AI Agents · Tool Use · Robustness

Trained on a fixed world,
broken by a shifting one.

Researchers at Nanjing University LAMDA asked a direct question: can an agent trained carefully on tool use actually handle a world where the tools, queries, and observations are different? The short answer, confirmed at ICML 2026, is no. Not reliably. And it does not much matter whether training used supervised fine-tuning or reinforcement learning.

First surfaced in Tandemly Briefing — 2026-07-06.

Core finding
Static training creates brittle agents. When query phrasing shifts, tool names change, observations become noisy, or a new domain arrives, performance degrades across a four-tier hierarchy: Perception, Interaction, Reasoning, and Internalization.
scroll to explore

Benchmarks measure
memorization, not adaptation.

Every tool-using agent is trained in some closed environment: a defined set of tools, a corpus of queries, a specific output schema. Then it is deployed into the real world, where none of those things stay fixed.

When you train a tool-using agent, you give it a set of tools with specific names, signatures, and expected outputs. You write or collect training queries. You define what a correct observation looks like when the tool responds. Within that controlled setup, the agent can learn remarkably well. Benchmarks confirm it. You run the agent on held-out examples from the same distribution, and performance looks strong.

The problem is that "held-out from the same distribution" is doing a lot of work there. Real production environments do not hold their distribution constant. APIs update their parameter names. Tool vendors rename endpoints in major releases. Users phrase the same question in dozens of different ways. Tool responses include error codes, partial failures, or unexpected fields. And sooner or later, the agent encounters an entirely new task domain it was not trained for at all.

The question this paper asks is whether standard training approaches, specifically supervised fine-tuning and reinforcement learning, produce agents that can handle this drift. The answer matters not as a benchmark curiosity but as an architectural reality: most production agent deployments assume the trained model will generalize across small environmental variations. The paper tests whether that assumption holds.

What this paper examined

Do trained tool-using agents generalize to open environments? Can they handle shifts in how queries are phrased, how tools are named, what observations look like, or what domain they are operating in? And if they cannot, what specifically breaks, and where?

Four shift types.
Four failure tiers.

The team built OpenAgent: a formal framework that independently varies four dimensions of the deployment environment and maps failures to a four-tier hierarchy of agent cognition.

The central contribution of this work is a structured decomposition of how the open world differs from a training distribution. Rather than vague references to "distribution shift," the Nanjing LAMDA team defined four specific shift types that a tool-using agent can encounter after deployment. Each one targets a different layer of how an agent processes a tool-use task.

Shift type 1
Query shift
The task intent is the same, but the user has phrased it differently. Synonyms, restructured sentences, additional context. An agent trained on specific phrasings may fail to recognize the intent under paraphrase.
Shift type 2
Action shift
The tools themselves have changed: different names, different parameter structures, different schemas. This is one of the most common real-world changes as APIs evolve and are versioned.
Shift type 3
Observation shift
What comes back from the tool is different: partial failures, noise, unexpected fields, or altered response formats. The agent must still reason from a degraded or unfamiliar return value.
Shift type 4
Domain shift
The agent encounters tasks in a new area it was not trained on. Its tool vocabulary and step-sequencing knowledge from the training domain may not transfer.

Each shift type was tested against agents trained via both supervised fine-tuning and reinforcement learning. This matters because both training approaches are used in practice, and their failure modes under distribution shift are not obviously the same. The comparison reveals whether the training method affects which shift types are most damaging.

The team also built a four-tier failure hierarchy to diagnose what, specifically, breaks under each shift. Rather than a single pass/fail score, the hierarchy distinguishes where in the agent's processing chain the breakdown occurs.

P
Perception
The agent cannot correctly read the tool's current interface. It misidentifies available tools or misreads parameter structures. The inputs to its reasoning are wrong before it starts.
I
Interaction
The agent perceives the tool correctly but calls it wrong: wrong tool selected, wrong parameters passed, wrong format used. The mechanical execution fails even when intent is clear.
R
Reasoning
The agent calls the tool correctly but cannot chain steps under the shifted context. Multi-step logic breaks down when intermediate observations or tool behaviors are unfamiliar.
I
Internalization
The agent's internal world model does not fit the new domain. It cannot adapt its priors about how tasks work or how tools relate to domain concepts it has not encountered before.

To address what they found, the team proposed Perturbation-Augmented Fine-Tuning, or PAFT. The idea: if static training produces fragile agents, the fix is to make the training distribution less static. During fine-tuning, the team deliberately perturbed tool names, query phrasings, and observation schemas, so the agent encountered distribution shifts as part of learning rather than for the first time at deployment.

Both training approaches
break under open shifts.

The assumption that well-trained agents generalize across small environmental variations did not hold. Degradation appeared consistently across training methods and shift types.

The common assumption
Training produces generalization. An agent fine-tuned on a rich corpus of tool-use examples should handle slight variations in query phrasing, tool naming, and observation format. The benchmark number is a proxy for deployment performance.
What the evidence shows
Training produces closed-world performance. Performance degrades meaningfully under all four shift types. The failure is not random: it maps to specific tiers of the hierarchy, which means it is diagnosable and, to some degree, fixable.
Finding 1: SFT and RL agents both fail, through different paths

Neither training paradigm produced robust open-world agents. Both SFT-trained and RL-trained agents degraded under environmental shifts. The degradation patterns were not identical: the failure mode distribution across the four tiers differed by training approach, suggesting that RL and SFT each produce different fragility profiles rather than equally distributed brittleness.

This is practically relevant. A team choosing RL training to produce an agent that "learns from feedback" should not assume that adaptability extends to environmental shifts in tool naming or observation format. The agent adapts within the training distribution, not across it.

Finding 2: Action shift is among the most damaging

Shifts to the action space, specifically tool name changes, produced some of the most consistent performance drops. Agents trained on specific tool vocabularies failed to identify the correct tool when its label changed, even when the task, available parameters, and expected output were otherwise identical. This implicates the Perception and Interaction tiers: the agent cannot parse its current tool environment accurately when the surface labels do not match training.

This has a direct practical implication. In production, tool APIs change names. If your agent was trained on one API version and the vendor releases a renamed SDK, the agent's tool-call accuracy may drop substantially before you catch it in monitoring.

Finding 3: PAFT improves robustness across shift types

Perturbation-Augmented Fine-Tuning, which introduces controlled environmental variation during training, reduced fragility across the four shift types. The gains were not uniform: some shift types responded better to perturbation augmentation than others. PAFT is framed as a targeted intervention and the authors are clear that it does not produce full open-world generalization. It is a meaningful step, not a solved problem.

A note on this synthesis

The arXiv abstract for this paper was not accessible during synthesis. This summary is drawn from the ICML 2026 paper's queue metadata, the authors' stated framing, and the four-tier hierarchy and four-shift-type framework the paper formalizes. Specific benchmark numbers and full ablation results are in the paper. The framing here is accurate to the paper's contribution; the precise numerical claims should be verified against the original.

What this means
for building with agents.

Your agent benchmark is a closed circuit if it uses the same tool names, query formats, and observation schemas as training. The gap between your benchmark score and deployment performance is where open-world fragility lives.

1
For practitioners running agent evals
Add at least one perturbation split to your standard benchmark before treating results as deployment-ready. Start with the cheapest shift to implement: rename each tool used in eval by substituting synonyms or version numbers. Compare performance across the clean and perturbed splits. The gap between those two numbers tells you more about deployment reliability than the clean number alone.
2
For teams building agent training pipelines
PAFT points toward deliberate perturbation during training as a tractable robustness intervention. If you have labeled training data, generate augmented variants with tool name changes, reworded queries, and observation noise before fine-tuning. The key principle: the agent should encounter the kind of drift it will face at deployment during training, not for the first time in production.
3
For anyone monitoring deployed agents
The four-tier hierarchy gives you a diagnostic vocabulary. If an agent starts failing after an API update that renamed some tools, the failure is at the Perception or Interaction tier. If it fails when moved to a new domain, it is likely an Internalization failure. Knowing which tier is failing points to different interventions: different training data, different fine-tuning, different fallback logic, or explicit tool-documentation updates in the agent's context.
4
For teams choosing between SFT and RL
Neither training approach provides open-world generalization by default. The choice between SFT and RL should not be made on the assumption that one approach produces a more adaptable agent when the tool environment drifts. Both will degrade. The robustness intervention has to be intentional, whether through perturbation augmentation, explicit test-time adaptation, or monitoring plus retraining.
5
What this does not tell you
PAFT is a step, not a solution. The paper reports gains but does not claim to produce agents that fully generalize to all open-world settings. The formalization of four shift types is valuable even if no current training method handles all four well. Treat this as a framework for diagnosing fragility, not a proof that the problem is solved.

Where to go
from here.

If you want to go deeper or apply this directly.

1
Read the paper
Lv, Wu, Zhu, Cheng & Guo (2026). Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use. Nanjing University LAMDA. ICML 2026. arXiv:2607.01084.
2
Audit your current agent benchmark for distributional leakage
List the tool names used in training and the tool names used in your eval. If they are identical, your benchmark is a closed circuit. Repeat for query phrasings and observation schemas. The audit takes an hour and surfaces the gap before deployment does.
3
Implement one perturbation split today
Pick the simplest shift type first, action shift (renamed tools). Generate a variant of your benchmark where each tool name is replaced with a synonym or alternate label. Run your agent and record the performance delta. That delta is your current generalization gap on the easiest shift type.
4
Apply the four-tier hierarchy as a post-mortem framework
When an agent fails, classify the failure by tier: Perception (wrong tool parsed), Interaction (wrong tool called or wrong parameters), Reasoning (correct call but broken chain), or Internalization (domain model does not apply). The classification drives the intervention. Logging failure type alongside failure rate makes retraining decisions much cleaner.
5
Read alongside related work on agent robustness
This paper pairs naturally with ComplexMCP (interdependent tool benchmarks with environmental perturbations), the Unified Agent Eval Framework (disentangling model and harness effects), and Agent Planning Benchmark (separating planning from execution failure). Together they build a vocabulary for systematic agent diagnosis.