The Robot World Model That Only Imagines Success — and How DreamTrue Curbed Its Optimism
A robot world model trained only on success keeps predicting success even when the gripper misses. DreamTrue post-trains it to imagine failures and grades them, cutting its own early version's interaction-defect rate from the 48% range to the 6% range.
The Robot That Left Its Model Alone and Fixed Everything Around It — PhysEvo
Do you rewire a genius's brain, or hand the genius better tools? A robotics project froze the model that drives the robot and rebuilt only the surrounding system from its own failure logs — and the exact same model's success rate climbed from 26% to 68%.
Should German RAG Think in German? The Payoff of Reasoning-Language Alignment
In a RAG setup where the question, the documents, and the answer are all German, aligning the model's thinking language to German too beat another forced language (French) on accuracy. But it never got past the model's native English reasoning (measured on a lightweight model).
The Agent That Improves Itself by Rewriting Its Own Rulebook — A Close Reading of Memento 3
Memento 3 learns a world model by editing only the "rulebook + code" in its external memory, leaving the underlying LLM frozen. Here is how it cleared all 25 public ARC-AGI-3 games, and what it implies for ONTLE's AI PM.
Leave the Brain Alone, Just Rewrite the Sticky-Note Memories — EngramEdit
How to fix a fact an LLM has memorized without retraining its brain. You leave the reasoning brain untouched and only touch up the memory sticky-notes, so the updated answer shows up even when you phrase the question differently — but it only works on models built with a particular memory architecture.
From Evidence to Action: Where Tool-Using AI Agents Break Down
A correct outcome isn't enough if the agent never secured the grounds for its action beforehand. Across the 656-task SafeActBench benchmark, failure often starts before execution—in the investigation that should ground an action—though grounded-execution errors and unfinished multi-step actions also appear.
Testing AI With Problems Written by AI: AutoSciBench
Agents are acing the test so fast that benchmarks go stale almost overnight. AutoSciBench watches how agents solve problems and uses what it sees to crank up the difficulty on its own.
Filling Every Gap Made It Worse — The Paradox of Cross-Tokenizer Distillation
In cross-tokenizer distillation, adding a span-probability loss that fills every comparison gap lowered accuracy in all 18 weighted configurations tested. What mattered wasn't coverage — it was the reliability of the signal.
Low-Precision RL, 5.4x Faster Without Losing Quality — Alibaba's TRACE
Alibaba's TRACE directly shrinks the rounding mismatch between training and rollout, holding BF16-level quality under FP4 reinforcement learning while boosting rollout throughput by up to 5.4x. Measured on four MoE models.
Make It Choose Before It Acts — The Real Bottleneck for AI Agents Was Selection
Sample several candidate commands and pick exactly one before running anything. What decided success wasn't how many candidates you drew, but how good the verifier doing the picking was.
Translating a Full Season, Fixing Itself as It Goes: How SMART Nailed Subtitle Consistency
Translate a full season of a drama and the same term of address keeps slipping from episode to episode. Without retraining the model, SMART rewrites its own working process mid-translation and took first place across all 15 translation directions.
When the Dish Flops, Whose Fault Is It? FloWright Trains Every Role Together
Multi-agent workflows have long trained only the designer. FloWright splits the failures in an execution trace by role and reinforcement-learns every role together, lifting a 4B model's accuracy by an average of +5.03 percentage points.