XferAgent: Cross-Agent Experience Transfer for Self-Improving LLM Agents
XferAgent tackles the problem of sharing reusable task experience across heterogeneous LLM agents. Instead of trapping effective strategies inside individual memories, it organizes local and shared experience graphs and reconciles cross-agent contributions with a similarity-margin merge rule that avoids unreliable unions. For each new task, it routes a connected, task-relevant subgraph and integrates it with local experience so agents benefit from collective exploration without sacrificing agent-specific context.
XferAgent is a competently engineered framework with a sensible graph-based design, but its empirical uplift is uneven. On GAIA with GPT-5.6 Luna it leads all baselines by roughly 4–5 percentage points, yet on the same benchmark with Qwen3-Next-80B-A3B-Instruct it edges AgentKB by less than a percentage point. The progressive-evaluation protocol tests cumulative recovery on repeated tasks rather than generalization to unseen tasks, which tempers the self-improvement claims.
The strongest aspect of the paper is the reconciliation rule for contributed experiences. The similarity-margin guardrail—merging only when the top candidate exceeds the runner-up by $\delta_{\mathrm{rec}}$—is a principled mechanism for handling heterogeneous agent inputs without forcing incompatible alignments. The multi-factor routing score that combines personalized PageRank graph relevance, exploration-boosted cross-agent outcomes, and contributor support is also well-motivated and clearly derived.
The evaluation contains serious confounds that weaken causal claims. First, the cross-framework transfer experiment conflates increasing agent count with increasing total training exposure, making it impossible to isolate whether gains come from framework heterogeneity or simply more trajectories. Second, the sensitivity analysis admits that varying two key parameters—$B$ and $\delta_{\mathrm{rec}}$—did not actually change logged routing decisions, rendering the reported accuracy swings for those knobs uninterpretable. Third, the progressive self-improvement results measure cumulative success over three passes on the same validation tasks, so gains may partly reflect task repetition rather than genuine experience evolution.
The evidence supports strong gains for GPT-5.6 Luna on GAIA, but the cross-agent advantage is much thinner elsewhere. On GAIA with Qwen, XferAgent scores 29.57% versus AgentKB’s 28.90% and No Memory’s 27.57%, a margin that could easily disappear with task variance. On ALFWorld the lead over the best baseline is only 0.4 percentage points, and on MBPP it is 1.7 points. No statistical significance tests or confidence intervals are reported, so readers cannot judge whether these differences are robust across folds.
The experimental protocols are documented in detail across multiple appendices, and frozen-memory evaluation is clearly defined. However, no code repository or prompts are released, the work relies on future model checkpoints such as GPT-5.6 Luna and Qwen3-Next-80B-A3B-Instruct, and many thresholds ($\theta_{\mathrm{edge}}$, $\theta_{\mathrm{rec}}$, $\delta_{\mathrm{rec}}$, $\beta_g$, $\beta_o$, $\beta_s$, $\lambda_u$) are tuned over a small 45-task sensitivity set with limited coverage, which would substantially block independent reproduction even if the models were available.
Large language model (LLM) agents can reuse experience from prior task executions to improve performance on subsequent tasks. Although existing methods enable cross-agent experience sharing, differences in agents' execution histories and task requirements make shared experiences difficult to reuse effectively. We propose XferAgent, a framework for cross-agent experience transfer that organizes reusable experience in local and shared experience graphs. XferAgent reconciles compatible contributions from heterogeneous agents while avoiding unreliable merges. For each new task, it selects shared experiences and integrates them with the target agent's local experience to support task execution. Experiments across diverse task domains show improvements over representative memory-based methods and effective experience transfer between heterogeneous agent frameworks. Cumulative task success also increases over successive task passes as agents update and reuse their experiences.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.