, privacy-centric approach can augment, rather than replace, the digital reading experience, offering a good alternative to cloud-dependent services. Therein, there remains a lack of amalgamate systems which combine progress-aware, privacy-preserving, and narrative eBook reading with AI support. It is in this very gap that the development of MeReader came to be; synthesizing said strands into an offline, book-constrained assistant focused on enhancing comprehension without compromising narrative continuity or user privacy.

PDF ↗

Daniel Reichenpfader, Kwaku Appiagyei, Manovah Nadarajah, Ahmed Reda Bahr, Irina Filchenko, Fabio Dennstädt

Common data elements (CDEs) support clinical data comparability but are usually defined manually. We investigated whether large language model agents could derive a CDE schema from sarcoma case report abstracts and apply it at scale. An agent proposed 72,909 candidate CDEs from 24,588 PubMed abstracts; agentic consolidation yielded 1,570 CDEs. We retained 53 supported by at least 50 abstracts and attempted evidence-linked extraction across the corpus. Processing succeeded for all retained CDEs in 79.2% of abstracts, including responses indicating non-mention. Two human reviewers approved 83.0% of retained CDEs for relevance, 67.9% for clarity, and 56.6% for correctness. In 100 sampled abstracts, exact agreement with the value extracted by the LLM was 94.9% among 1,116 pairs with a non-empty answer (abstract-level bootstrap 95% CI, 93.0-96.4%). Errors concentrated in anatomical and temporal fields. These findings support agent-assisted schema drafting followed by expert curation before large-scale extraction. Abstract

Common data elements (CDEs) support clinical data comparability but are usually defined manually. We investigated whether large language model agents could derive a CDE schema from sarcoma case report abstracts and apply it at scale. An agent proposed 72,909 candidate CDEs from 24,588 PubMed abstracts; agentic consolidation yielded 1,570 CDEs. We retained 53 supported by at least 50 abstracts and attempted evidence-linked extraction across the corpus. Processing succeeded for all retained CDEs in 79.2% of abstracts, including responses indicating non-mention. Two human reviewers approved 83.0% of retained CDEs for relevance, 67.9% for clarity, and 56.6% for correctness. In 100 sampled abstracts, exact agreement with the value extracted by the LLM was 94.9% among 1,116 pairs with a non-empty answer (abstract-level bootstrap 95% CI, 93.0-96.4%). Errors concentrated in anatomical and temporal fields. These findings support agent-assisted schema drafting followed by expert curation before large-scale extraction.

PDF ↗

arXivcs.CL · cs.AI · cs.LG

Interactive Program Induction with Language Model Priors

Robin Shing Moon Chan, Ryan Cotterell

Program induction refers to the process in which a learner aims to recover a program from a few given input-output examples. This paradigm broadens programming access by letting users demonstrate the rule-based behavior they want to apply to unseen data, instead of having to implement it themselves. However, recovering a program from finite input-output examples is fundamentally ambiguous, as a finite set of input-output examples is theoretically consistent with infinitely many programs whose output diverges on unseen inputs. We view this problem from a Bayesian perspective, where, given a prior p0p^0 over programs, identifying the target program means concentrating posterior probability mass on it by observing additional input-output examples. In the interactive program induction setting, the learner chooses maximally informative inputs and queries an oracle, such as the user, for their outputs. We derive a greedy query-selection algorithm and show that it terminates after O(1γlog⁡1/p0(f⋆))\mathcal{O}(\frac{1}γ\log 1/p^0(f^\star)) informative oracle queries, and returns the target program f⋆f^\star up to approximate equivalence when no other class outweighs it under the prior. The algorithm uses a language model as the prior under a PAC-style relaxation of program equivalence, where the query-selection objective is estimated using importance sampling. On the CodeARC benchmark, the algorithm improves accuracy over the LM only agent by up to 22 percentage points while observing up to 10.5 fewer input-output examples. Abstract

Program induction refers to the process in which a learner aims to recover a program from a few given input-output examples. This paradigm broadens programming access by letting users demonstrate the rule-based behavior they want to apply to unseen data, instead of having to implement it themselves. However, recovering a program from finite input-output examples is fundamentally ambiguous, as a finite set of input-output examples is theoretically consistent with infinitely many programs whose output diverges on unseen inputs. We view this problem from a Bayesian perspective, where, given a prior p0p^0 over programs, identifying the target program means concentrating posterior probability mass on it by observing additional input-output examples. In the interactive program induction setting, the learner chooses maximally informative inputs and queries an oracle, such as the user, for their outputs. We derive a greedy query-selection algorithm and show that it terminates after O(1γlog⁡1/p0(f⋆))\mathcal{O}(\frac{1}γ\log 1/p^0(f^\star)) informative oracle queries, and returns the target program f⋆f^\star up to approximate equivalence when no other class outweighs it under the prior. The algorithm uses a language model as the prior under a PAC-style relaxation of program equivalence, where the query-selection objective is estimated using importance sampling. On the CodeARC benchmark, the algorithm improves accuracy over the LM only agent by up to 22 percentage points while observing up to 10.5 fewer input-output examples.

PDF ↗

Models

Loading
EVIDENCE / COMPARISON

Benchmarks

Explore every benchmark

Loading

These charts contain model-linked results. The charts above also include unlinked source profiles. Compare scores within each evaluation. Model quality, usage, transparency and system performance measure different things.

Install TensorTrace