We Asked Opus-5.5 to Edit π_0.5’s Weights. AutoResearch Made It Happen.

1 Our Idea and Research Setup#

We asked whether a Frontier Model could understand the parameters of a trained robot policy and, through a small number of edits, make the robot execute actions and complete tasks faster. We gave this problem to Opus 5.5, asking it to derive weight changes from mechanisms and objectives.

We provided the objective and environment; Opus 5.5 proposed hypotheses, designed experiments, and selected a solution. It worked through Claude Code with one NVIDIA GeForce RTX 4090, a trained ${\pi }_{0.5}$ [2] Training-Time RTC checkpoint, model and inference code, and recorded YAM demonstrations and rollouts. Opus never operated the robot or received online physical feedback; its research used only existing recordings. We conducted the subsequent physical tests on YAM arms. It surveyed 136 historical rollouts (5 task types) and 154 human demonstrations (3 task types). For editing, it selected a separate set of 12 demonstrations (6 task types) and 12 rollouts (4 task types) from the available data, covering coin insertion, moving-block picking, stacking, bottle placement, and clock manipulation.

We prohibited backpropagation, gradient-based optimizers, and LoRA, while allowing forward-pass experiments and closed-form fitting. We asked it to change as few parameters as possible, preserve the production inference and control programs, and deliver an independent checkpoint that could be rolled back. A 20%–30% reduction in task time was the exploratory target; a stable 5%–10% improvement would also be meaningful. It needed to check smoothness and cross-task behavior, and compare against the original policy, action-amplitude scaling, trajectory time scaling, and random weight edits.

The resulting method developed through mechanistic hypotheses, controlled experiments, and revisions. We are interested not only in performance after editing, but also in how Opus identified problems, tested explanations, and changed direction after failure. Section 2 presents its research process; Section 3 reports our additional real-robot and mechanistic experiments. We close by considering what this case suggests about AutoResearch in robotics.

Here, an action chunk is a policy prediction of 50 future action steps. Each action step contains 12 joint targets for the two arms and 2 gripper targets. The baseline action execution rate is 30 Hz; Section 3.1 also tests 35 and 40 Hz. We execute 32 steps per cycle. When updating the plan, the first 6 steps of the old plan serve as conditioning actions, or the RTC prefix. We call an abrupt change in target position at a plan switch a chunk jump; this describes a change in commands, not an instantaneous movement of the physical arm.

2 Opus 5.5’s Hypotheses, Experiments, and Solution#

2.1 Locating the Opportunity: Inference Waiting or Action Execution?#

Opus 5.5 first checked whether task duration was dominated by waiting for model inference. It read the deployment configuration and compared inference times, action commands, and measured joint states across 136 historical RTC trajectories. Each inference took approximately 105–109 ms, substantially less than the roughly 1.073 s plan-update period. The recordings also showed no obvious idle intervals dominated by waiting for the next inference. It then used action speeds and pauses in 154 demonstrations as a reference and turned to the actions generated by the policy.

This diagnosis focused the investigation on saving time during action execution. Opus 5.5 pursued two lines of inquiry: first, whether anomalies at old-to-new chunk transitions originated in generation; second, whether subsequent actions could advance farther while preserving the task path. The first led to prefix repair, and the second to tempo acceleration through a Q/K projection edit; we refer to their combination as the combined edit.

2.2 Repairing Chunk Jumps in the Retained Action Prefix#

By comparing model outputs with executed commands, Opus 5.5 found that the first 6 steps were supplied as conditioning actions, but the generated outputs deviated from those old actions. Deployment usually discarded only the first 4 expired steps, so deviations in steps 5 and 6 entered execution. The cause was that the prefix was not restored again after the final flow-matching integration step.

Opus 5.5 computed a weight-update matrix $\mathrm{\Delta }W$ and saved $W^{\prime }=W+\mathrm{\Delta }W$ in a new checkpoint. Opus edited $W$, the linear-layer weight mapping time features to scale and shift in the final normalization layer (Figure 1; final_norm_1.Dense_0.kernel). The $1024\times 3072$ matrix $\mathrm{\Delta }W$ specifies an increment for every weight. The goal is nearly zero prefix updates while preserving outputs at other sampling times, with inference code unchanged.

The original time-modulation process is shown in Figure 1, using $W$ at the highlighted location. The prefix uses generation-time condition $t=0$; other actions use $1.0,0.9,\dots ,0.1$. This is not physical execution time. Time encoding and time_mlp_in/out map $t$ to 1024-dimensional features $c_t$; $c_0$ is not zero. For one action token, let $x_t$ be its hidden features after 18 Transformer blocks, and $m_t$ the three groups of modulation values from the time branch. Features in this subsection are row vectors; $u,d$ are column vectors. The original model adjusts $x_t$ using scale and shift: $$\begin{array}{rl}{m}_t& =[{\mathrm{scale}}_t,{\mathrm{shift}}_t,{\mathrm{gate}}_t]=c_{t}W+b,\\ x_{\mathrm{out},t}& =\mathrm{RMSNorm}(x_t)\odot (1+{\mathrm{scale}}_t)+{\mathrm{shift}}_t,\\ v_t& =x_{\mathrm{out},t}{W}_{\mathrm{out}}+b_{\mathrm{out}}.\end{array}$$(1) These are existing model operations, which Opus used to derive the edit. Here, $\odot$ means elementwise multiplication; gate is unused. $\mathrm{RMSNorm}(x)=x/\sqrt{\mathrm{mean}(x^2)+{10}^{-6}}$ contains no additional learned scale; time modulation is written separately. The linear layer action_out_proj, with $W_{\mathrm{out}},b_{\mathrm{out}}$, has $W_{\mathrm{out}}\in {\mathbb{R}}^{1024\times 32}$ and produces the sampling update $v_t\in {\mathbb{R}}^{32}$; it is not the time MLP. The prefix repair replaces only $W$, preserving $b,W_{\mathrm{out}},b_{\mathrm{out}}$.

Step 1: choose targets that give zero prefix updates. Primes denote edited quantities. Opus targeted scale $-\mathbf{1}$ and solved for the minimum-norm $s^{\ast }$ that gives zero after the linear action output layer: $$\begin{array}{rl}{\mathrm{scale}}_0^{\prime }& =-\mathbf{1},{\mathrm{shift}}_0^{\prime }=s^{\ast },s^{\ast }{W}_{\mathrm{out}}+b_{\mathrm{out}}=0,\\ x_{\mathrm{out},0}^{\prime }& =\mathrm{RMSNorm}(x_0)\odot \mathbf{0}+s^{\ast }=s^{\ast },v_0^{\prime }=0.\end{array}$$(2) The Transformer output $x_0$ need not be zero: the target scale removes its influence, while the target shift makes the final features equal to $s^{\ast }$ and cancels the output bias.

Step 2: compute the target modulation change. Subtracting original values from targets gives $d^{\mathsf{T}}$; adding it back verifies the desired result: $$\begin{array}{rl}{d}^{\mathsf{T}}& =[-\mathbf{1}-{\mathrm{scale}}_0,s^{\ast }-{\mathrm{shift}}_0,\mathbf{0}],\\ m_0+d^{\mathsf{T}}& =[{\mathrm{scale}}_0,{\mathrm{shift}}_0,{\mathrm{gate}}_0]+d^{\mathsf{T}}=[-\mathbf{1},s^{\ast },{\mathrm{gate}}_0].\end{array}$$(3) Step 3: solve for $u$ to apply the increment only to the prefix. Opus stacked the time features at the ten sampling times and at the prefix condition $t=0$ as rows of $A\in {\mathbb{R}}^{11\times 1024}$. It sought 1024 coefficients $u$ giving $c_{0}u=1$ for the prefix and $c_{t}u=0$ at the ten other times. Stacking these targets into $r$ yields $$\begin{array}{rl}A& ={\mathrm{stack}}_{\mathrm{rows}}(c_0,c_{1.0},c_{0.9},\dots ,c_{0.1}),r=(1,0,\dots ,0{)}^{\mathsf{T}},\\ u& =\underset{z}{\mathrm{argmin}}\parallel Az-r{\parallel }_2^2\text{(minimum-norm solution)}.\end{array}$$(4) A dot product of 1 adds the full change $d^{\mathsf{T}}$ to the prefix; a dot product of 0 adds nothing. Thus $c_0$ comes from the original model, whereas $u$ comes from this solve.

Write the weights and check the effect. The desired change after applying the new weights is $(c_{t}u)d^{\mathsf{T}}$. By associativity, this equals $c_t(u{d}^{\mathsf{T}})$, so $u{d}^{\mathsf{T}}$ can be added directly to the weight matrix: $$\begin{array}{rl}& \mathrm{\Delta }W=u{d}^{\mathsf{T}},W^{\prime }=W+\mathrm{\Delta }W,W_{ij}^{\prime }=W_{ij}+u_i{d}_j,\\ & c_t\mathrm{\Delta }W=c_t(u{d}^{\mathsf{T}})=(c_{t}u)d^{\mathsf{T}},\\ & m_t^{\prime }=c_t(W+\mathrm{\Delta }W)+b=m_t+(c_{t}u)d^{\mathsf{T}}.\end{array}$$(5) Row $i$ of $\mathrm{\Delta }W$ is $u_i{d}^{\mathsf{T}}$: all rows are scaled copies of the same vector, leaving only one independent direction. This is a rank-one update.

Across 36 demonstration observations and 24 historical execution observations, the prefix joint error fell from approximately 0.095 rad to $4.49\times {10}^{-4}$ rad. The maximum change in subsequent actions was approximately 0.0034 rad, with almost no change in path progress. Opus 5.5 therefore retained this prefix repair and continued searching for a speedup method. Our code audit and experiments on this issue appear in Section 3.2.

2.3 The Action-Tempo Hypothesis and RoPE Position Intervention#

The prefix repair did not make subsequent actions advance faster. While reading the model code, Opus 5.5 proposed a hypothesis for acceleration: “My leading hypothesis is that action timing depends on RoPE positions.” We elaborate on this hypothesis as follows: the model may have learned an implicit relationship between action-position spacing and action displacement; increasing the internal position spacing while keeping the physical control interval unchanged may increase useful displacement over the same duration.

Figure 2 summarizes the intervention: change the coordinates used to read the action sequence, then test whether predicted actions advance faster.

Implementation: change the Q/K rotations associated with action positions.

Let $i=0,\dots ,49$ index actions and $P$ denote the number of unmasked context tokens. Action indices here and in the subsequent offline formulas are zero-based. The internal positions change, while the physical execution interval remains $\mathrm{\Delta }t=1/30$ s: $$p_i=P+i,p_i^{(s)}=P+si,t_i^{\mathrm{e}\mathrm{x}\mathrm{e}\mathrm{c}}=t_0+i\mathrm{\Delta }t.$$(6) RoPE encodes position through rotations of Q/K [3]: $$q_i=W_Q^{\mathsf{T}}{x}_i,k_i=W_K^{\mathsf{T}}{x}_i,Q_i=R(p_i)q_i,K_i=R(p_i)k_i.$$(7) Here, $x_i\in {\mathbb{R}}^{1024}$ is the column-vector feature after the layer’s adaptive normalization and before Q/K projection. $W_Q,W_K\in {\mathbb{R}}^{1024\times 256}$ denote the weights for one query head and the shared key head; the implementation has eight query heads and one key/value head. Within each 256-dimensional head, channels $m$ and $m+128$ form a rotation pair, for $m=0,\dots ,127$. For one such pair $(a,b)$, $$\left[\begin{array}{c}{a}^{\prime }\\ b^{\prime }\end{array}\right]=\left[\begin{array}{cc}\cos \theta & -\sin \theta \\ \sin \theta & \cos \theta \end{array}\right]\left[\begin{array}{c}a\\ b\end{array}\right],\theta =p{10000}^{-m/128}.$$(8) $R(p)$ denotes the rotations of all 128 channel pairs. Holding the content vectors before rotation fixed, the inner product between actions $i,j$ changes from the original expression to $$Q_i^{\mathsf{T}}{K}_j=q_i^{\mathsf{T}}R(j-i)k_j,(Q_i^{(s)}{)}^{\mathsf{T}}{K}_j^{(s)}=q_i^{\mathsf{T}}R(s(j-i))k_j.$$(9) The intervention therefore increases relative position spacing between actions. It also changes how action queries read the unchanged image, text, and state keys. Through softmax and the weighted aggregation of V, these changes affect subsequent action generation. The intervention does not directly scale V or joint targets.

Test: does the position intervention increase progress along the original route?

Opus 5.5 ran two sets of controls: (1) on 36 demonstration observations and 24 historical rollout observations, it changed action positions in all 18 layers to $P+si$, testing $s=0.8,1.25,1.5,2.0$; (2) on 36 demonstration observations, it kept RoPE unchanged and accelerated only the first six conditioning actions by factors of 1.25, 1.5, and 2.0. Each intervention and the original model used the same observations and initial noise.

It first resampled the original predictions at different speeds and found the speed that most closely matched the intervened joint trajectory. The RoPE intervention with $s=1.25$ corresponded to approximately $1.12\times$, whereas accelerating only the prefix corresponded to approximately $1.01\times$. It therefore continued to investigate the position intervention.

Opus 5.5 then tested: at the same action step, does the edited model move farther along the route planned by the original model? At a fixed action execution rate, the same step corresponds to the same elapsed execution time. It performed forward kinematics and connected the positions of the original TCP poses into a reference route. Each edited position was matched to the nearest point on that route. Dividing the cumulative distance along the route to that point by the original model’s distance at the same action step gives the path-progress ratio; the mean nearest distance from edited positions to the route gives the path deviation (Figure 3).

The calculation excludes the first six conditioning steps. Let $c_{\mathrm{r}\mathrm{e}\mathrm{f}}(u)$ be the linearly interpolated reference TCP route from original steps 6–49 (in meters), $L_{\mathrm{r}\mathrm{e}\mathrm{f}}(u)$ its cumulative arc length from step 6, and $e_r$ the edited position at step $r$. Progress is evaluated only at $r=21,37$; deviation uses all positions in $\mathcal{I}=\{6,\dots ,37\}$. Matching the code, progress searches the grid ${\mathcal{U}}_p=\{6+0.01k:k=0,\dots ,4300\}$; deviation searches ${\mathcal{U}}_d$, comprising 880 equally spaced points over $[6,49]$: $$\begin{array}{rl}{u}_r^{\ast }& =\arg \underset{u\in {\mathcal{U}}_p}{min}\parallel e_r-c_{\mathrm{r}\mathrm{e}\mathrm{f}}(u){\parallel }_2,r\in \{21,37\},\\ g_r& =\frac{L_{\mathrm{r}\mathrm{e}\mathrm{f}}(u_r^{\ast })}{max\{L_{\mathrm{r}\mathrm{e}\mathrm{f}}(r),{10}^{-9}\}},\\ \overline{d}& =\frac{1000}{|\mathcal{I}|}\sum _{r\in \mathcal{I}}\underset{u\in {\mathcal{U}}_d}{min}\parallel e_r-c_{\mathrm{r}\mathrm{e}\mathrm{f}}(u){\parallel }_2.\end{array}$$(10) The denominator floor is ${10}^{-9}$ meters, avoiding division by zero when reference progress vanishes; the factor of 1000 converts $\overline{d}$ to millimeters. Table 1 reports $g_{37}$ and $\overline{d}$ (mm). Arms whose original motion covers less than 15 mm within the comparison interval are excluded. Eligible arms are averaged within each observation, followed by the median across observations.

Table 1 shows that $s=1.25$ and $1.5$ increase progress, whereas the benefit largely disappears at $2.0$ and deviation increases. Accelerating only the prefix has little effect. We hypothesize that the relationship between action positions and displacements learned at a fixed action execution interval may limit the benefits of increasing position spacing. This is an interpretation of the non-monotonic trend, rather than an established cause.

Opus 5.5 selected $s=1.25$ as its main target. This intervention changes the position rule within the action expert’s attention, leaving its weights unchanged. To preserve the production inference code, it used this model as a direct RoPE teacher and attempted to internalize the intervention into new weights with the same architecture as the original model.

2.4 First Attempt: Fit Only the Final Action Output Layer#

The minimal-change test: can editing only the last layer reproduce the teacher’s final actions? The original output weight is $1024\times 32$: the first 14 channels represent the two arms and grippers, and the remaining 18 are padding. Opus 5.5 fitted only the first 14 columns, preserving the other columns and all biases. Below, $B\in {\mathbb{R}}^{1024\times 14}$ denotes those columns and $b_0\in {\mathbb{R}}^{14}$ the corresponding bias; $A_i^T,z_{k,i},v_{k,i}\in {\mathbb{R}}^{14}$ contain those 14 channels in normalized action space. $A_i^T$ is the teacher’s final output, $z_{k,i}$ the candidate state at iteration $k$ and position $i$, and $h_{k,i}\in {\mathbb{R}}^{1024}$ the column-vector feature after final time-conditioned normalization, entering the output layer. The remaining generation time is ${\sigma }_k=1-k/10$: $$\begin{array}{rl}{v}_{k,i}^{\ast }& =\frac{z_{k,i}-A_i^T}{{\sigma }_k},v_{k,i}=B^{\mathsf{T}}{h}_{k,i}+b_0,\\ z_{k+1,i}& =z_{k,i}-0.1v_{k,i},k=0,\dots ,9,i\ge 6.\end{array}$$(11) $v^{\ast }$ is the target update toward the teacher’s final action, not the robot arm’s velocity. Opus 5.5 fitted $B$ using least squares regularized toward the original weights, with bias $b_0$ fixed. The sampler iterates ten times to produce the final actions (Figure 4).

Reconstruction on fitting inputs and transfer to new inputs were checked separately. Let $a_{i,j},a_{i,j}^T$ be the candidate’s and teacher’s final joint targets, with $j$ ranging over the 12 joints. The joint error for each observation is $$E_{\mathrm{R}\mathrm{M}\mathrm{S}}=\sqrt{\frac{1}{32\times 12}\sum _{i=6}^{37}\sum _{j=1}^{12}(a_{i,j}-a_{i,j}^T{)}^2}.$$(12) This squares the errors, averages them, and takes the square root, giving an error in radians; results are then averaged over observations. To check for abrupt changes in motion, it computed jerk, the rate of change of acceleration, from TCP positions ${\mathbf{p}}_i$: $${\mathbf{j}}_i=\frac{{\mathbf{p}}_{i+3}-3{\mathbf{p}}_{i+2}+3{\mathbf{p}}_{i+1}-{\mathbf{p}}_i}{\mathrm{\Delta }{t}^3},J=\sqrt{\frac{1}{29}\sum _{i=6}^{34}\parallel {\mathbf{j}}_i{\parallel }_2^2}.$$(13) The reported ratio is $J_{\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{t}}/J_{\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{g}}$, averaged over eligible arms within each observation. It excludes chunk transitions and is interpreted together with path progress and deviation.

After three fitting rounds on 24 demonstration observations, joint RMS was 0.0050 rad. On another 12 demonstration and 16 historical rollout observations, the error increased to 0.0239 rad, mean path deviation was 8.256 mm, and the jerk ratio was 3.245. Reconstruction quality did not carry over to the test inputs, so Opus 5.5 abandoned output-head fitting and turned to the attention layers.

2.5 Moving Inside Attention: Fitting Q/K Projections#

Opus 5.5 moved the fitting target from final actions to Q/K representations: a known rotation directly supplies the targets, which the new weights approximate (Figure 5).

Step 1: construct intermediate representation targets. To retain the candidate’s native RoPE $R(P+i)$ while approximating the teacher’s $R(P+si)$, an additional rotation must be supplied by the projection output: $$R(P+i)R({\delta }_i)=R(P+si),{\delta }_i=(s-1)i.$$(14) Targets are therefore constructed separately for Q and K: $$y_i^Q=R({\delta }_i)W_{Q,0}^{\mathsf{T}}{x}_i,y_i^K=R({\delta }_i)W_{K,0}^{\mathsf{T}}{x}_i.$$(15) Each target $y_i$ comes from applying a known rotation to the original Q/K of the same hidden feature $x_i$, without human labels for faster actions.

Step 2: collect features and fit a closed-form solution. Opus 5.5 first checked whether the feature-collection forward pass altered the original output. Joint RMS was approximately 0.00037–0.00044 rad, below its chosen tolerance of 0.002 rad. It transposed the column vectors, using $x_i^{\mathsf{T}}$ and $y_i^{\mathsf{T}}$ as individual rows of $X\in {\mathbb{R}}^{N\times 1024}$ and $Y\in {\mathbb{R}}^{N\times 256}$, then solved $$\begin{array}{rl}{W}_{\mathrm{f}\mathrm{i}\mathrm{t}}& =\arg \underset{W}{min}(\parallel XW-Y{\parallel }_F^2+{\lambda }_{\mathrm{a}\mathrm{b}\mathrm{s}}\parallel W-W_0{\parallel }_F^2)\\ & =(X^{\mathsf{T}}X+{\lambda }_{\mathrm{a}\mathrm{b}\mathrm{s}}I{)}^{-1}(X^{\mathsf{T}}Y+{\lambda }_{\mathrm{a}\mathrm{b}\mathrm{s}}{W}_0),\\ \mathrm{\Delta }W& =W_{\mathrm{f}\mathrm{i}\mathrm{t}}-W_0,{\lambda }_{\mathrm{a}\mathrm{b}\mathrm{s}}={\lambda }_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}\frac{\mathrm{tr}(X^{\mathsf{T}}X)}{1024}.\end{array}$$(16) The first term brings the new projection closer to the target; the second limits its departure from the original weights. Here, $\parallel \cdot {\parallel }_F^2$ is the sum of squared entries, $I$ is the identity matrix, $N$ is the number of collected feature rows, and ${\lambda }_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}$ is the candidate regularization coefficient, scaled by feature energy. The implementation solves a linear system rather than explicitly computing an inverse and uses no backpropagation. Each round recollects features from the candidate and accumulates $X^{\mathsf{T}}X$ and $X^{\mathsf{T}}Y$ across rounds before solving; $X,Y$ denote the accumulated samples. V projection weights remain unchanged. The shared weights do not directly receive the action index, so they rely on position information already present in the features to approximate the additional rotation.

Step 3: inspect generated actions and refine the fit. The initial fit edited Q/K in all 18 layers. After two rounds, normalized action RMS against the teacher on fitting observations fell from 0.03694 to approximately 0.00298. This measured final actions, not Q/K feature errors. The refined fit preserves layer 0 and fits only layers 1–17; its corresponding teacher also changes RoPE only in those layers. The rationale was that layer 0’s inputs have not yet undergone action-attention mixing and may lack the information needed for a shared projection to distinguish action indices.

Step 4: validate the refinement on held-out observations. The tempo edit used 48 observations from 12 demonstrations and 12 historical rollouts for fitting, reserving another 36 time points for testing. The time points do not overlap, but come from the same trajectories. After three fitting rounds, on the shared test set, the progress ratio increased from approximately 1.17 for the initial version to 1.20, deviation fell from 1.26 to 1.08 mm, and the jerk ratio fell from 1.47 to 1.37. Opus 5.5 therefore retained the tempo edit and proceeded to test edit strength and its combination with prefix repair.

2.6 Adding Parameter Constraints and Testing the Combination with Prefix Repair#

Opus 5.5 controlled the fitted change with an editing strength $\alpha$: $$W(\alpha )=W_0+\alpha \mathrm{\Delta }W,\mathrm{\Delta }W=W_{\mathrm{f}\mathrm{i}\mathrm{t}}(s=1.25)-W_0.$$(17) $\alpha =0$ gives the original Q/K projections, and $\alpha =1$ gives the full tempo edit. This is a different parameter from the positional scale $s=1.25$ used during fitting. When it tested $\alpha$ from 0.25 to 1.5, progress initially increased, then saturated as path deviation grew. Restricting the edit to rank 16, 64, or 256 gave progress ratios of approximately 1.06, 1.10, and 1.16, below the approximately 1.20 achieved by the full tempo edit.

It then compared prefix repair, tempo edit, and their combination on the same 36 test observations, adding trajectory time scaling and action-amplitude scaling as baselines, both with a factor of 1.21445. The path and jerk metrics in Table 2 include 35 valid observations. The chunk-jump metric is the maximum adjacent TCP displacement around the splice between the first 4 steps of the old plan and steps 5–13 of the new plan, taking the larger value across the two arms. It includes ordinary displacement between adjacent actions.

The combined candidate increased path progress and reduced chunk jumps, but its within-chunk jerk ratio rose to 1.363. Time scaling had lower deviation and jerk, but retained the original chunk jumps. A random weight direction produced 24.05 mm path deviation and a jerk ratio of 31.07, failing to preserve the original behavior. Opus 5.5 ultimately selected the combined edit with $\alpha =1$. Since it did not compare against prefix repair combined with time scaling, this does not establish that editing outperforms every post-processing alternative.

2.7 Deliverables and Research Cost#

After checking saving, rollback, and native reloading, Opus 5.5 delivered an independent checkpoint that modified 1.16% of all model parameters. All 8 test inputs produced valid outputs. Its initial research run took 1 hour, 50 minutes, and 42 seconds, with 193 tool calls (163 Bash, 18 Write, 8 Read, and 4 others) and 187 API requests. Usage records were available for 181 requests, totaling 184,614 output tokens, with 433,426 cache-write tokens and 36,783,281 cache-read tokens. The 6 requests without usage records are not counted as zero; cache reads include repeated context. These figures exclude subsequent real-robot analysis and preparation of this article.

3 Our Follow-up Experiments and Analysis#

We next measured actual task time and chunk jumps, and analyzed internal effects through attention-region interventions and prefix-handling controls. These are our additional experiments and are not included in the cost of Opus 5.5’s initial research run.

3.1 Real-Robot Experiments: Task Time and Action Quality#

We compared the original Training-Time RTC policy with the combined edited policy ($\alpha =1$) on bottle placement, coin insertion, and block stacking. Within each task, the two policies were evaluated on the same object layout, with three executions per condition. Only bottle placement was tested at multiple action execution rates. (1) Bottle placement: weight-editing gains persist across execution rates. Beyond the earlier 30 Hz tests, we compared both policies at 35 and 40 Hz. At each rate, both policies used matched settings and the same object layout, with three successful executions per condition. Time runs from formal execution start to the fourth bottle’s release inside the bin, measured using acquisition timestamps.

We kept both checkpoints fixed and shortened only the execution interval between action points, without an additional RoPE intervention or weight fit. The 32-step execution setting, six-step conditioned prefix, and 100 Hz low-level control loop stayed fixed; the same number of steps therefore spans less physical time at higher rates.

Increasing the execution rate accelerated both policies, while the edited policy retained an additional benefit at all three matched rates. In Table 3, the edited policy at 35 Hz approximately matched the original at 40 Hz. When both ran at 40 Hz, the edited policy still finished 2.15 s earlier on average. This supports a change in how the policy generates actions whose benefit can combine with faster execution. Consistency here means a benefit at every tested rate, not an identical effect size.

Chunk jumps decreased at all three rates; the overall mean fell from 30.88 to 3.14 mm, a reduction of 89.8%. We used the same calculation throughout: between formal start and the fourth bottle release, locate adjacent command ticks across each plan switch, compute their TCP distance with forward kinematics, and take the larger value across the two arms. We average within each trial, then equally across the three trials per condition. Initial plan entry and transitions through an empty plan are excluded. This metric measures the command step at a switch. Table 4 reports each rate; the aggregate weights all three rates equally rather than averaging percentage reductions.

Faster execution also comes with larger residual chunk jumps. For the edited policy, mean jump rises from 1.43 to 3.09 and 4.89 mm at 30, 35, and 40 Hz, while the reduction relative to the original narrows from 94.5% to 90.7% and 85.2%. The edited policy still has substantially smaller command jumps at each matched rate, but the results reveal a trade-off between execution speed and command continuity.

Per-trial completion times from the earlier 30 Hz cohort appear in Table 5; Table 6 further reports action quality for those six 30 Hz recordings.

The edited policy finished 0.96 s earlier on average, a 4.3% reduction in time. This compares three successful executions per policy under the same layout and does not estimate a change in success rate.

Within the same completion interval of those six 30 Hz recordings, we computed the TCP distance between adjacent commands at every plan switch, taking the larger value across the two arms. We then computed statistics per trajectory and averaged them over the three trajectories in each group (Table 6). The original and edited groups contained 60 and 58 switches, respectively; the initial entry into a plan was excluded.

Mean chunk jump decreased by 94.5%, while end-effector tracking error remained almost unchanged. The command jerk P95, computed from third-order differences of the complete command trajectory at 30 Hz, also decreased by 77.4%. It includes improvements at plan switches, unlike the offline jerk metric that measures only within-chunk behavior.

(2) Coin insertion: the speedup occurred mainly during transport after grasping. We divided the action into stages using sustained closing and opening events in the measured gripper states. Table 7 shows that the total time from left-gripper closure to handover release fell from 5.85 s to 4.64 s, a 20.7% reduction. Within this interval, the transport stage from left-gripper closure to right-gripper closure shortened by 22.1%; the time from the start to grasping and the time spent releasing during the handover were almost unchanged.

The 20.7% measures the reduction in the stage from grasping to handover release, which we report separately from the full-task completion time in the bottle experiment.

(3) Block stacking: no clear speedup was observed. All 3 tests in each group ultimately succeeded, giving 3/3 successes for both policies. In the original policy’s second execution, a failed stacking attempt was followed by a new grasp and successful completion, resulting in a longer total time of 28.27 s. Table 8 retains the completion times of all six executions. To compare execution speed without retries, we exclude that run when computing the mean. The remaining 2 original runs and the 3 edited runs averaged 19.46 and 19.28 s, respectively: a reduction of only 0.19 s (approximately 1.0%), with overlapping individual durations.

These real-robot results show that Weight Editing did produce gains in our tasks: bottle placement became faster with smaller chunk jumps, and the transport stage in coin insertion also accelerated. However, this remains an initial attempt, with limited speed gains. This is consistent with the hypothesis and offline results in Section 2.3: the learned relationship between action positions and displacements may provide a way to adjust the pace of actions.

3.2 Repairing Prefix Consistency in Parameter Space#

Unexpectedly, prefix repair brought our attention to a bug left unhandled in the official public Training-Time RTC (TT-RTC) code: when actual latency is shorter than the preset prefix, conditioning actions modified by the final integration step may still enter execution. We first briefly introduce TT-RTC.

TT-RTC addresses how a robot can keep moving while waiting for inference of the next chunk and then switch smoothly to that chunk. The top row of Figure 6 shows the remaining actions $A_1,A_2,\dots$ from the previous chunk. When the model starts inferring the next action sequence at $A_1$, the robot continues executing the current chunk. We use the first six steps $A_1$–$A_6$, expected to execute during the wait, as the conditioning prefix, allowing the model to generate subsequent new actions $B_7,B_8,\dots$. The fixed-latency simulation in the official TT-RTC code uses equal conditioning length and latency. Taking the six-step example in Figure 6(a), the new result arrives just as the old chunk reaches $A_6$. The program discards the first six steps of the new output and starts executing at $B_7$, completing the chunk transition [4], [5].

Figure 6(b) shows the problem when the result returns early. Model computation takes only four action steps, and the robot has reached only $A_4$ of the old chunk, so it still needs $A_5,A_6$. However, the conditioning prefix still participates in integration updates during flow-matching sampling. With the ten sampling iterations we use, prefix changes made by integrations 1–9 are replaced by the supplied old actions before the next iteration. After the tenth integration, however, the result is returned directly, without another replacement. Consequently, the new result may contain $A_5^{\ast },A_6^{\ast }$ rather than the planned $A_5,A_6$. Switching chunks at this point sends the robot two modified actions, producing a jump. Under the setup in Figure 6(a), the official TT-RTC code discards the entire prefix, so this issue is not exposed.

For deployments requiring IK, we typically reserve a conservatively longer prefix to cover solver-latency fluctuations. Larger fluctuations can also create a larger gap between this reserved length and the latency of a particular inference call. When a result returns early, more conditioning actions remain to be executed. The original OpenWAM [6] implementation uses Mink differential IK wrapped by i2rt, solving quadratic programs (QPs) iteratively. Action decoding, including IK, has a median duration of 22.9 ms and a maximum of 921.2 ms. The subsequently used DLS gate adjusts motion regularization according to the Jacobian’s smallest singular value and limits joint changes; its duration still ranges from a median of 49.3 ms to a maximum of 177.7 ms.

Figure 6(c) shows the repaired transition: even when the new result arrives after $A_4$, the next executed actions are still the planned $A_5,A_6$. Opus 5.5’s prefix repair makes the sampling updates for these conditioning actions close to zero, greatly reducing the changes introduced by the final integration step. It edits the time-modulation weights of the action expert’s final normalization layer, final_norm_1.Dense_0.kernel. Under the current setup, its effect is nearly equivalent to executing actions[:6] = condition[:6] before returning: restoring the first six steps to the supplied old actions. Opus approximated this operation by editing parameters.

We subsequently added controls on the same 36 observations (12 local demonstration observations and 24 historical rollout observations), fixing random seeds and sampling steps. We compared shortening the conditioning prefix, using prefix repair alone, and directly restoring the prefix, all without the tempo edit.

Table 9 shows that, compared with the original 19.16 mm, the prefix weight edit and the code fix that directly restores the prefix reduce the median transition-step distance to 1.80 and 1.88 mm, respectively, with very small changes to subsequent actions. This supports their approximate equivalence for prefix repair. Shortening the conditioning prefix to four steps also improves this particular comparison, but does not eliminate the same issue when actual latency is shorter than the conditioning length.

By editing time-modulation weights to make updates to the conditioning prefix close to zero, Opus repaired in parameter space a bug left unhandled by the official public code: when actual latency is shorter than the preset prefix, conditioning actions altered by sampling updates may still enter the execution sequence. This prefix protection is especially important for deployments requiring IK, where a margin must be reserved for latency fluctuations. We did not specify this defect; Opus discovered it during its research and produced a parameter-level repair under the constraint of leaving inference code unchanged.

3.3 Mechanism Analysis: Which Attention Changes Preserve the Progress Gain?#

We compared the original model, the direct RoPE teacher, and the tempo-edited model on the same eight bottle-task observations, fixing the initial noise, RTC conditions, and number of sampling steps, with prefix repair disabled. First, the tempo-edited model’s attention to context was closer to the teacher’s: its distributional distance, measured by JSD, was 0.00173, below its distance of 0.00792 from the original model (lower means more similar). This supports that fitting preserved part of the position intervention’s internal effect.

Figure 7 marks two regions: the lower-right block contains actions reading actions, and the lower-left block contains actions reading images, text, and state. At each layer, using the same current hidden features $X$, we compute projections with the original ($r=0$) and edited ($r=E$) Q/K weights according to Equation (7) and apply the original RoPE, obtaining the pre-softmax scores: $$L_{ij}^r=\frac{(Q_i^r{)}^{\mathsf{T}}{K}_j^r}{\sqrt{d}},L^r=[L_{\mathrm{c}\mathrm{t}\mathrm{x}}^r\mid L_{\mathrm{a}\mathrm{c}\mathrm{t}}^r],r\in \{0,E\}.$$(18) Here, $i$ is an action query, $j$ is a key, and $d=256$; ctx denotes images, text, and state, while act denotes actions. Cached context keys remain unchanged. The two regional interventions are $$\begin{array}{rl}{A}_{\mathrm{a}\mathrm{c}\mathrm{t}\text{-}\mathrm{o}\mathrm{n}\mathrm{l}\mathrm{y}}& =\mathrm{softmax}([L_{\mathrm{c}\mathrm{t}\mathrm{x}}^0\mid L_{\mathrm{a}\mathrm{c}\mathrm{t}}^E]+M),\\ A_{\mathrm{c}\mathrm{t}\mathrm{x}\text{-}\mathrm{o}\mathrm{n}\mathrm{l}\mathrm{y}}& =\mathrm{softmax}([L_{\mathrm{c}\mathrm{t}\mathrm{x}}^E\mid L_{\mathrm{a}\mathrm{c}\mathrm{t}}^0]+M).\end{array}$$(19) $M$ is the original attention mask ($0$ at readable positions and $-\mathrm{\infty }$ at masked positions). The combined scores are normalized together, followed by $AV$ and the remaining forward pass; V projection weights remain unchanged. Because the softmax denominator is shared, changing one block’s scores can also change the other block’s probabilities.

Table 10 shows that the step-37 progress ratio changes from 1.140 with the full tempo edit to 1.087 with the action-only intervention and 1.039 with the context-only intervention. Intervening on scores between actions preserves more of the offline progress gain.

Closing Thoughts#

We expected a moment of magic: Claude would intuitively change a few weights, and the robot would move faster. That is not what happened. We believe the key to this Weight Editing result was the Frontier Model’s AutoResearch ability: understanding mechanisms, finding problems, testing hypotheses, and revising its approach after failure. Its understanding of parameter space made the edits possible; the research process helped it find edits worth making.

This raises a more important question: as Frontier Models increasingly internalize robotics knowledge, how will the next generation, and the generations beyond it, change the way we ask questions, test ideas, and advance research? This case cannot yet provide an answer, but it makes these questions concrete and leaves us excited about what comes next. As models take on more exploration and validation, we need to think more carefully: how can we distill robotics needs into research tasks suited to AutoResearch? And how can we identify more meaningful scientific or engineering problems?

We hope this article offers a glimpse into research on Frontier Models for Robotics.

References

1. Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab. Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights. MIT CSAIL, 2026. https://rikiyat.github.io/blog/weight-space-programming/.

2. Physical Intelligence et al. ${\pi }_{0.5}$: a Vision-Language-Action Model with Open-World Generalization. arXiv:2504.16054, 2025. https://arxiv.org/abs/2504.16054.

3. Jianlin Su et al. RoFormer: Enhanced Transformer with Rotary Position Embedding. arXiv:2104.09864. https://arxiv.org/abs/2104.09864.

4. Kevin Black, Allen Z. Ren, Michael Equi, Sergey Levine. Training-Time Action Conditioning for Efficient Real-Time Chunking. arXiv:2512.05964v2. https://arxiv.org/html/2512.05964v2.

5. Physical Intelligence. real-time-chunking-kinetix, commit 9296f31d62d5. https://github.com/Physical-Intelligence/real-time-chunking-kinetix. We inspected src/model.py and src/eval_flow.py.

6. Yuran Wang et al. OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining. arXiv:2609.07398, 2026. https://arxiv.org/abs/2609.07398.

@misc{gong2026roboticsweightediting,
  author = {Gong, Ziyang and Zheng, Kaifeng and Huang, Yingrui and
            Chen, Wendi and Tian, Jiaguo and Yang, Junkai and
            Liu, Minghuan and Yang, Zhenjie},
  title = {{We Asked Opus-5.5 to Edit $\pi_{0.5}$'s Weights.
            AutoResearch Made It Happen.}},
  year = {2026},
  url = {https://cuzyoung.github.io/robotics-weight-editing/en/}
}