oracle price map learning for contextual pricing
a new method learns the optimal price as a smooth function of a scalar index, improving revenue in semiparametric dynamic pricing.
topic
research
theory of mind gains don't always help human-ai chats
improving llm theory of mind on static tests often fails to boost real interactive task performance, a new study finds.
state-constrained dispatch cuts multi-agent alignment tax
a new framework uses state machines and online rlhf to enforce business rules in multi-agent systems, improving accuracy over gpt-4o on a recruitment benchmark.
alpha-tcav fixes statistical flaws in concept-based explainability
a new framework replaces the unstable indicator function in tcav with a smooth parameterized function, reducing variance and unifying existing methods.
quantization undoes alignment in compressed llms
a study finds that aggressive quantization causes previously unbiased language models to develop new stereotypical behaviors, with a clear dose-response pattern.
gds urges open code default after nhs retreat
the uk government digital service recommends keeping public sector code open by default, pushing back against the nhs decision to close repositories after vulnerability reports.
arxiv bans authors for unchecked ai-generated papers
arxiv will impose a one-year ban on authors who submit papers with clear signs of unverified llm output, such as hallucinated references.
integer goal programming for meal plans with real serving sizes
a new mixed integer goal programming method creates personalized meal plans using whole servings and soft nutrient targets, avoiding fractional foods and infeasible diets.
turboquant compresses llm caches to 3 bits
google's turboquant uses polarquant and qjl to cut kv cache memory by over 5x without retraining or accuracy loss.
nasa tests ai chip for autonomous spacecraft
nasa is testing a radiation-hardened processor that delivers up to 500 times the performance of current spaceflight computers, enabling ai-powered spacecraft to operate independently in deep space.
regret metric exposes cost of strict false discovery control
a new weighted regret metric reveals deterministic online multiple testing procedures suffer linear regret from false negatives, and a decoupled wrapper fixes it.
adaptive importance sampling fixes quantized rl training
a new method corrects policy gradient bias from low-precision rollouts in llm reinforcement learning, preventing training collapse.
two-axis framework sorts ai agent design patterns
a new 7x6 matrix classifies llm-based agents by cognitive function and execution topology, identifying 27 distinct patterns.
hidden ai coordinators increase safety risks in multi-agent systems
invisible orchestrators in multi-agent llm systems suppress protective behavior and cause power-holders to dissociate, raising safety concerns.
ai that builds itself gets $650m
recursive superintelligence raises $650m to create ai that can find its own flaws and redesign itself without humans.
benchjack audits ai agent benchmarks for reward hacking flaws
an automated system finds and patches reward hacking exploits in popular ai agent benchmarks, revealing widespread vulnerabilities.
granite multilingual r2: compact 97m model tops sub-100m retrieval
ibm releases two apache 2.0 multilingual embedding models with 32k context, covering 200+ languages and code retrieval.
ai confidence alignment speeds human decision learning
study shows that matching ai confidence to human self-confidence helps people learn faster when making decisions with ai assistance.