AI Agents Automate Work, TTS Models Shrink

🤖 Agents

AI Agents Reshape Knowledge Work

The Rundown

Perplexity's latest research reveals how its AI agents are fundamentally altering knowledge work. By analyzing production data from its Search and Computer products, the company found that its autonomous agents perform significantly more work per user session. Computer, for instance, handles 26 minutes of autonomous work per session, a stark contrast to Search's 33 seconds. This automation extends to task decomposition and execution, tasks users previously managed manually. The shift is evident in follow-up query patterns, which move towards higher-order tasks like verification and extension. Crucially, this autonomy boosts execution quality, with per-query dissatisfaction rates dropping by 55% on Computer compared to Search. The implications for efficiency are profound: matched tasks see completion times slashed from 269 minutes to just 36 minutes, representing an estimated 87% reduction in time and a 94% cost saving compared to human-assisted search. Beyond speed, these agents expand the scope of achievable work, enabling users to tackle more complex, cross-disciplinary, and composite tasks that were previously impractical.

The details

• Perplexity's Computer product performs 26 minutes of autonomous work per user session, compared to 33 seconds for Search.
• Per-query dissatisfaction rates are 55% lower on Computer than on Search due to enhanced execution quality.
• Completion time for matched tasks is reduced from 269 minutes to 36 minutes, an 87% decrease.
• Estimated time and cost savings for matched tasks reach 87% and 94%, respectively, compared to humans using Search alone.

Why it matters

This study provides concrete evidence of AI agents' high-impact impact on productivity. Startups can leverage these findings to build tools that automate complex workflows, reduce operational costs, and unlock new service offerings by empowering users to tackle more ambitious projects.

🧠 AI Systems

DuMate-DeepResearch: Auditable Multi-Agent Systems

The Rundown

Addressing limitations in current deep research agents, DuMate-DeepResearch introduces a multi-agent framework built on the Qianfan Agent Foundry. This system tackles challenges like long-horizon planning, single-agent bottlenecks, hallucination risks, and process auditability. DuMate decouples the core agent logic from an extensible tool ecosystem, ensuring every decision and tool invocation is traceable. Key innovations include a graph-based dynamic planning strategy that refines research roadmaps iteratively through reflection and parallel branching. A recursive two-level execution design delegates complex sub-tasks to inner Search Agents, isolating noisy retrieval and stabilizing long-horizon execution. Furthermore, a rubric-based test-time optimization mechanism generates task-specific quality criteria to scaffold evidence-grounded synthesis and adaptive stopping. Across two benchmarks, DuMate-DeepResearch sets new current best results, achieving the best overall score of 58.03% on DeepResearch Bench and 61.95% on DeepResearch Bench II, leading in information recall and analysis.

The details

• DuMate-DeepResearch achieves a 58.03% overall score on the DeepResearch Bench, a new current best.
• On DeepResearch Bench II, the system reached 61.95% overall, ranking first in information recall and analysis.
• The framework uses a graph-based dynamic planning strategy for iterative roadmap refinement.
• A recursive two-level execution design isolates sub-tasks to inner Search Agents for stability.

Why it matters

This framework offers a blueprint for building more reliable and transparent AI systems for complex research tasks. Startups can adapt these multi-agent principles to develop specialized research assistants or data analysis tools that offer auditable workflows and improved accuracy.

🔊 Text-to-Speech

dots.tts: Continuous Autoregressive Text-to-Speech

The Rundown

The dots.tts technical report introduces a 2B-parameter continuous autoregressive text-to-speech (TTS) foundation model. This model operates in a continuous latent space, differentiating itself from existing approaches through three key innovations. First, it employs an AudioVAE trained with multiple objectives to create a semantically structured and prediction-friendly speech space. Second, full-history conditioning in the flow-matching head ensures long-range consistency and minimizes drift during generation. Third, reward-free self-corrective post-training enhances robustness and acoustic quality. Trained on a large multilingual corpus, dots.tts achieves top performance on Seed-TTS-Eval, with Word Error Rates (WERs) of 0.94% (zh), 1.30% (en), and 6.60% (zh-hard), alongside SIM scores of 81.0, 77.1, and 79.5, respectively. For efficient inference, CFG-aware MeanFlow distillation enables low-latency speech generation, with first-packet latencies of 85ms in output streaming and 54ms in dual-streaming modes. The code and checkpoints are released under the Apache 2.0 license.

The details

• dots.tts achieves a 0.94% WER on the zh test set and a 1.30% WER on the en test set.
• The model offers first-packet latencies of 85ms in output streaming and 54ms in dual-streaming modes.
• It features a 2B-parameter continuous autoregressive architecture for TTS.
• The model demonstrates strong generation stability, voice cloning, and emotional expressiveness.

Why it matters

This advancement in TTS technology offers significant opportunities for startups in content creation, accessibility tools, and personalized user experiences. The low latency and high quality make it ideal for real-time applications and scalable voice services, potentially disrupting existing market players.

📚 LLM Training

SETA: Sparse Subspace Sharing for Continual Learning

The Rundown

SETA (Mixture of Sparse Experts for Task Agnostic Continual Learning) addresses the plasticity-stability dilemma in continual learning for LLMs. This framework uses adaptive sparse subspace decomposition into task-specific and shared expert modules. Unlike standard updates where parameters compete, SETA isolates knowledge into unique experts for specific patterns and shared experts for common features. This structure is maintained via adaptive elastic anchoring and a routing-aware regularization that protects shared knowledge at both weight and routing levels. A unified gating network automatically retrieves the correct expert combination during inference. Experiments across diverse benchmarks show SETA achieving competitive or superior performance compared to current best continual learning baselines. It demonstrates strong retention of early-task knowledge and improved backward transfer on LLaMA-2 7B and Qwen3-4B models.

The details

• SETA achieves competitive or superior overall performance against current best continual learning baselines.
• The framework demonstrates strong retention of early-task knowledge on LLaMA-2 7B and Qwen3-4B.
• It utilizes adaptive sparse subspace decomposition into task-specific and shared expert modules.
• Routing-aware regularization protects shared knowledge at both weight and routing levels.

Why it matters

SETA's approach to continual learning is vital for startups building AI products that need to adapt and learn over time without forgetting previous capabilities. This could enable more dynamic and evolving AI assistants or specialized models that continuously improve with new data.