Agents that learn from video, hand off thinking, and prove their guardrails

Learning without retraining

Tutorial Agentic Learning, or TAL, lets language-model agents learn continuously from instructional videos without parameter updates, according to the paper. It consolidates tutorial videos into a persistent, structured memory called a tutorialbook, built through a two-stage adversarial process, and it retrieves information through a Table-of-Contents tree that ties retrieval directly to the model's reasoning. The point is practical: an agent that can keep what it watched can reuse it, rather than starting from scratch each time. m:…5837

The test bed is narrow but concrete. The VidLearn benchmark consists of 72 chess tutorial videos spanning 12 openings, and on it TAL achieved up to 113% relative gains on training splits and 91% relative gains on test splits compared with vanilla multimodal models. In head-to-head games, TAL reached a stronger opening-phase position than its vanilla counterpart in 69% of games. This suggests video learning works where the skill is procedural and demonstrable, like a chess opening. It does not show the same result for skills that cannot be watched and copied. m:…5837

Passing the reasoning baton

Most multi-agent methods today use a group thinking setup, the MACR paper notes, where agents generate solutions independently. MACR instead introduces a relay-based framework that dynamically transfers reasoning between agents, using confidence-guided relay triggering and a bidding mechanism to select the next agent. Relay collaboration, the paper says, lets heterogeneous agents integrate complementary insights across reasoning stages, so the agent best placed for the next step takes over. m:…3896

The reported gains are modest on accuracy and large on cost. MACR achieved a 3.3% accuracy improvement on MathVista and 3.9% on MMStar, while consuming 6.0x less token overhead on HallusionBench and MMStar compared with other multi-agent strategies. This suggests the relay idea earns its keep less through smarter answers and more through cheaper coordination: fewer tokens spent on agents duplicating each other's work. The authors say the source code is available, which lets other teams test whether the relay advantage holds outside these benchmarks. m:…3896

Guardrails that can be checked

LeanGuard tackles the other side of tool-using agents: stopping them from doing what policy forbids. It is a verified runtime that blocks policy-violating tool calls, using Lean4 to write policies that check tool calls before they reach the environment. Its policies combine state predicates with scoped past-time temporal operators, and its decision procedure is proved equivalent to LeanLTL semantics. In plain terms, the guardrail is written in a language with proof behind it, not as another prompt asking the model to behave. m:…4159

The reported results are strong on the benchmarks the authors chose. LeanGuard reduced direct prompt-injection success on DTap from 57.9% to 5.9%, and on Tau-Bench unguarded agents executed at least 33 forbidden calls while LeanGuard executed zero. Blocking alone can break tasks, so the paper also tests repair: proof-guided repair proves an admissible edit for 88 of 252 denials in a median of 2 milliseconds, and feedback mechanisms raise task success for four of five models over bare denial. This suggests verified blocking plus a route back to a permitted action is more useful than denial alone. m:…4159

Where agents still fall short

IdeaTaste is the counterweight to the optimism above. The benchmark tests language-model research proposals against future publications, using a literature snapshot frozen at 2025, and the study analyzed more than 90,000 proposals across six contemporary language models. LLM agents reached 26% of the human level on citation matching and 18% on idea matching, and the paper says AI agents rarely arrive at the concrete approaches pursued by human researchers. AI proposals trailed 2026 papers most in expected success and least in presentation, and proposals whose ideas no human paper shares fared worst in expert review. m:…3635

Two findings limit easy fixes. Neither richer field context nor larger populations substantially closed the performance gap, the paper says, though AI populations can cite as diversely as a human community. Taken together, this suggests agents are getting better at learning procedures, sharing reasoning and obeying checked guardrails, while still struggling to originate research ideas humans would actually pursue. For builders, that split matters: trust agents with watched procedures and verified tool calls before trusting them with open-ended invention. m:…3635