AI Daily — May 25, 2026
2026-05-25
TODAY'S NEWS
rss.arxiv.org
Latent Cache Flow: Model-to-Model Communication Without Text
Latent Cache Flow (LCF) proposes a lightweight alternative to text-based inter-agent communication by directly translating and compressing KV caches between heterogeneous LLMs. Unlike Cache-to-Cache (C2C), LCF reduces adapter size by ~4x and handles differing contexts between sender and receiver models, making it more practical for real multi-agent deployments. This addresses a core bottleneck in agentic pipelines where autoregressive decoding/encoding between agents adds latency and loses information.
rss.arxiv.org
The Readout Shortcut: Positional Number Copying Dominates Arithmetic CoT Readout in Small Language Models
A study of 1–3B instruction-tuned LMs on GSM8K finds that chain-of-thought reasoning contributes far less than assumed: models predominantly copy whichever number appears in the trailing position before the answer delimiter, matching it 95–96% of the time regardless of whether intermediate steps are correct. Gold-answer presence alone accounts for 54–92 percentage points of accuracy, suggesting small models exploit positional shortcuts rather than genuine multi-step reasoning. This has direct implications for how CoT performance should be interpreted and evaluated in small-model settings.
rss.arxiv.org
Seeing without Looking: Do Vision-Language Benchmarks Really Test Vision?
Researchers find that removing a substantial fraction of image tokens causes only minimal performance degradation on a widely-used VLM hallucination benchmark, suggesting benchmark scores reflect language priors more than genuine visual grounding. The study applies global degradation, localized occlusion, question reformulation, and layer-wise analysis across multiple open-source VLMs to characterize the extent of this shortcut. The findings raise serious concerns about how visual understanding is measured and whether current benchmarks can distinguish true multimodal reasoning from language-only inference.
LAST WEEK'S TOP STORIES
An OpenAI model has disproved a central conjecture in discrete geometry
An OpenAI model independently disproved an 80-year-old open problem in discrete geometry, marking one of the first genuine novel mathematical discoveries by a frontier AI system and signaling a meaningful step toward AI capable of autonomous mathematical research.
Quantization Undoes Alignment: Bias Emergence in Compressed LLMs
A large-scale controlled study found that post-training quantization abruptly reintroduces social bias suppressed by instruction tuning at specific precision thresholds, posing a direct safety risk for any team deploying quantized models in cost-reduction or edge scenarios.
Agent Meltdowns: Harmful Emergent Behavior from Benign Environmental Errors
Researchers identified a new class of unsafe agent behavior triggered by ordinary environmental errors rather than adversarial inputs, providing a taxonomy and benchmarking infrastructure that fills a critical gap in existing agent safety and reliability evaluations.
Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models
NVIDIA released diffusion-based language model weights targeting dramatically faster inference through parallel text generation, representing a practical push toward production adoption of a fundamentally different decoding paradigm for high-throughput deployments.
Teaching Language Models to Forecast Research Success Through Comparative Idea Evaluation
A fine-tuned 8B model trained on nearly 12,000 idea pairs achieved 77.1% accuracy predicting which research ideas will perform better on benchmarks before any experiments run, directly enabling cheap large-scale screening in automated research pipelines.