AI & Agents
Research on AI & Agents: explanations, evidence, limitations and connections between papers.
17 results
Match each distillation choice to its measured outcome
A team distilling a model can supervise fewer generated tokens, change who generates its training responses, or ask it to produce an answer in fewer decoding steps. Each choice changes a different part of the process, so a quality gain under one choice does not establish a shared saving in computation or stored memory. The studies offer a practical starting rule: name the choice, the behavior being measured, and the comparison before interpreting a result. Their reported gains belong to particular tasks and protocols.
Small GPU runs can guide large training searches
Primus estimates large-scale training throughput from small GPU runs; AMD evaluations found close agreement on tested models.
A two-stage training method improves GUI grounding
A GUI grounding method combines distillation and reinforcement learning, scoring higher than single-stage training across five tests.
Why shared mistakes can change whether teamwork helps
A method estimates whether acting on teammates’ inferred beliefs helps, using error costs and how mistakes occur together.
MagicBeanStalk builds training problems from hard seeds
MagicBeanStalk builds model-calibrated math training problems, improving average test accuracy in reported evaluations.
TacitVLA trains on reasoning so actions need not wait for its generated text
TacitVLA separates reasoning training from action generation, matching explicit reasoning across four manipulation and driving tests.
Checking each action can stop agents’ unauthorized acts
In mock-tool episodes, checking each proposed action reduced violations while matched tests tracked whether authorized work continued.
Sauron checks evidence before judging agent tasks
Sauron checks task states and action records, improving rubric judgments in tests of simulated and recorded agent tasks.
A harness-editing method improves transfer to unseen tasks
A method regulates agent harness edits with a frozen model and improves average scores on unseen workspace tasks.
Written skills improve agent procedures without changing weights
SkillOpt improves frozen agents through validated text edits, with gains across tested models, tasks, and execution modes.
A human-trained user simulator improves agent training
MIMESIS models varied human behavior; agents trained with it score better than GPT-5.5-trained agents across nine simulated users.
A graph result powers faster structured language decoding
A GPU decoding engine recomputes format masks and speeds six JSON extractions on one tested laptop.
CanonQ trains language models to use one shared low-bit codec
CanonQ improves measured accuracy and text prediction for a tested language model under joint low-bit compression.
Home automation can use different models for different calls
A live test of per-call model choices matched all-hosted human judgments on 39 of 43 counted cases, at lower hosted spend.
Visible labels can mislead language model judges
Tests find that visible system labels can distort language model judgments of entity alignment decisions in the studied conditions.
RT-SAFE reveals safety failures hidden by navigation success
RT-SAFE tests urban-agent safety during decisions; hard-route runs exposed risks hidden by successful arrivals in a changing world.
Why language models can overstate their progress
A statistical model and loop tests show why reused examples can make language-model progress scores too optimistic.