AI & Agents
Research on AI & Agents: explanations, evidence, limitations and connections between papers.
33 results
Place evidence at the decision it needs to inform
A robot can misread a grasp, a game can compile while behaving badly, and a tool-using agent can approach a change it cannot take back. Intermediate evidence matters only through the decision that receives it. Across four studies, traces inform robot control, website evaluation, later robot actions, or review of a drafted tool call. Their reported outcomes describe different tasks and measures; the useful first step is to identify the decision each trace is meant to support.
A causal method traces ad engagement to chosen labels
D-VAE estimates how marketing labels affect engagement through generated ads, with a small RED user experiment showing higher engagement for its recommendations.
Separate GPU paths speed prompt and token processing
FuTransformer tailors Llama-style inference to Hopper GPUs; tests show faster paths with a measurable INT4 text-prediction cost.
A graph monitor transfers hallucination detection across language models
LEMON uses attention graphs to detect hallucinated tokens on unseen models, outperforming probability scores in tested settings.
Canary contracts reveal where safety evidence gets lost
Canary contracts compare repaired framework paths to show when safety evidence reaches an enforcement decision.
Dia calls other models only when reasoning answers disagree
Dia uses answer disagreements to trigger independent solving, outperforming full chain-of-thought in most tested model and test settings.
Arcade lets agents request tools from a changing catalog
Arcade’s tool recommendation system ranked complete tool sets more often than BM25 in a same-harness test of tool queries.
WebCoder trains one agent to handle changing frontend projects
WebCoder trains a frontend agent to meet changing requirements, with higher scores than its base model on two evaluated task suites.
A customer-delivery benchmark finds agents struggle to build effective systems
FDE-Gym finds that AI agents struggle to discover customer constraints and turn gathered facts into effective delivery systems.
Attention statistics can be extracted without storing full matrices
A Triton backend extracts attention summaries without full matrices; added features improved detection in some tested dataset–proxy pairs.
A three-level audit tests when safety cues sway guard decisions
A three-level audit finds edited cues changed one tested guard’s decisions, with effects varying by configuration.
A language model grows exact answers by checking new code
A two-loop system expands exact question answering with checked operations, tested on synthetic merchant, business-ledger, and clinical data.
LAGO reuses specialist agents to solve coding tasks
On 150 shared code-repair issues, LAGO resolved more tasks than two rerun systems using the same model.
Scoring options separately keeps choices order-stable
Prism-1 scores choices independently, keeping answers stable under shuffling across five task sets and two tested models.
RoleQ aligns ternary weight scales with how model units work
RoleQ-T converts language models to ternary weights, scoring above tested alternatives on academic knowledge questions for Qwen3-8B-Base.
ATLAS improves model attribution when domains and versions shift
ATLAS compares source-family text patterns and improves attribution when both writing domains and model versions change.
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.