absorb.md

Elad Gil

Chronological feed of everything captured from Elad Gil.

paper / eladgil / Jun 12

Massey Products as the Algebraic Backbone of Spherical T-Duality for Iterated Sphere Bundles

Cavalcanti demonstrates that the cohomological data encoded in the Gysin sequences of iterated sphere bundles can be reformulated in terms of a vanishing Massey product. When such a bundle is equipped with an integral cohomology class of transgressive degree one, it admits a T-dual iterated sphere bundle—explicitly the one corresponding to the same Massey product read in reverse. This establishes a clean algebraic duality for a broad class of geometric fibrations relevant to differential geometry, algebraic topology, and string theory.

mathematicsalgebraic-topologydifferential-geometryt-dualitymassey-products
“For a class of iterated sphere bundles, the cohomological data contained in their Gysin sequences can be repackaged into data for a vanishing Massey product.”
paper / eladgil / Jun 12

Modular Multi-Domain Digital Twins for Sustainable 6G Network Management

This paper introduces a modular, multi-domain Digital Twin (DT) architecture for 6G network management, overcoming the scalability and coordination challenges of monolithic DTs. The proposed framework enables intent-driven automation by exposing DT capabilities as services within a multi-domain orchestration layer, featuring a DT Orchestrator that dynamically composes domain-specific DT modules. A green-networking use case demonstrated significant reductions in grid consumption through optimized solar energy utilization, validating the framework's effectiveness in sustainable network management.

6g-networksdigital-twinnetwork-managementorchestrationsustainabilitygreen-networking
“A modular, multi-domain Digital Twin architecture is proposed for managing future 6G networks, addressing the limitations of conventional single-domain management in distributed and heterogeneous environments.”
paper / eladgil / May 30

Joint Encoder-LM Optimization on Mel Spectrograms Fixes Autoregressive Speech Generation Failure Modes

MELD proposes a discrete latent variable model operating directly on mel spectrograms, where the encoder and autoregressive speech language model are optimized jointly rather than independently. This end-to-end training objective addresses a fundamental misalignment in current speech LM pipelines, where separately trained encoders produce representations not optimized for downstream generation tasks. The approach outperforms both codec-based and mel-spectrogram-based baselines on zero-shot TTS and STT benchmarks, while also eliminating known failure modes in autoregressive mel modeling — specifically prolonged silence generation and word omissions.

speech-language-modelsmel-spectrogramdiscrete-latent-variablestext-to-speechspeech-to-textautoregressive-modelsaudio-processing
“Separately optimized encoders in current speech language models produce suboptimal representations for downstream tasks because they are unaware of the autoregressive objective.”