Blog
Research on AI agents that write code, explore ideas, and improve themselves. Experiments examine their capabilities, creativity, and limits.
-
Comparing creative research answers with CREST
Caesar's answers won a strong preference from AI judges on shared creative tasks, with public explanations that readers can inspect and challenge.
-
CREST: comparing papers from five AI scientists
AI judges favored Sakana Agentic Scientist v3's released papers, but its two-paper collection limits broader claims about the system.
-
Values Shape What Self-Improving Coding Agents Build
Short, persistent instructions changed the algorithms coding agents developed, with guidance favoring exploration generally outperforming guidance favoring small, verified changes.
-
Deriving agent values from research papers and auditing their sources
Verified papers supplied proposed agent guidance, but checking citations did not establish that the guidance would improve performance.
-
An AI-generated solver improved 21 historical circle-packing values Record
Established methods yielded independently checkable improvements across many circle counts; a follow-up starting from the finished solver found no additional records.
-
An AI-generated solver improved the 27-circle packing record Record
A solver developed for a smaller packing problem improved a historical best-known result, with saved coordinates for independent checks.
-
An AI-generated solver exceeded the winning AHC039 contest score Record
Starting with a simple rectangle, an agent developed a solver that surpassed the contest winner's official score after the competition had ended.
-
Thirty 2048 runs separated a capability effect from generalization
Instructions encouraging exploration improved 2048 performance on the familiar board size, without a detectable advantage on a larger board.
-
A metric-focused agent passed fresh 2048 games and failed a larger board
Fresh 2048 games missed a board-size failure, showing why generalization tests need to be chosen before results are known.
-
How AI agents rewrote their values across 41 exploratory runs
In these runs, agents always revised editable values, and uncapped files expanded from brief guidance into lengthy instruction documents.
-
Auditing an 11-of-14 tally in agent bin-packing experiments
An apparent advantage in packing performance weakened under audit, while the agents' different approaches to building solvers remained visible.
-
Pre-registering AI agent experiments at the run level
More test problems do not make up for too few agent runs, even when the analysis is written down before scoring.
-
Values in self-improving agents: five design hypotheses
A proposal asks how starting values could shape the choices an agent makes as it improves its own tools and methods.
-
Two AI agents differing by one values file, with no reliable winner
One agent explored broadly while another built a connected project with extensive checks; the values comparison remains inconclusive.
-
Caesar illustrates reports using images from cited sources
Source images can give report illustrations recognizable details while carrying mistakes from their descriptions into the finished picture.
-
How Caesar's web exploration graphs grow and reconnect
Saved web explorations show frequent returns to earlier pages and large differences in how new routes and connections develop.
-
Notes from well-connected pages scored higher in a Caesar comparison
Notes from well-connected pages received higher ratings, without showing that context from nearby pages caused the advantage.
-
Human and AI evaluations gave different pictures of Caesar's advantage
People showed a modest preference for Caesar's summarized answers, offering weaker and less consistent support than AI judges gave its full reports.
-
Measuring model-family preference in AI judges
A strict AI judge can still favor its own model family, even when its scores initially suggest the opposite.