Two papers from Yann LeCun's group press the same point from opposite sides this week: training objective matters more for a model than size. One paper has robots planning in layers, while the other has a small text model aligning with vision better than far larger language models. The remaining papers cover agents that operate over long periods, plus the cost of letting them reason and remember without checks.
Layered planning raises Visual AntMaze success from 18% to 73%
H-JEPA, from Yann LeCun's group, builds a stack of world models where each level looks farther forward in a separate simplified view of the scene. Planning in H-JEPA runs top down, with the upper level pushing toward the goal and passing predictions down as subgoals for the shorter horizon below. Upper H-JEPA levels shed fast unpredictable detail and hold the slower state that guides the task.
Across four simulated navigation and manipulation environments the H-JEPA hierarchy beat a flat JEPA. H-JEPA raised success on Visual AntMaze from 18% to 73% with a three-level stack and less planner compute, and with inverse-dynamics supervision extended the approach to real-robot videos from DROID with better offline planning fidelity at lower planner compute. For builders of robots or agents that act over long horizons, H-JEPA gains came from structure rather than a larger planner.
Small text model matches vision models on world representation
Text-trained models and image-trained models appear to converge on a shared inner view of the world as scale grows, a claim called the Platonic Representation Hypothesis. Randall Balestriero and Yann LeCun state it as a theorem: JEPA-family models trained apart on separate modalities reach the same representation, save for one rotation, when each pair contains two views of one scene. Next-token prediction follows the same pattern with weak pairs, and the authors present that weakness as the reason language models meet vision only at scale.
Good views demand 42 times fewer training samples for equal alignment, and a 109M-parameter text JEPA matches the DINOv2 and CLIP image encoders better than any language model in the comparison, to 72B parameters. A single rotation turns the separately trained encoders into an image captioner. Should the finding stand, cheaper multimodal models follow from better pairs rather than added parameters.
Agents fail less when lessons link back to evidence
An agent that keeps notes from past tasks can learn the wrong lesson and then fail jobs solvable with no memory at all. WikiTrace stores each lesson as a short insight linked to a summary of why the attempt succeeded or failed and to the raw trajectory. When a lesson is vague, clashes with another or comes from a failed task, the WikiTrace agent follows links with terminal tools to fill in or correct the insight before acting.
On five agent benchmarks WikiTrace beat existing systems, including a version given the same lessons without links, while cutting the failures that memory itself causes. For builders shipping agents with long-term memory, the WikiTrace rule keeps evidence beside the lesson. The code is public.
Reasoning models shorten answers while holding accuracy steady
Large reasoning models often continue deliberating after gathering enough information to respond, and each surplus token adds inference cost. LoCo locates overthinking points inside a model's own chain of thought, freezes the prefix, creates shorter verified continuations from that point and trains on the resulting pairs through direct preference optimization, teaching the model when to keep reasoning and when to conclude. Built only from self-generated mathematical traces, LoCo shortens completions by 41.7% to 45.4% on mathematics and by 23.4% to 36.3% on code despite math-only training while holding or lifting accuracy, a direct saving on the inference bill for providers serving a reasoning model.
Agents keep improving infrastructure over a full day
Infrathon measures whether agents can sustain gains on real AI infrastructure across extended runs. Six models completed 150 runs under 24-hour budgets for more than 10,000 GPU-hours in total, and 135 of the 150 runs posted further gains during the second half of the budget. Model order changed with evaluation length, and briefer checks can miss later gains and select the wrong agent.
Claude Opus 5.5 lifted the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%. Barriers to sustained improvement were task violations, unreliable testing and premature completion.





