Singapore

The paper explores how video models trained for classification tasks represent nuanced, hidden semantic information that may not affect the final outcome, a key challenge for Trustworthy AI models. Through Explainable and Interpretable AI methods, specifically mechanistic interpretability techniques, the internal circuit responsible for representing the action&#39;s outcome is reverse-engineered in a pre-trained video vision transformer, revealing that the &quot;Success vs Failure&quot; signal is computed through a distinct amplification cascade. While there are low-level differences observed from layer 0, the abstract and semantic representation of the outcome is progressively amplified from layers 5 through 11. Causal analysis, primarily using activation patching supported by ablation results, reveals a clear division of labor: Attention Heads act as &quot;evidence gatherers&quot;, providing necessary low-level information for partial signal recovery, while MLP Blocks function as robust &quot;concept composers&quot;, each of which is sufficient to generate the entire &quot;success&quot; signal. This distributed and redundant circuit in the model&#39;s internals explains its resilience to simple ablations, demonstrating a core computational pattern for processing human-action outcomes. Crucially, the existence of this sophisticated circuit for representing complex outcomes, even within a model trained only for simple classification, highlights the potential for models to develop forms of &#39;hidden knowledge&#39; beyond their explicit task, underscoring the need for mechanistic oversight for building genuinely Explainable and Trustworthy AI systems intended for deployment.

AAAI 2026

Attention Gathers, MLPs Compose: A Causal Analysis of an Action-Outcome Circuit in VideoViT

and transparency; adversarial attacks & robustness; safety

robustness & trustworthiness

explainability

interpretability

We are pleased to announce the Thirty-Eighth AAAI Conference on Artificial Intelligence (AAAI-24), which will be held in Vancouver, British Columbia at the Vancouver Convention Centre – West Building from 20-27 February, 2024.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-24 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.

We expect for AAAI-24 to be an in-person conference – one author of all accepted papers will be expected to present work in person unless there are exceptional circumstances that prevent this.

In order to access the AAAI-24 event page you need to register [here](https://aaai.org/aaai-conference/registration/)

AAAI 2024

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. 

The paper explores how video models trained for classification tasks represent nuanced, hidden semantic information that may not affect the final outcome, a key challenge for Trustworthy AI models. Through Explainable and Interpretable AI methods, specifically mechanistic interpretability techniques, the internal circuit responsible for representing the action's outcome is reverse-engineered in a pre-trained video vision transformer, revealing that the "Success vs Failure" signal is computed through a distinct amplification cascade. While there are low-level differences observed from layer 0, the abstract and semantic representation of the outcome is progressively amplified from layers 5 through 11. Causal analysis, primarily using activation patching supported by ablation results, reveals a clear division of labor: Attention Heads act as "evidence gatherers", providing necessary low-level information for partial signal recovery, while MLP Blocks function as robust "concept composers", each of which is sufficient to generate the entire "success" signal. This distributed and redundant circuit in the model's internals explains its resilience to simple ablations, demonstrating a core computational pattern for processing human-action outcomes. Crucially, the existence of this sophisticated circuit for representing complex outcomes, even within a model trained only for simple classification, highlights the potential for models to develop forms of 'hidden knowledge' beyond their explicit task, underscoring the need for mechanistic oversight for building genuinely Explainable and Trustworthy AI systems intended for deployment.

workshop paper

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-26 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to AAAI 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://aaai.org/conference/aaai/aaai-26/registration/) first.

**Online Registration Form**: https://aaai.getregistered.net/conference-2026 

Registration Required

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

Semantic representations of rhythmic structures are important for AI-driven music generation and choreography. South Asian classical dance, such as Bharatanatyam, relies on intricate rhythms that guide choreography and improvisation. These rhythms are expressed through Nattuvangam, a vocal and percussive form that uses rhythmic syllables (Solkattus) and cymbal cues (Talam). Despite its pedagogical importance, Nattuvangam is rarely documented in digital form, which limits systematic study and teaching. We present the first curated dataset of Nattuvangam recordings that capture diverse Solkattu patterns and cyclic Talam structures. Each clip is analyzed using handcrafted and learned features, including onset envelopes, inter-onset intervals, tempograms, and Mel-spectrogram embeddings. These representations allow machine learning models to identify, cluster, and retrieve rhythmic motifs across performances. The dataset serves as a pedagogical tool and supports computational exploration of Solkattu patterns in relation to Talam, revealing the structural principles underlying Nattuvangam. This work establishes a foundation for studying Nattuvangam as both a standalone and performative art form, bridging cultural teaching with AI-based rhythm analysis in low-resource contexts.

Low-Resource Rhythm Learning of South Asian Beat Structures: Machine Learning Approaches to Nattuvangam

Sleep disorders, particularly insomnia, and mental health conditions affect a significant fraction of adults worldwide, posing serious mental and physical health risk. Music therapy offers promising, low-cost, and non-invasive treatment, but current approaches rely heavily on expert-curated playlists, limiting scalability and personalization. We propose a low-cost generative system leveraging recent advances in diffusion models to synthesize music for therapy. We focus on insomnia and curate a dataset of waveform sleep music to generate audio tailored to sleep. To ensure real-world feasibility, we optimize our system for training and use on a single GPU, balancing quality and efficiency through extensive ablation studies. We show through subjective human evaluations that our generated music matches or outperforms existing baselines in both perceived quality and relevance to sleep therapy, while using only a fraction of the computational cost.

A Novel Diffusion Model Based Approach for Sleep Therapeutic Music Generation

Neural codec language models have revolutionized speech synthesis but face significant challenges when adapted to music generation, particularly in achieving precise timbre con- trol while preserving melodic content. We introduce Neural Code Language Model for Controllable Timbre Transfer (NCLMCTT), a novel architecture that enables zero-shot instrument cloning through direct audio conditioning without explicit timbre learning. Our approach combines a 385M-parameter transformer for coarse musical structure modeling with a specialized upsampler for fine timbral detail, achieving flexible control through 1-5 second reference audio segments. We establish the first comprehensive benchmark dataset for controllable timbre transfer evaluation, comprising 62,500 high-fidelity samples across 50 synthesizer presets with ground truth targets. Extensive experiments demonstrate sub- stantial improvements over the TokenSynth baseline: 27.1% reduction in SI-SDR, 50.9% in Mel Distance, and 59.4% in STFT Distance, while maintaining strong melodic coher- ence (Chroma Similarity: 0.85). Our method achieves robust zero-shot generalization, with performance on unseen instrument presets matching that of seen presets. Ablation stud- ies confirm that extended reference audio duration (40.8% improvement), cross-attention mechanisms (11.9% improvement), and increased model capacity contribute meaningfully to overall performance. By separating melodic content from timbral characteristics and enabling implicit timbre control, NCLMCTT provides both immediate practical value for music creators and a methodological foundation for advancing controllable neural audio synthesis.

Neural Codec Language Model for Controllable Timbre Transfer in Music Synthesis

With recent advances in automatic speech recognition (ASR), large language models (LLMs), and text-to-speech (TTS) technologies, spoken dialogue systems (SDS) have become widely accessible. However, most existing SDS are limited to conventional spoken responses. We present SingingSDS, a cascaded SDS that responds through singing rather than speaking, fostering more affective, memorable, and pleasurable interactions in character-based roleplay and interactive entertainment scenarios. SingingSDS employs a modular ASR-LLM-SVS pipeline and supports a wide range of configurations across character personas, ASR and LLM backends, SVS models, melody sources, and voice profiles, tailored to different needs in terms of latency, quality, and musical style. SingingSDS is available as a plug-and-play web demo, featuring modular, open-source code that supports customization and extension. Demo: https://huggingface.co/spaces/espnet/SingingSDS. Code: https://github.com/SingingSDS/SingingSDS. Video: https://youtube.com/playlist?list=PLZpUJJbwp2WvtPBenG5D3h09qKIrt24ui&si=7CSLWAYWcfkTEdqe.

SingingSDS: A Singing-Capable Spoken Dialogue System for Conversational Roleplay Applications

This paper traces how “control” is socially coded in contemporary music technology, and offers an account of the circuit of control to bridge the conceptual gap between the sociopolitical and the technical. The analysis redirects attention from the controllability of intelligent systems to the social formations through which control is allocated and recognized. To frame this shift, the paper theorizes the gear economy—a lens set alongside the much frequented concept gig economy in sociology—to explain how AI-music tools (e.g., Suno, Udio) are legitimated through existing music gear markets. The claim is coalition rather than addition: AI tools fold into pre-existing markets that price and credential controllability. Formally, the paper adopts a triptych that deliberately echoes Deleuze’s influential “Postscript” in title and cadence: Historical, Logic, Programming to progress its analysis. Rather than constructing a new vocabulary of control in music technologies, this paper demonstrates that the parameters and regimes of control—technical, economic, and political—are themselves social indices. The aim is a design-oriented ethical intervention that speaks to technologists, musicians, and the general public, rendering these arrangements visible and reconfigurable and redirecting control toward humans so that AI music tools augment rather than shrink creative agency.

Postscript on the Musics of Control

We present Artificial Dancing Intelligence (ADI), an interactive neural music visualizer that is accessed through a web app, but performs inference entirely on local devices. Our approach enables anyone to visually express music, while leveraging the expressive, and sometimes unpredictable dynamics of self-organized systems. ADI uses an audio stream’s average energy (known as RMS) to modulate a neural cellular automata (NCA) to produce visual patterns that move and ’dance’ along with the audio stream in real time. Through the web interface, users can adjust the relationship between the music’s energy and the NCA system to create unique visual performances out of any music audio file. ADI achieves smooth, real-time responsiveness on modern consumer devices.

Artificial Dancing Intelligence: Neural Cellular Automata for Visual Performance of Music

Artists and rights holders face growing concerns about unauthorized use of their copyrighted works in training generative models. We introduce TS-RaMIA, a practical auditing framework enabling creators to test whether their symbolic music has been used without authorization. Unlike existing likelihood-based approaches that are confounded by piece length and density, TS-RaMIA exploits \emph{structural tokens}—bar lines, positions, and tempo markers encoding musical phrasing—through sample-level analysis and rigorous debiasing. Our method combines (i) length matching and conditional calibration to remove spurious confounders, (ii) tail-of-top- aggregation on structural tokens to amplify sparse memorization signals, and (iii) a lightweight meta-attacker fusing statistical cues via composer-stratified cross-validation. Evaluated on a 67M-parameter REMI Transformer trained on MAESTRO pieces, TS-RaMIA achieves AUC 0.826 and TPR@1%FPR 14.6%, while a debiased baseline drops to AUC 0.563. Cross-representation validation on NotaGen (ABC notation) yields comparable performance (AUC 0.73, TPR@1%FPR 8.9%), demonstrating transferability. We release our code at https://anonymous.4open.science/r/TS-RaMIA-63F0.

TS-RaMIA: Membership Inference Attacks for Symbolic Music Generation Models

Multimodal Large Language Models (MLLMs) claim “musical understanding,” yet most evaluations conflate listening with score reading. We benchmark three SOTA LLMs (Gemini 2.5 Pro, Gemini 2.5 Flash, and Qwen2.5-Omni) across three core music skills: Syncopation Scoring (rhythm perception), Transposition Detection (melody perception), and Chord Quality Identification (harmony perception). Moreover, we separate three sources of variability: (i) perceptual limitations (by contrasting audio recordings vs. symbolic MIDI inputs), (ii) exposure to prior examples (zero- vs. few-shot manipulations), and (iii) reasoning strategies (Standalone, Chain of Thought, LogicLM). For the latter we adapt LogicLM, a framework combining LLMs with symbolic solvers to perform structured reasoning. In LogicLM, LLMs act as perceptual formulators, generating strict, machinecheckable schemas (onset grids, interval sequences) that deterministic solvers execute with self-refinement. Our results reveal a clear perceptual gap: models perform near ceiling on MIDI but show substantial accuracy drops on audio. Reasoning and few-shot prompting o!er minimal gains. This is expected for MIDI, where performance reaches saturation, but more surprising for audio, where LogicLM, despite near-perfect MIDI accuracy, remains notably brittle. Among models, Gemini Pro achieves the highest performance across most conditions. Transposition yields the highest accuracies across models, while Chord Identification scores slightly below Syncopation. Overall, current systems reason well over symbols (MIDI) but do not yet ”listen” reliably from audio, with reasoning strategies having little impact over accuracy. Our method and dataset make the perception–reasoning boundary explicit and o!er actionable guidance for building robust, audio music systems.

Downloads

Next from AAAI 2026

Low-Resource Rhythm Learning of South Asian Beat Structures: Machine Learning Approaches to Nattuvangam

Similar lecture

On the Actionability of Outcome Prediction