AI Research Pulse

Small encoders can replace prompted LLM choices, but the reading setup and initialisation matter:

• Per-option scoring beat joint scoring by 5.8...

Two recent papers take different routes to self-improving agents: one builds a competitive system, the other improves the improvement process...

Frontier black-box agents excel at long-horizon reasoning, but they have no internalized knowledge of how a specific environment unfolds. The proposed...

A new benchmark asks whether a web agent's acceptance rule still works when requirements change. AgentProofBench pairs tests of a frozen rule with...

The common view that constrained decoding hurts reasoning looks more nuanced in this new preprint. Its causal ablations across eleven open-source...

Child AI safety is typically judged at the model level. A study of 1,331 publicly reported child-related AI incidents argues that model-level...

Headline hallucination detection scores may not be directly comparable. Existing methods are often compared across benchmarks that define and measure...

Mixture-of-experts transformers use only a sparse subset of weights in each forward pass. This work explores a second conditional computation...

Video formats built for sequential playback rely on inter-frame references, which become costly when training samples frames at random.

• Current...

The new Continual Learning Exam tests whether LLM agents can genuinely improve from accumulated experience across 600 progressive levels in...

Progress-aware advantage shaping markedly boosts critic-free training for long, multi-step agent tasks. ProAS uses an Agentic Reference Graph to...

DynaFlow moves beyond fixed collaboration structures by letting the main agent decide when to delegate and synthesize executable workflows on...

• OPTIC stores one token-shared code tuple decoded per layer via affine transforms fitted to the output path.
• Memory savings: ~100× PLE compression...

Static benchmarks assume fixed tools and workflows, yet real ecosystems constantly evolve as services disappear, get replaced, or spawn more efficient...

Qwen-Image-PE shows that teaching models visual design principles transforms underspecified user requests into coherent image-generation...

Principia forms comparative causal hypotheses from inputs and outcomes, then synthesizes recurring relationships into inductive principles that refine...

Sparse autoencoders extract human-readable linguistic features from black-box AI text detectors, delivering stronger out-of-distribution performance...

Anthropic's S-1 filing explicitly ties model misuse, undetected capability jumps, and deliberate model restrictions to potential revenue loss and...

The full technical guide details gradient checkpointing, FP16 mixed precision, and 8-bit AdamW to slash memory use by 30-40% while hitting 2,000...

HiSTEVE 1.0 explores harnessing human-interface VLA agents for open-world games.