The Archive

Task-vector merge collapse in LLMs: variance-based prediction and repair using pre-merge interference scoring.

KV²: query-agnostic KV-cache compression via selective reconstruction for reusable long-context inference.

LLM agents exhibit systematic source preference in search/recommendation across domains; mitigation strategies proposed.

Study of LLM calibration in investigation tasks: frontier models identify root causes (84%) but overstate evidence confidence (97% in 9B model).

San Antonio City Councilmember Ric Galvan remembers when data centers first arrived in his city in the 2000s, looking like relatively unassuming office buildings. But that's changed in recent years, as data centers have grown into massive "hyperscale" facilities spanning millions of square feet and housing the physical infrastructure for AI. There are now more than a dozen data centers in Galvan's roughly 55-square-mile district, and he expects that number to grow to around 20 over the next few years if proposed projects are completed. They're clustered along two main roads, concentrating nui...

Madison, a server in New York City, greets every table by asking about each diner's allergies. Lately, there have been some close calls. "Sometimes people will tell me they have a shellfish allergy, and I'll come back, and they won't ask me any questions," says Madison, who asked that her last name be withheld to protect her identity. "And they'll order a fish dish that comes with a broth that is made of shellfish." When she tells them a dish contains an allergen, she's had guests push back, saying ChatGPT disagrees. "They just want to talk to ChatGPT," Madison explains. "You're allergic to s...

On-policy distillation analysis reveals student model collapse into repetitive text despite performance gains; implicit reward model misalignment mechanism identified.

Amazon is calling for people to support AI data center projects, or risk irreparable harm to the US economy and national security. In a more than 3,000 word blog posted today, Amazon web services CEO Matt Garman pushed back on public concerns around the impact that data centers may have on jobs, power demands, and the environment, saying "our nation can't afford to lose" the race for AI dominance. "With any change, there will be important questions raised, but there will also be misinformation and outright lies, and in the age of social media and 24/7 news, myths take hold faster than ever be...

Hindsight-guided distillation on ZebraMap rare disease dataset: 1.5B student slightly outperforms 8B teacher on filtered subset with statistical significance.

LiBRA algorithm removes AI-image watermarks while preserving image quality via latent-space bidirectional optimization; evaluates watermark robustness.

Defensive Policy Gradient algorithm achieves O(ε^{-3}) sample complexity for variance-reduced REINFORCE without variance assumptions; establishes lower bounds.

Three methods (contrastive steering, steering vector networks, LoRA-predicting hypernetworks) for few-shot author-style transfer in scientific abstracts show style-content trade-off.

Benchmark of classical (Gurobi), photonic (QCI Dirac-3), and simulated quantum solvers on QUBO feature selection for fraud/default detection across 13 methods.

ImmuneAgent integrates multimodal reasoning, meta-learning, and wet-lab feedback for broadly neutralizing antibody discovery from BCR repertoires across virus families.

Probing video Diffusion Transformers reveals kinematic and rigid-body physical quantities are linearly decodable from internal activations despite poor physics reasoning benchmarks.

EvoRiskBench: evolving security benchmark for workspace agents via EP-Path-EF framework mapping entry points to runtime effects; nine entry-point categories, five effect categories.

TSGuard: real-time missing-data imputation for streaming time series using graph-aware temporal models with constraint validation.

SLIM benchmark reveals latent-world-model failures on multi-object pushing tasks with language goals; encoder lacks action sensitivity.

German monolingual RAG study finds reasoning-language alignment affects retrieval; accuracy degrades when reasoning language differs from target.

Page-EntroKV: entropy-weighted KV-cache eviction framework for grouped-query attention that prevents inflated cache overhead and preserves retrieval heads.

Safe streaming flow planners aligned with execution dynamics enable real-time replanning of diffusion-based trajectories under safety constraints.

Dictyostelium literature retrieval benchmark for model-organism domain, evaluating organism-specific terminology and sparse biological corpora.

LLM textual gradients for trading-strategy optimization; addresses myopia and temporal robustness in backtesting feedback.

Conformal-prediction calibration for RL-driven NAS corrects distribution shift as search policy improves; online coverage steering guarantees.

ParaGeo decomposes paralinguistic variation in speech via frozen LM K/V projections; 80-control probe on GLM-4-Voice.

Trigger-tag mechanisms for misuse detection in open-weight LLMs lack systematic robustness analysis; practical security implications explored.

Streaming VLMs can anticipate future visual events and dynamically adjust computation without retraining via learned foresight mechanisms.

AMSC framework uses Wasserstein task embeddings to find and compose relevant prior policies for continuous adaptation in lifelong RL.

Empirical study of trade-offs between structured pruning and fault tolerance in DNNs for space applications with bit-level robustness.

Production system distills LLM signals into eight interpretable matching dimensions for resume-vacancy recruitment ranking.