-
SwanBench-Speech Benchmarks Long-Form Speech Generation Across 17 Scenarios
SwanBench-Speech is a benchmark for the part of speech generation that single-utterance demos hide: long form. Most TTS evaluations score short clips on naturalness and prosody. What breaks when you push to minutes of continuous speech — drift, identity loss, expressive monotony — they tend not to measure. ## 1,101 samples across 17 scenarios The… Continue reading
-
NVIDIA’s AXPO Closes the Thinking-Acting Gap in Multimodal Agentic RL
AXPO — Agent Explorative Policy Optimization — targets a quiet failure mode in how multimodal agents are trained with reinforcement learning: the Thinking-Acting Gap. Standard RL recipes like GRPO end up underweighting the tool-call moments that decide whether the agent succeeds. ## Why GRPO drops the ball on tool use The diagnostic data is sharp.… Continue reading
-
Alibaba’s Qwen 3.7-Max Runs 35 Hours Straight Across 1,158 Tool Calls
Alibaba announced Qwen 3.7-Max at the Apsara Cloud Summit, with API rollout already started. It’s positioned as Alibaba’s most capable agent model so far — built for long-running, multi-step workflows rather than single-prompt question answering — and ships with a 1M-token context window and a native extended-thinking mode. ## The 35-hour demo The standout number… Continue reading
-
A Power Law for LoRA Memory Pinpoints When Models Memorize Verbatim
A team from Zhejiang University and Alibaba turned LoRA into something it usually isn’t — a measurement instrument — and used it to pin down how much a model can actually memorize through low-rank fine-tuning. The headline is a Parametric Memory Law: a power-law relationship between loss reduction, effective parameters, and sequence length. ## A… Continue reading
-
LaRA Detects RL Post-Training Contamination by Watching Layer-Wise Representations
LaRA targets a fast-growing eval-integrity problem: as models go through reinforcement learning post-training, benchmark questions can quietly leak into the training data — and the model passes the test by remembering rather than reasoning. Layer-wise Representation Analysis is the proposed detector. ## Looking inside, not outside Most contamination detection compares answers to suspected sources from… Continue reading
-
LlamaIndex Rewrites LiteParse in Rust for Up to 100x Faster Document Parsing
LlamaIndex shipped LiteParse v2.0, a complete Rust rewrite of its open-source document parser that claims up to 100x faster parsing. It runs entirely on your machine — no cloud, no LLM, no API key — and is aimed at the unglamorous step every RAG and agent pipeline hits first: turning a PDF, DOCX, or scanned… Continue reading
-
LiteCoder-Terminal Scales Terminal Agent Training to 11,000 Trajectories
LiteCoder-Terminal is a scaling effort for one of the harder agent settings: the actual command line. The release moves the data side from a sub-1,000-trajectory preview to 11,255 agent trajectories collected across multiple scaffolds, and broadens the task mix beyond what previous terminal-agent datasets covered. ## Multi-scaffold, three new task categories The earlier preview trained… Continue reading
-
Artisan’s Ava 2.0 Drops AI BDR Pricing 10x and Goes Self-Serve
Artisan announced general availability of Ava 2.0, a ground-up rebuild of its AI BDR — the autonomous sales rep that runs the outbound pipeline without a human in the loop on each step. The launch also slashed the entry price by 10x and dropped the gate: $250/month instead of $2,500/month, self-service onboarding in under 10… Continue reading
-
Firecrawl /monitor Pings Your Agent Only When Web Pages Actually Change
Firecrawl launched /monitor, a feature that pings your AI agent the moment a webpage you care about actually changes — instead of having the agent re-scrape and re-summarise the whole page on every check. ## A goal in English, a webhook on change You point it at a URL, describe in plain English what to… Continue reading
-
Baidu NAVA Treats Audio and Video as One Signal Instead of Two Streams to Align
NAVA — Native Audio-Visual Alignment for Generation — comes out of Baidu’s ERNIE research group and stakes a position the field has been edging toward: audio and video should be learned together as one signal, not two separate streams stapled at the seam. ## Native, not stitched Most audio-visual generation pipelines today are bolted together.… Continue reading
Top AI Product
Every day, hundreds of new AI tools launch across Product Hunt, Hacker News, and GitHub. We dig through the noise so you don't have to — surfacing only the ones worth your attention with honest, no-fluff reviews. Explore our latest picks, deep dives, and curated collections to find your next favorite AI tool.
Top AI Product 