Abstract Compression of Conversational Audio Context for LLM-based ASR
Research on abstract compression of conversational audio context for LLM-based automatic speech recognition.
arXiv cs.CL — Computation and Language·21 Oct 2026
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Research on abstract compression of conversational audio context for LLM-based automatic speech recognition.
Research paper proposes moving LLM agent memory architectures beyond atomic fact extraction to improve long-term dialogue reasoning.
A research paper analyzes the combined effects of dimensionality reduction and quantization on text embedding compression.
Researchers propose a hidden-state probe method to replace LLM prompt metadata filtering in RAG architectures.
Researchers show Group Relative Policy Optimization improves automatic speech recognition trained on synthetic voice data for banking domains.
Researchers introduce StartLux-Decision, reasoning models from 0.8B to 27B parameters using confidence-guided deliberative inference.
Researchers identify safety risks in agentic AI memory caching, showing failure counters fail to detect stale reasoning after fact changes.
A new benchmark, REPOCOD, reveals LLMs struggle with real-world repository-level coding tasks despite high benchmark scores.
Researchers introduced CETBench, showing LLMs are brittle at checking whether functionally equivalent code programs match.
Researchers propose DiDi-Merging, a dynamic model merging method using differentiable rank selection to manage storage budgets.
Researchers introduced Penumbra, an adversarial search framework to evaluate agent compliance with complex financial regulatory obligations.
An arXiv paper argues LLM agents cannot reliably determine when rule verification requires external human judgement at runtime.
Researchers test training-free activation steering to store agent experiences as vectors, challenging assumptions about text-based memory.
Researchers examine scaling laws for agentic reinforcement learning, highlighting how synthesized tasks and environments reshape data design.
Researchers propose a counterfactual evaluation framework for LLM-generated data repairs by separating target records from applied updates.
Researchers introduced HarnessPro, a framework for optimizing AI agent harnesses to improve performance while reducing execution costs.
Researchers introduce WeAct, a method for dynamic weight switching during agent execution to mitigate catastrophic forgetting in long-horizon tasks.
Researchers introduced LLM-Attacker, a unified red-teaming framework to evaluate text reconstruction attacks on split LLM architectures.
Difflet is a serving engine designed to optimize diffusion model inference performance on AWS Trainium2 hardware.
Researchers propose MultiAgent-OPD, an end-to-end training method for hierarchical multi-agent systems using on-policy distillation.
Researchers introduced SAFE-Mem, a framework to secure agentic LLM memory using fine-grained authority and governance-aware retrieval.
A new research paper systematizes the evaluation of open-text legal answers using large language models, deriving eight desiderata.
A new research paper explores long-horizon trajectories for terminal agents executing complex professional tasks and gathering evidence.
Researchers propose a method to programmatically construct challenging browser-use agent tasks by environment interventions rather than new datasets.
Researchers extend refusal steering vectors to three open-source Mixture-of-Experts models, analyzing how safety alignment is managed during inference.
Researchers introduce TokenSwap to benchmark and mitigate the 'modality gap'—inconsistent outputs in multimodal LLMs for equivalent inputs.
Researchers propose SINT-Flow, an LLM-based framework with five operators for fully automated, end-to-end schema integration across diverse input tables.
Research introduces boundary-enforced Physics-Informed Neural Networks (PINNs) with a priori error bounds for elliptic Dirichlet problems.
Research finds Chain-of-Thought (CoT) reasoning in LLMs complicates refusal control, making activation steering less effective than with non-CoT models.
DiffEEG, a 9.6M-parameter self-supervised diffusion model, is proposed for EEG seizure detection, addressing data scarcity and class imbalance.