How can continual agent merging acquire new expert behavior without forgetting?

Merging experts one at a time risks erasing earlier skills

Batch agent merging assumes every specialized expert is available at once, so the merged model can be built in a single pass. Continual agent merging instead receives experts sequentially and must fold each new one into the current merged model without keeping the earlier experts around. The paper identifies two coupled difficulties in this setting: a merge in parameter space or a feature subspace does not guarantee that the merged model actually reproduces the incoming expert's behavior on interaction trajectories, and each update toward the new expert can disturb interactive behavior that was already integrated.

The trajectory point is what makes agents different from ordinary task models. In a multi-turn interaction, the merged model's own actions determine the observations it sees next, so its path can diverge from the expert's path. Copying an expert's weights or task vector therefore does not ensure the expert's decisions appear at the states the merged agent actually visits.

Gates are learned from the merged model's own rollouts

BACAM's core move is to learn parameter-wise merging gates from candidate-generated trajectories under expert-guided behavioral supervision. A gate is a per-parameter value in [0,1] that controls how strongly the incoming expert's parameters are blended into the current model. Because the training states are decision points extracted from trajectories the candidate model produces, the supervision targets the behavior the merged agent will actually need, not behavior observed only on the expert's own path.

The paper reports that training inputs are unlabeled and that teachers supply token-distribution targets, using a Top-32 representation with distillation temperature 1.0. After a fixed number of refresh rounds, the learned gates are folded into the model weights, so no extra inference-time parameters or computation remain. Each stage uses three rollout-refresh rounds, with 256 training states per task stream per round and 24 probe trajectories per task stream.

Two budgets separate learning from interference

The method adds two controls that operate at different levels. Task-level stability-plasticity control governs how much the merge may move while still satisfying the old-task objective, using a stability budget epsilon of 5×10⁻³. Tensor-level conflict-aware update budgets then act on individual weight tensors: a conflict score combines the direction of the old-task gradient with gradient magnitudes, producing a budget between 0 and 1 that shrinks when an update opposes an old task and that old task is sensitive to the change.

The budget is applied only to positive gate updates, which move parameters toward the new expert; negative updates, which move the gate back toward the current model, are left unrestricted. Taking the minimum budget across old tasks gives each tensor its most restrictive limit. The paper is explicit that these budgets describe optimization behavior and are not a direct measurement of a tensor's importance for task success.

Success-rate gains and what they do not mean

Across web shopping, tool use, information retrieval, and embodied interaction, BACAM reports an average success rate of 62.82%, exceeding the strongest evaluated merging baseline by 21.69 percentage points. That average weights the five reported success-rate columns equally, counting ALFWorld in-distribution and out-of-distribution separately; it is not a pooled success rate over instances, which would weight the larger search evaluation set much more heavily.

The metric definitions matter when reading the table. Success rate requires full task completion, while WebShop's task score gives partial credit for satisfied shopping constraints, so a high task score with lower success rate can mean the agent found a suitable product but missed a remaining constraint. Interaction length (Round) is also not a performance measure on its own: failures stay in the average, and a shorter interaction can reflect early termination rather than efficiency.

Order still changes how much survives

The paper's own order experiments show that continual merging is not order-invariant. In the WTSA order (WebShop, Tool, Search, ALFWorld), WebShop success rate moves from 78.00 after the first stage to 77.20 after all four, and ALFWorld reaches 88.57 in-distribution and 86.57 out-of-distribution. In the ASTW order, ALFWorld starts at 98.57 in-distribution and ends at 96.43, but WebShop only appears at the final stage and reaches 63.00. In TWAS, WebShop falls from 67.80 at stage two to 56.80 at stage four, and ALFWorld drops from 88.57 to 84.29 in-distribution.

These stage-by-stage numbers are the clearest evidence that the arrival order shapes which capabilities survive. The authors name robustness to expert merging order as a key direction for future work, which places the 62.82% headline figure inside a specific default four-expert order rather than establishing order-independent behavior.

Where the behavioral supervision shows up

The qualitative examples locate the gain at decisive interaction steps. In a WebShop case, BACAM recognizes that the required 52-by-84-inch size is still unselected and chooses it before purchase, scoring 1.00, while WUDI-Merging reselects the color and omits the size for 0.80. In a tool-use case, BACAM resolves the company ticker before retrieving stock information, whereas the comparison treats the request as a history lookup. In search, BACAM resolves a country clue before searching for its largest national park; in ALFWorld, it checks another location for the lamp instead of repeating invalid commands at the desk.

These examples are illustrative rather than a measured breakdown of errors, but they match the method's premise: the gate is trained on states the merged model generates, so the behavior it learns is the behavior needed at the merged model's own decision points.

What the 62.82% average actually decides

The practical reading is a priority, not a blanket claim. BACAM's advantage is largest when experts arrive sequentially and cannot be retained, and when the merged agent must act correctly on its own multi-turn trajectories rather than merely hold expert-like weights. The 21.69-point gap over the strongest evaluated baseline is measured on the paper's four-task average under its default order, with baselines that differ in whether they use task instances, first-turn prompts, or teacher outputs, and with data-collection and update schedules that are method-specific rather than budget-matched.

Within that scope, the result that should change a reader's comparison is the combination of the 62.82% average with the per-order stage tables: BACAM preserves earlier capabilities far better than the evaluated continual and general merging baselines in the reported orders, but the size of what survives still depends on when each expert arrives. Order robustness, not the headline average alone, is the open question the paper leaves for anyone deploying sequential expert merges.

Sources used in this answer