Machine Learning
-
Representation Forcing Lets Unified Multimodal Models Skip the External VAE
Representation Forcing (RF), from the University of Hong Kong and ByteDance Seed, attacks a quiet bottleneck inside unified multimodal models: the external VAE everyone leans on to bridge from latent to pixels. RF replaces that crutch by having the model internally predict — and then use — its own high-level visual representations to generate images… Continue reading
-
LongTraceRL Mines Tiered Distractors From Search-Agent Traces for Long-Context RL
LongTraceRL trains long-context reasoning the way an agent actually experiences a long context — by reusing what real search agents do — and grades the model with entity-level rubric rewards instead of a single yes/no on the final answer. ## Tiered distractors from agent traces Long-context RL has been bottlenecked by sparse rewards and easy… 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
-
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
-
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
-
Qwen-VLA Adds a Trajectory Decoder to Qwen for Cross-Embodiment Robot Control
Qwen-VLA extends Qwen’s vision-language stack past perception into actual robot control — adding a DiT-based action decoder that turns understanding into continuous actions and trajectories. The goal is one model that handles many robot bodies and tasks, not a separate VLA per platform. ## Cross-embodiment by prompt The hard part of generalist robot models is… Continue reading
-
CollectionLoRA Packs 50 Image-Effect LoRAs Into a Single Adapter
CollectionLoRA tackles a quietly painful problem in customized image editing: as the number of effect LoRAs you want to ship grows, storing and dynamically loading them all turns into deployment overhead nobody planned for. Its answer is to fold up to 50 different effect LoRAs — plus few-step generation — into a single LoRA. ##… Continue reading
-
LearnWeak Trains Small Computer-Use Agents on Their Own Failures
LearnWeak is a training framework for a frustrating reality: small, open computer-use agents — the ones that click and type through apps — stay noticeably weaker than big closed models, and just throwing more synthetic training data at a domain barely moves the needle. Its fix is to stop training broadly and start training on… Continue reading
-
Bidirectional Evolutionary Search Lets Language Models Improve Themselves
Bidirectional Evolutionary Search (BES) is a new framework for self-improving language models that attacks a quiet weakness in how models search for answers: expanding from a single starting point keeps producing similar candidates. BES pairs a forward search that breeds variety with a backward search that creates feedback. ## Forward evolution, backward decomposition In the… Continue reading
Top AI Product 