LiveAsset turns user corrections into reusable checks for AI agents

Learning more than a correction

An AI agent can produce an unsatisfactory result even when it follows the literal wording of a request. Users often explain the problem without spelling out the solution. The challenge is to infer the missing expectation without treating every correction as a universal rule.

Guanming Liu and colleagues at Fudan University and Lenovo propose LiveAsset, a control layer around a language-model agent. It learns reusable “intent assets”: explanations of the user’s expectations, checkable requirements and relevant user-provided facts.

The motivation includes an analysis of real coding conversations from SWE-chat. From 1,000 sampled follow-up records, the researchers retained 331 corrective or clarifying messages. Of these, 135 identified a discrepancy without explicitly stating a desired requirement. This analysis informed the experimental feedback simulator; it was not a real-user evaluation of LiveAsset itself.

How the system works

LiveAsset examines three pieces of evidence together: the agent’s original result, the user’s correction and the revised result. Comparing them helps identify what the correction meant. As further corrections arrive, the system updates its assets while retaining earlier requirements that have not been contradicted.

The paper illustrates this with a product-review script. A user asks how a hair styler helps on busy mornings, then rejects an invented claim about minutes saved. The resulting asset preserves both expectations: explain the practical routine benefit, and ground personal timing claims in actual experience. The example illustrates the mechanism rather than an experimental result.

For a new request, LiveAsset selects relevant groups of intents. Before each reply reaches the user, a model-based checker assesses which requirements apply and whether they are met. It asks the same agent to revise unmet requirements while preserving satisfied ones, then checks again. Revision stops when nothing remains unmet or the number of satisfied requirements and facts stops increasing.

Importantly, checks occur after tool calls have taken effect. This is not a mechanism that guarantees prevention of an unwanted external action before it happens.

What the experiments measured

The researchers generated learning conversations using simulated users that progressively disclosed requirements and corrected the agent’s work. The simulator could access private task specifications and evaluation criteria, but these were not directly included in the interaction histories supplied to LiveAsset.

Testing then proceeded without those simulated corrections. Three benchmarks—π-Bench, PERMA and Vita—replayed original requests from reset environments. Two others, PAST-Bench and SkillEvolBench, tested transfer to changed tasks.

Comparisons included an agent without prior experience, complete interaction histories, reusable skills, Mem0 memory and TellOnce’s rule-based controls. Two agent models, GPT-6 Luna and DeepSeek V4.1-Flash, were evaluated. Each evaluation target was run twice, with independent scoring using GPT-6.1 Sol and benchmark-specific criteria.

Stronger compliance, uneven task gains

LiveAsset achieved the highest satisfaction of requirements exposed during learning on all three replay benchmarks with both models. With GPT-6 Luna on π-Bench, satisfaction reached 92.5%, compared with 73.8% for TellOnce and 80.1% when the full interaction history was supplied.

However, satisfying learned requirements and completing the whole task were distinct measures. On PERMA with GPT-6 Luna, LiveAsset achieved 95.4% requirement satisfaction, but its task-completion score was 85.7%, below the full-history approach’s 92.9%.

Transfer results also require qualification. LiveAsset beat the skill, memory and rule-based baselines on overall scores in both transfer benchmarks with both models. It did not win every comparison: full history performed better on PAST-Bench with DeepSeek, and the skill baseline led one composition-test category with GPT-6 Luna.

Removing explanations, surrounding interaction context or runtime correction reduced replay performance, supporting the contribution of each component within these experiments.

Limits and research status

The learning feedback was simulated, so the results do not establish effectiveness with real users whose preferences may be inconsistent or change over time. Model-based checking also remained fallible: some outputs passed internal checks while still failing external requirements.

The additional control comes at a computational cost. Across the three replay benchmarks, GPT-6 Luna runs averaged approximately 206,000 tokens with LiveAsset versus 154,000 for the base agent, including auxiliary calls. These token totals do not establish response latency or monetary cost.

The supplied work is an arXiv preprint, not established peer-reviewed validation. Its central finding is promising but bounded: contextual explanations plus repeated checking improved reuse of corrections in the tested settings, without guaranteeing correct intent inference or flawless execution.