AI & Agents
Research on AI & Agents: explanations, evidence, limitations and connections between papers.
56 results
A gate checks memories before consequential actions
AMV-Guard checks action-linked memories when evidence can change a decision, lowering verification cost in simulated tool-use tests.
Trace the evidence path before trusting an agent decision
A simulator can produce the needed value while an agent answers from a different file. A code warning can sound convincing while its cited program relationship is wrong. A safety monitor can run while a detected result fails to reach its blocking decision. The common task is to follow evidence to the decision it must support. Each study tests a specific connection and bounds its conclusion to the evidence it observed.
AgentTriage makes agent security diagnoses replayable
AgentTriage makes agent security diagnoses replayable from recorded events, while abstaining when required evidence is missing.
Models can spot changed guidelines without fully updating advice
ChronoEvo finds models can recognize updated guideline advice without fully carrying its important changes into recommendations.
Twice gives search agents credit for individual turns
Twice links agent turns to shared later sampling, improving seven-test exact-answer averages over GRPO for two Qwen models.
Scaling laws estimate safety-probe error from size and labels
Safety-probe scaling laws estimate error from model size and labels, with next-model predictions tested across three model families.
Simulated tests predict safeguards on deployment tasks
Simulated control evaluations predicted LinuxArena safety more accurately than attacks run in BashArena, a separate test setting.
Trace2World builds browser app worlds from interaction traces
Trace2World builds browser app environments from real interaction traces and improves workflow fidelity over a task-local builder.
Match overlap evidence to the decision it will change
A repair model may have seen the patch it is asked to produce. A pooled brain-scan repository may give one person two identifiers. A math training problem may resemble a benchmark question while posing a different task. Each case creates pressure to act on a possible overlap, and each makes a mistaken decision costly. The evidence needed for that decision depends on what is being compared. The studies use code history, reviewed images, and judgments about problem content to move from possible matches toward conclusions within their tested settings.
A testbed measures coding-agent gains on science models
Agent4Sci measures coding-agent gains on scientific ML tasks, while audits and separate test scoring guard against evaluation gaming.
Chat refusals often fail to prevent unsafe tool calls
A matched-request study finds chat refusals often fail to prevent unsafe tool calls across five tested language models.
World models can miss events in frames that look accurate
A transition-level diagnostic finds reward and ending events often drive return error in reliable Atari world-model comparisons.
A benchmark measures when reassurance leaves advice goals unmet
A diagnostic test finds that conversational models vary in leaving practical goals unmet across authored advice scenarios.
Aligning context fields improved industrial app ranking
EASformer aligns app histories with candidate context, improving offline ranking and eCPM in two tested online scenarios.
Agent S4 evolves computer-use tools from task experience
Agent S4 evolves reusable computer-use tools from task attempts, improving results in tests of general and scientific software tasks.
Score-guided sampling improves sparse attention
Stratified Sparse Attention uses score-guided sampling to improve sparse attention results in selected long-context language-model tests.
Inflated tool listings sometimes mislead agent buyers
A simulated tool-market study finds inflated seller listings can lower buyer task success in severe cases.
WildAWM creates computer worlds for cross-interface agent training
WildAWM creates cross-interface computer tasks; trained Qwen models improve on average across eight external task tests.
Tandem speeds hybrid phone-agent model serving
Tandem adapts hybrid model serving to changing phone-agent requests, reducing average step latency on a Snapdragon 8 Elite trace.
A price sweep reveals how seller cues sway AI shopping
Price-equivalent auditing measures how much seller cues offset product prices, revealing differences hidden by equal-price choices.
A model can see a failed step yet trust the agent’s record
EviGate checks agent decision states and improved task success across three tested settings with calibration from each matching domain.
How simulator report delivery changes agents’ answers
On verified PHREEQC questions, report retrieval lowered accuracy for two GPT agents; naming the report improved evidence use.
How shared memory can lead agents into the wrong account
A shared-memory audit separates credential exposure from account misuse and finds owner wording reduced misuse in controlled tests.
Place each trace where it can inform the next decision
A robot may need to choose the next object to move, an application judge may need to see what happens after a click, and a tool-using agent may need review before changing a record. Each can produce a trace, but the trace helps a different decision. The studies suggest placing evidence where that decision occurs. One robot method feeds reasoning into control; another moves text generation off the immediate action path. Application interaction informs a judge, while a grounded tool-use draft reaches a reviewer before a change. Their reported task success and timing results belong to their separate tests.