Scientific feedback can rewrite weight-editing plans: HypoLoop cuts Burgers relative error from 0.040194 to 0.016559
Synopsis
Scientific prediction models keep exposing new errors as experiments proceed, yet weight editing has typically refit the same parameter subset with the same objective. HypoLoop lets experimental feedback revise both which weights change and how errors drive the correction: fixed, numerically adaptive or LLM-guided policies propose an editing plan, numerical solvers compute the weight patch, and internal validation selects the deployed predictor. On Burgers, reallocating a matched editable-coordinate budget alone lowered relative error from 0.040194 to 0.016559 under DeepSeek and from 0.030975 to 0.016195 under Claude.
Interpretation
Weight editing no longer keeps one fixed set of parameter locations and one fitting objective; each round of experimental feedback re-decides where to edit and how to fit. Earlier practice repeatedly fit the same parameter subset with the same objective as feedback accumulated, so both choices could go stale as new error patterns appeared. On Burgers, dynamic and fixed editable locations shared the same proposal method, numerical fitting procedure and acceptance rule with a fixed editable-coordinate budget, and relative error fell from 0.040194 to 0.016559 under DeepSeek and from 0.030975 to 0.016195 under Claude.
Each round hands the model structure, error statistics from returned observations and the editing history to a proposal policy that emits an executable editing plan. The plan has two parts: a set of editable native coordinates, and a fitting configuration specifying sample weights, loss, regularization, reference preservation and patch scale. A numerical solver fits the weight patch from that plan, and internal cross-validation on experimental-group or temporal folds evaluates candidates; only candidates meeting the cross-validation improvement threshold and the allowed loss change on accumulated feedback are deployed, otherwise the original weights are kept.
At the same coordinates, the fitting rule alone can change the direction of the patch. DIODE feedback showed systematic depth overestimation, so the accepted plan shifted weight toward positive residuals. With the same 33 editable coordinates, smooth MAE, ridge coefficient 0.01 and zero preservation penalty, uniform weighting gave 0.894701 and positive-residual weighting gave 0.775221, a 13.35% drop in internal cross-validation AbsRel; reversing the residual emphasis raised error to 1.437018.
Perspective
The demonstrated scope is 22 scientific tasks, each with its own pretrained model, metric and feedback stream, evaluated separately under DeepSeek and Claude; methods are compared with Base, Adam, Memory and Calib under the same feedback, and matched-budget location comparisons are run on Burgers and QP.
Results are mixed across tasks: ESOL RMSE falls relative to the strongest feedback baseline under DeepSeek but rises under Claude, and on some tasks HypoLoop family members trail the baselines. Several relative-change figures in Table 1 are not given in the prose, and the prose description of Burgers reallocation reports only the two services' error values.
