Repair the Core, Not the Symptom: Planning-Graph Repair for Long-Horizon Agents

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Repair the Core, Not the Symptom".

Jane: Here's the gist of the paper. The gist The study proposes WM-SAR, a spectral subgraph repair method that estimates node-edge amplification to select a compact region for LLM repair, which achieves stronger longhorizon stabilization and root-cause recovery under compact token budgets than engineering correctors.

Tom: Problem Formulation Long planning naturally becomes a graph, and replanning the whole graph after each error is not viable because it consumes context budget and forces the LLM to retrieve irrelevant causes The need for world-model correctors arises because when a failure appears, the corrector must decide which part of the graph to revise.

Jane: A strong default is an engineering LLM corrector, which can be simple and model-agnostic but often misses small subgraphs that re-amplify error.

Tom: Engineering Baselines A family of engineering correctors includes pointwise error scans, TopK and window selection, local graph expansion, cascade repair, and full-context LLM repair These methods optimize variants of Pv∈R e(v), which is a useful attribution signal but not a stability objective.

Title and authors: Tom: So we've looked at the mechanism behind WM-SAR; now let's talk about what this paper is actually called and who did it. The paper is titled "Repair the Core, Not the Symptom: Planning-Graph Repair for Long-Horizon Agents."

Jane: That title really sets the tone, suggesting that fixing just the symptom isn't enough; you need to fix the core structural issue causing problems.

Lu: The authors are Xinyuan Song, Zekun Cai, and their team from Emory University and The University of Tokyo. They're working on connecting world models to planning graphs.

Meng: I see they are focusing on how these planning graphs handle failures when the context budget is tight, which is a real production problem for me.

Tom: It’s about showing that instead of using expensive full graph replanning after every error, we can select a much smaller region to fix.

Jane: The implication here is that we can achieve high horizon stability without paying the massive context cost associated with repairing much larger neighborhoods, which is what those other methods end up paying for later on <ref:2607.01767#pg3>.

Lu: They introduce a strong family of engineering correctors as a baseline, including pointwise error scans, TopK and window selection, local graph expansion, cascade repair, and full-context LLM repair.

Meng: That's what we see in the existing work; they are trying all the standard ways to fix things locally before proposing their method.

Tom: And they show that while those engineering correctors can work on small graphs, they have a limitation: they spend budget by scanning visible evidence and might miss the small subgraph that re-amplifies error.

Jane: That's the problem they are trying to solve; those methods often fail because fixing one thing doesn't stop the error amplification path from becoming active again later on.

Lu: So, this paper is essentially positioning WM-SAR as a way to move beyond those local fixes by optimizing for actual stability metrics like rho(B) -reduction, rather than just an attribution signal.

Meng: If we can deploy this kind of repair mechanism efficiently in production, it means agents could stay stable much longer while maintaining lower token usage for those recovery steps.

Tom: So they are moving the optimization goal from simply finding where the error is most visible to finding where the structure is most unstable over time.

Jane: And that difference matters a lot when you're dealing with long horizon tasks where small initial errors can snowball into big failures later on.

The paper's summary: Tom: Now let's talk about how they actually put this into practice in the summary of "Repair the Core, Not the Symptom: Planning-Graph Repair for Long-Horizon Agents." They use WM-SAR by using GEAF times kappa seeding and growing that residual relief term rho rel to produce a compact causal context for LLM repair.

Jane: The main shift they are showing is in the repair objective itself, changing the focus from which single node has the most error to identifying which connected subgraph, when repaired, most reduces future error growth over time.

Lu: They achieve this by implementing this objective with GEAF times kappa seeding and rho rel growing to produce that compact causal context for LLM repair by finding a region R such that R K rho(B G f R) connected.

Meng: So they are defining this specific causal region by looking for the one where we have the minimum residual amplification, defined as R V: G fR connected, R K rho(B G f R).

Tom: Essentially, they are using this criterion to find that specific causal region that actively lowers residual amplification through a mathematical definition.

Jane: It’s a solid piece of work showing how to achieve strong long-horizon stabilization and root cause recovery under tight token budgets when using WM-SAR.

Lu: They also provide some future direction, though they note that their uncertainty term Unc(v) has no effect at their default hyperparameters on the synthetic dataset, so datasets with meaningful uncertainty signals are needed to validate that component.

Meng: I agree; needing more data where uncertainty is actually meaningful is what will help them fully unlock the potential of this method for real-world deployment because those scenarios need that signal.

Tom: So to wrap up this summary, "Repair the Core, Not the Symptom" shows how they build repair methods based on structural stability rather than just local error scores in agent systems.

Jane: It’s a paper that clearly shows a path toward stabilizing those long-horizon agent systems without needing expensive full graph replanning.

The paper's improvements: Tom: Let's look at the specific improvements they highlight, because it’s not just one thing; it’s several ways they suggest better approaches to planning-graph repair. They show we can replace visible error scans or local graph expansion with something that targets the amplifier itself.

Jane: It really boils down to moving away from just fixing a single high-error node toward selecting a small connected region whose repair stabilizes future rollouts, which is much more robust for long tasks.

Lu: The core idea is finding that specific connected path whose removal lowers the residual amplification, which flattens the error curve over a sustained planning horizon H.

Meng: From my side, it’s appealing because they show you can get that high relief regime using only a small number of nodes, averaging eight point three, without paying the massive context cost of full graph LLM repair <ref:2607.01767#pg3>.

Lalam: I’m confident in this direction; the paper grounds this repair objective in a mathematical concept—the residual spectral radius—which gives us a solid way to control future rollout error.

Tom: So, WM-SAR uses GEAF times kappa seeding and growing that rho rel term to produce a compact causal context for LLM repair, which is what makes it stronger than the engineering baselines they analyzed.

Jane: That’s the main mechanism they use to achieve this compact context, which is what makes it stronger than the engineering baselines we usually see when we compare them.

Lu: This connects the residual spectral radius directly to the idea that this method controls future rollout error, which is what we need for long-horizon agents.

Meng: It’s about finding that specific causal region that actively lowers residual amplification through a mathematical criterion, which is what makes it stronger than the engineering baselines they usually see.

Tom: So to wrap up these improvements, "Repair the Core, Not the Symptom" shows how we build repair methods based on structural stability rather than just local error scores in agent systems.

Jane: It’s a solid piece of work that shows how to achieve strong long-horizon stabilization and root cause recovery under tight token budgets with WM-SAR.

Lu: We still need to see validation on datasets with meaningful uncertainty signals because the method's uncertainty term doesn't have an effect at their default hyperparameters on the synthetic dataset.

Conclusion: Tom: So to wrap up, "Repair the Core, Not the Symptom: Planning-Graph Repair for Long-Horizon Agents" shows how we can build repair methods based on structural stability rather than just local error scores in agent systems.

Jane: It really boils down to moving away from just fixing a single high-error node toward selecting a small connected region whose repair stabilizes future rollouts.

Lu: The core idea is finding that specific connected path whose removal lowers the residual amplification, which flattens the error curve over a sustained planning horizon.

Lalam: I’m confident in this direction; it shows a path toward stabilizing those long-horizon agent systems without needing expensive full graph replanning.

Tom: So we've seen that WM-SAR uses GEAF times kappa seeding and growing that rho rel term to produce a compact causal context for LLM repair.

Jane: That’s the main mechanism they use to achieve this compact context, which is what makes it stronger than the engineering baselines we usually see.

Lu: It’s about implementing this objective with GEAF times kappa seeding and rho rel growing to produce that compact causal context for LLM repair.

Meng: The big improvement they highlight is that you can perform more efficient debugging by selecting a causal region that actively lowers residual amplification through the criterion R V: G fR connected, R K rho(B G f R).

Tom: So we’ve seen that it’s about finding that specific causal region that actively lowers residual amplification through a mathematical criterion.

Tom: So if you're interested in how we move from local error fixes to structural stability, we'll be looking at some papers on graph learning techniques next.

Jane: Thanks for joining us today, and we’ll be back next time to talk about whatever new research is coming out on arXiv.