The First Query Sets the Search Trajectory

Featured

01 Panoramic Memory, On-Demand Views

AlayaVista separates two jobs inside a world model. A panoramic state preserves the whole scene, while a local renderer generates only the camera’s current view. The model expands one perspective into a 360-degree scene prior, then updates that latent state as the camera moves.

A latent viewport renderer extracts the requested angle on demand. A perspective refiner restores detail, suppresses artifacts, and increases resolution. The system avoids generating a complete spherical video at every step or explicitly reconstructing a 3D scene. Chunked autoregression and few-step distillation help reduce interactive latency.

MUGEN supports this design with 1,318 hours of real panoramic video at 4K resolution or higher, plus semantic and geometric annotations. Separating global memory from current observation is the compelling idea. Long rollouts must still show whether it controls occlusion-related forgetting, drift, and latency.

Key takeaways:

• A panoramic latent state may preserve off-screen content more reliably than forcing a local video model to remember a long history.
• On-demand rendering concentrates compute on the region the user can actually see.
• Evaluate long-term consistency, fast camera motion, and end-to-end latency instead of focusing only on individual frames.

Source: AlayaVista: Streaming World Modeling from Panoramic States to Perspective Video

02 Deep Search Can Fail On Move One

A deep-search agent can retrieve the right material and still fail to connect the evidence. Question's Gambit splits initial retrieval into clue extraction, complementary queries, result merging, and reranking. This creates context that later search steps can aggregate and verify.

The method raises BrowseComp-Plus answer accuracy from 83.1% for a strong baseline to 90.5%. Initial querying is a strategy decision that shapes the entire exploration path, not a one-off prompt edit. Improving branch coverage at the start may beat adding more reasoning inside the search loop.

Key takeaways:

• When deep search fails, first check whether the initial queries covered complementary clues.
• Treat initial retrieval as a separate module so results can be merged and reranked consistently.
• Track first-round recall alongside final accuracy to distinguish path failures from retrieval failures.

Source: Question's Gambit: The First Move Matters in Agentic Deep Search

03 Legal RAG Needs Sufficient Evidence

Legal RAG must ask more than whether cited material is relevant. It must decide whether the evidence supports the conclusion and whether insufficient evidence calls for abstention. This work converts existing legal QA datasets into contextual preference data and introduces LegalRewardBench for noisy or incomplete retrieval settings.

With direct preference optimization, the best data combination improves results by up to 25.6 percentage points over the baseline. Training mostly on criminal law from Victoria, Australia, also produces a 16.2-point gain on US housing-law questions.

Performance depends heavily on how the preference data is built. Length balancing alone changes results substantially, suggesting that reward models may learn superficial signals. Domain reward models can support evaluation and training, but they cannot replace human review, evidence audits, or real-world validation.

Key takeaways:

• Evaluate citation relevance and evidence sufficiency as separate properties in legal RAG.
• Control superficial signals such as answer length when constructing preference data.
• Cross-jurisdiction gains do not prove that a model can make dependable legal judgments.

Source: Building Legal Reward Models for Grounding and Abstention

04 Internet Training Has Two Communication Bottlenecks

Low-end GPUs training across the internet face two competing traffic flows. Data parallelism synchronizes gradients, while pipeline parallelism transfers activations. Optimizing either path alone can leave the other blocking the entire run.

Amazon's system uses two asynchronous loops. A fast loop compresses both traffic types to maintain throughput. A slow loop occasionally runs uncompressed forward and backward passes outside the critical path. Their delayed calibration signals help a spectral-correction optimizer denoise gradients affected by masking.

Compressing pipeline traffic alone delivers up to 9× higher throughput. Compressing both flows exceeds 40× on roughly 200 Mbps links while matching uncompressed training in domain adaptation and continual pretraining. Deployment tests should reproduce these gains under the target network and workload.

Key takeaways:

• Measure gradient synchronization and activation transfer together when evaluating decentralized training.
• Asynchronous calibration supports aggressive compression without blocking the main training loop.
• The 40× result comes from a specific network condition and needs retesting on each target setup.

Source: Communication-Efficient LLM Adaptation over Decentralized GPU Meshes