Abstract

AI agents are rapidly transforming the software engineering landscape, supported by a wave of new frameworks. Our review of the 10 most representative general-purpose agentic AI frameworks reveals a maturing but fragmented landscape. While some frameworks provide enterprise-ready capabilities, most remain immature, with limited memory models, basic reasoning, and uneven ecosystem support. Significant improvements are needed before these frameworks can enable reliable and large-scale adoption in practice

Abstract

AI-enabled systems are subjected to various types of runtime uncertainties, ranging from dynamic workloads, resource requirements, model drift, etc. These uncertainties have a big impact on the overall Quality of Service (QoS). This is particularly true in the case of Language Model (LM) enabled systems where the autoregressive nature of token generation introduces variability in latency, energy usage and response quality. These systems, powered by LLMs, are either resource-intensive (if run on-prem) or raise privacy/cost concerns (if leveraged using APIs). While deploying a Small Language Model (SLM) can be resource-efficient, it often falls short in addressing the diversity and scale of real-world requirements. To this, we argue that, rather than relying on any one SLM, leveraging a coordinated fleet of SLMs, each with specialized strengths can enable systems to dynamically adapt to shifting contexts and workload patterns. However, realizing the full potential of such an approach demands intelligent orchestration and continuous adaptation. To this end, we introduce CALM , a self-adaptive orchestration mechanism based on MAPE-K. Our approach continuously monitors user queries, analyzes the QoS metrics of the SLMs, identifies the optimal SLM to be used, routes the query to the identified SLM and further to enhance the effectiveness and efficiency, leverages caching and scheduling to decide the SLMs to be kept in memory. Our evaluation shows that CALM reduces latency by approximately 40% and energy consumption by 50%, while preserving domain-specific task performance when compared to single-LLM baselines.

Abstract

The automorphism group of a code is the group of permutations that map a code to itself. Berman codes are a class of binary linear codes characterized by two integer parameters n≥2 and m≥1, and this class includes the Reed-Muller codes as well. The class of Berman codes and their duals were recently shown to achieve the capacity of the binary erasure channel. A number of abelian codes that arise from the intersection and subspace sums of Berman and Dual Berman codes were also identified recently, for odd n≥3. A subclass of these abelian codes was shown to have good short block-length performance for AWGN channels, with efficient decoding algorithms. In this work, we identify the exact automorphism group for Berman codes and their duals. Further, we find the exact automorphism group for the above mentioned abelian codes, when n≥5. In the case of such abelian codes with n=3, we present partial characterizations of the automorphism groups for a large collection of parameter choices, and complete characterizations for a few.

Abstract

Text-to-speech systems are continuously improving, making the evaluation of perceptual quality increasingly challenging. Existing MOS prediction methods reduce speech quality to a single scalar value, without capturing how different listeners perceive audio across dimensions like acoustic fidelity, prosodic naturalness, phonetic intelligibility, and semantic coherence. We present MDMOS, a multi-dimensional framework where audio is processed through four complementary perceptual expert heads, each combining layer-grouped WavLM-Large representations with hand-crafted signal-processing features. An ordinal KL loss provides each head with independent supervision, preventing head collapse. A quality-adaptive fusion module dynamically weights the expert scores, while orthogonality regularisation and variance-weighted training account for rater uncertainty. Experiments on SOMOS and BVCC demonstrate strong MOS prediction accuracy, achieving system-level SRCC of 0.925 and competitive performance with 0.944 respectively.

Abstract

Text-to-Speech (TTS) technology has achieved remarkable progress with the advent of deep learning, low-resource indigenous languages still remain at the margins of this advancement. In the Indian context, indigenous languages such as Bhili, Gondi, Santhali, and Mundari, spoken by millions of people, are critically underrepresented. This paper addresses this gap by developing neural TTS systems for these languages through the creation of a comprehensive speech-text paired corpus combining translated text from resource-rich languages and original compositions in indigenous languages. We trained and fine-tuned Glow-TTS models using baseline and transfer learning approaches. Subjective and objective evaluation metrics show that Bhili achieved the highest Mean Opinion Score (MOS) when fine-tuned from a Hindi-pretrained model. Gondi, Santhali, and partly Mundari performed better when trained from scratch, underscoring the need for language-specific modeling in structurally diverse languages.

Abstract

This paper presents a novel framework for fine-grained classification of emphasis in spontaneous speech by moving beyond traditional binary stress detection, which rely on handcrafted features like MFCCs, pitch and formant frequencies. By utilizing a lecture mode multi-speaker speech corpus annotated with two stress levels: `High' and `Low', we propose a methodology that utilises a fusion of acoustic embeddings along with contextual linguistic features from pretrained Self-Supervised models. This feature set is processed by a Temporal CNN-BiLSTM network to learn the intricate temporal patterns of stress. The results demonstrate that the proposed approach outperforms the traditional baselines with an accuracy of 98% on the Stressed vs Unstressed task and 96% on the Low vs High-stress task. The framework achieved an accuracy of 95% on the multi-class (Unstressed vs Low vs High) task, validating its ability to perform real-world nuanced detection of emphasis at varied levels and speakers.

Abstract

Masonry structures, both contemporary and historic, often require strengthening to meet modern safety demands and resist seismic actions. Conventional retrofitting techniques are frequently associated with drawbacks, including increased mass, compatibility issues with substrates, and intrusive application procedures. Composite-based solutions have therefore gained prominence as effective alternatives. Among these, fibre-reinforced polymers (FRPs) were initially adopted due to their superior mechanical properties and ease of handling; however, their dependence on epoxy resins introduces limitations related to durability, cost, and performance under adverse environmental conditions. These limitations have driven the development of textile-reinforced mortar (TRM) systems as compatible and sustainable strengthening solutions. This study presents a critical synthesis of TRM-based strengthening for masonry structures, integrating experimental evidence, field applications, and existing design provisions. The mechanical behaviour, bond-controlled interaction at the TRM-masonry interface, and durability characteristics of TRM systems are examined. Existing design provisions from international guidelines are systematically compared and critically reinterpreted within a limit-state framework to evaluate the in-plane and out-of-plane behaviour of TRM-strengthened walls. The framework explicitly accounts for failure mechanisms, bond behaviour, and effective strain limits governing reinforcement efficiency. Key findings indicate that the effectiveness of TRM strengthening is governed by interface behaviour, reinforcement configuration, and anchorage conditions, rather than textile tensile capacity alone. Field applications are critically analyzed to extract design-relevant insights, demonstrating improved strength, deformation capacity, and crack control while maintaining compatibility with masonry substrates. Emerging natural fibre-based TRM systems are also identified as promising alternatives for sustainable strengthening. The study provides design-oriented insights for improving the reliability and applicability of TRM retrofitting strategies.

Abstract

We present a multi-stage style change detection system for the PAN 2026 Multi-Author Writing Style Analysis
shared task, which requires identifying sentence-level authorship transitions within documents. The proposed
pipeline combines heterogeneous transformer encoders, ensemble fusion, and document-level sequential re-
finement. Three sentence encoders (DeBERTa-v3-base, RoBERTa-large, and MPNet-base) are first fine-tuned
using the CoSENT ranking objective to learn style-sensitive sentence representations. Their predictions are
subsequently combined through difficulty-specific MLP boundary classifiers trained with Focal BCE loss and
R-Drop regularization to address severe class imbalance. The resulting probabilities are fused using Optuna-
based ensemble optimization and further refined using a bidirectional LSTM operating over complete document
boundary sequences. Compared to earlier paragraph-level formulations of the task, the sentence-level granularity
of PAN 2026 produces substantially longer transition sequences per document, making sequential modeling
considerably more informative. On the official hidden test set evaluated through TIRA, the submitted system
achieves macro-F1 scores of 0.9930, 0.7218, and 0.7573 on the easy, medium, and hard subsets respectively,
while maintaining close agreement with validation performance across all difficulty levels.

Abstract

utomatic pronunciation assessment (APA) of L2 learners is
critical in computer-assisted language learning. Existing ap-
proaches typically rely on canonical phonemes, its aligned
boundaries and large labeled data, which is costly and cumber-
some. To address these, we propose ALFreeD, a segmentation-
free APA framework guided by teacher's speech without the
need of canonical phonemes using minimal labeled data. In
this, pronunciation quality is assessed by quantifying deviation
between learners' and teacher's utterances. First, frame-level
embeddings are extracted using HuBERT, and deviation scores
are computed between them after obtaining their temporal cor-
respondence with DTW. Next, utterance level pronunciation de-
viation vector is obtained by suppressing variabilities other than
pronunciation quality with i-vector framework and is passed to
a MLP for predicting quality. Experiments on SpeechOcean762
demonstrate competitive performance with recent E2E methods
using minimal labeled data

Abstract

Text-to-image diffusion models can generate violent, harmful, or
weapon-related content, creating a difficult trade-off between safety
alignment and visual quality. Existing preference-based approaches
typically optimize a weighted combination of safety and quality
objectives, which can become unstable when the two objectives conflict.

We present BiDPO, a bilevel direct preference optimization
framework for safety-first diffusion alignment with two disjoint LoRA
adapters. A lower-level adapter is trained first on safety-ranked image
pairs, and an upper-level adapter is then trained on quality-ranked
pairs constructed from safety-gated candidates. This separation reduces
direct interference between safety and quality updates while keeping the
base model frozen.

We also study a failure mode that we call upper-level safety
drift}, where quality-focused updates reduce the safety gains obtained
at the lower level. To mitigate this effect, we evaluate several
controls, including chained references, gap-weighted pair mining, and
adaptive detector weighting. On SD1.5, the best BiDPO variant achieves
a lower unsafe-prompt safety-risk score than our joint baseline
(0.235 vs. 0.248) while slightly improving quality under our
evaluation pipeline, and it also outperforms the strongest comparison
methods we tested. Across controlled experiments, BiDPO shows a
consistent safety advantage over matched joint training. These results
suggest that a bilevel training structure can provide a stronger
safety-quality trade-off than a single weighted objective.