The Model as Its Own Teacher, and Why Splitting Crowds Wins

Featured

01 Training: This Teacher Isn't Stronger. It Just Knows the Ending.

Outcome rewards score once, at the end of a trajectory. Dozens of intermediate decisions have to split that single scalar between them, which makes token-level policy updates mostly guesswork. SEED fills that supervision void by hanging a dense token-level distillation signal inside the RL loop.

Where the teacher comes from is the part worth asking about. It isn't a larger model. It's the same policy model reading the trajectory it just finished and writing out "skills" in natural language — reusable procedures, the observation that decided the outcome, the failure mode to avoid. SEED feeds those skills back into the context and scores the same batch of actions twice, once with them and once without. That probability gap becomes per-token supervision, optimized jointly with the outcome reward. Self-evolving here means one model both collects and analyzes trajectories, so a stronger policy also writes better skills.

The teacher's edge is information, not capability. It has hindsight: it knows whether the trajectory succeeded and which step broke it. That draws a clean boundary on returns. If the failure was "didn't realize that was the move," a sentence of natural language can pull it back. If the model simply lacks the ability, no amount of hindsight conjures it — the same conclusion the on-policy distillation breakdown reached a few days ago. Distillation moves existing capability; it doesn't add new capability. The abstract offers "consistent gains" and better sample efficiency without hard numbers. Experiments cover both text and visual agent tasks, and the code is open, so running a baseline beats reading the abstract when you want to know if it's worth it on your own task.

Key takeaways: - SEED turns hindsight into a supervision signal: the same model reads its own trajectory, summarizes it as skills, and the with-skill/without-skill probability gap fills what the outcome reward leaves empty. No larger external teacher needed. - Before judging whether this helps you, sort your failures into "didn't think of it" and "can't do it." Hindsight only fixes the first. - Code is open and both text and visual agent tasks are covered, but the abstract gives only qualitative conclusions. Deploying it means running your own baseline comparison.

Source: SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning

02 Agent: Search Agents Spin in Place. A Bigger Model Won't Help.

Retrieval agent failure has a fixed shape. One search turns up no useful evidence, the model can't tell which direction to pivot, so it swaps in a synonym and searches again. Interaction history keeps growing while task progress gets harder to track, and the budget runs out with the answer still partial.

SearchOS puts the blame on progress that exists only inside conversation history — implicit, easy to lose, impossible to share across agents. Its fix moves progress outside the model. The task gets modeled as a relational table to fill: discover entities, complete fields, attach source evidence to every value. "Which cells are still empty" becomes a coverage query rather than an act of recall. The other half logs search strategies already tried and failed, then puts middleware between the model and its tools to watch for stalling and budget exhaustion. Interrupting turns into an external judgment instead of something the model has to notice about itself.

On WideSearch and GISA it leads single-agent and multi-agent baselines across every metric compared. How much of that comes from explicit state versus the parallel scheduling shipped alongside it isn't separable from the abstract. More throughput means more searches on its own, so the ablation matters before anyone draws conclusions. The design principle transfers regardless: pull "what's covered, what's already failed" out of context and into queryable external state. That lives in your harness layer, and it doesn't wait on a better model.

Key takeaways: - Circling happens because progress exists only in conversation history. Externalizing "covered" and "failed" into queryable state is a harness-layer change you can make yourself. - Stall detection belongs in middleware between the model and its tools. Don't expect the model to notice it's going nowhere. - The benchmark lead mixes in extra searches from parallel scheduling. What explicit state contributes on its own needs ablation data.

Source: SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

03 Robotics: RxBrain Puts Language and Imagination in One Sequence

Embodied planning has run on two tracks. Vision-language models (VLMs) decompose tasks in text and make decisions, world models predict future frames, and engineering glue connects the two. Tencent's previous embodied paper argued the efficiency case. RxBrain goes upstream of that, to how a plan should be represented at all.

It deletes the interface. Language and visual imagination alternate inside a single plan sequence, with language handling task decomposition, constraints, ordering, and decision logic while visual imagination anchors each step to a concrete physical state. The benefit is direct: every step of the plan already carries what the world looks like after execution, so nothing extra has to check that language decisions and physical states line up.

The cost is equally direct. One sequence means errors chain. Get an intermediate state wrong in imagination and the language reasoning after it keeps building on a false premise, and the abstract never says how that error gets bounded. It also reports real-robot results without large-scale action-data pretraining, but the word used is "promising," so the experiment scale in the full paper decides that claim. For teams building embodied systems, whether the single-sequence route survives accumulated error matters more than the numbers on a self-built benchmark.

Key takeaways: - One plan sequence carries both language and visual imagination, so every step aligns with a physical state and the VLM-to-world-model integration work disappears. - Chained error is the price. A wrong imagined intermediate state contaminates all downstream language reasoning, and that's the first question to ask about this route. - "Real-robot performance without large-scale action-data pretraining" is the claim most worth verifying. The abstract says promising, so don't read it as settled.

Source: RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination

04 Interpretability: Asking About Subgroups Beats Asking About the Whole

Intuition says splitting one population-level question into a dozen subgroup questions and reweighting should only compound the error. This paper cuts populations into finer and finer subgroups with a binary tree, feeds each to the model, and aggregates by prior weights. The result runs the other way: finer splits produce estimates closer to real human survey data. The authors call it the "macro fallacy."

The starting logic is plain. If the prompt genuinely specifies a condition, and the output genuinely estimates that conditional distribution, then partitioning the population any valid way and reweighting should recover the same overall distribution. Mainstream frontier models violate that identity across multiple domains. Subgroup-level knowledge is in there. It just doesn't propagate up when you ask about the whole.

What makes this self-check useful is that it needs no ground truth. Split the same question by age once and by region once, and the two aggregates should match. When they don't, what you're holding isn't a conditional distribution — it's the model's rough impression of how "that kind of person" answers. Teams doing synthetic survey research and persona simulation should run this before deciding whether to trust their numbers.

Key takeaways: - Run a consistency check before estimating population distributions with an LLM. Aggregate under several different partitions, and if the numbers disagree, don't feed them into decisions. - When you need overall proportions, take the fine-grained path. Asking subgroups and summing lands closer to real data than asking about the whole. - The test needs no ground truth and any pipeline can run it today, which also says frontier models are far from saturated on this dimension.

Source: Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models