CKT-WAM: Parameter-Efficient Context Knowledge Transfer Between World Action Models
1. Paper overview
In one sentence: A learned context interface improves a frozen student's manipulation performance, but keeps the teacher in the inference loop and leaves important adapter implementation details unresolved. e-probleme-teachere-compressione-routinge-main-resultse-reale-budgete-layout-conflict
| At a glance | What to know |
|---|---|
| Research problem | Source description Heterogeneous WAMs can have incompatible latent representations and action heads, making output imitation or layerwise feature matching restrictive. The authors seek an inexpensive trainable interface that lets a smaller model consume useful teacher representations without redesigning its generative backbone. e-problem |
| Core mechanism | Source description A teacher observation encoder, learnable-query compression, an always-active generalized branch and sparse specialized branches construct transferable context. Only the CKT module is optimized. e-interfacee-compressione-routing |
| A key reported result | LIBERO-Plus zero-shot robustness: 86.1 Reported total success rate (%). Training on all four LIBERO suites; evaluation under camera, robot, language, light, background, noise and layout shifts. Cosmos-Policy 82.2; π0.5 85.7. Reader-computed gains: 3.9 and 0.4 percentage points. Highest listed aggregate, not dominance in every category. Language is 86.7 versus VLA-JEPA's 88.1, contradicting the surrounding claim of best language performance. e-protocole-main-results |
| Reading caution |
Core contributions
- Source description
A teacher observation encoder, learnable-query compression, an always-active generalized branch and sparse specialized branches construct transferable context. Only the CKT module is optimized. e-interfacee-compressione-routing
- Reader analysis
Appending context to conditioning keys and values preserves the student's visual rotary-coordinate scheme. This is an interface property, not proof of preserved predictive accuracy or unchanged hidden representations. e-injectione-positional
Figure 1. Teacher knowledge enters through a compact conditioning interface. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the frozen teacher on the left and follow its arrow to hidden-state selection. The two compressor query banks feed a generalized branch and specialized branches. The dotted router connections select which specialists contribute; Equation (6) specifies normalized weighting over the selected set. The generalized output stays active and joins the specialized context in C_A. Follow the long downward arrow into the student's context-injection block: E_i remains image conditioning, while E_t, the separator and C_A form the textual conditioning sequence. Fire and snowflake symbols mark trainable modules and frozen backbones. Equations (8)–(9) establish that injection changes cross-attention keys and values, not the visual self-attention token stream. e-interfacee-teachere-compressione-routinge-injectione-main-resultse-layout-conflict
What it supports. The interface lets a frozen student respond differently without changing its backbone weights: its attention receives new, observation-dependent conditioning values. The teacher supplies intermediate features once per observation, not a generated future trajectory. The main formulation therefore transfers a representation through conditioning rather than requiring the student to imitate teacher actions.
Where the evidence stops. The inset says 85.4%, while Tables 1–3 report 86.1%; use the tables for results. Main-text shared compression and Appendix B.4's per-adapter compression also disagree. The faithful diagram does not resolve that implementation conflict.
2. Motivation
2.1 The problem and the proposed response
Heterogeneous WAMs can have incompatible latent representations and action heads, making output imitation or layerwise feature matching restrictive. The authors seek an inexpensive trainable interface that lets a smaller model consume useful teacher representations without redesigning its generative backbone. e-problem
2.2 What this reading follows
CKT-WAM asks how one world action model can help another when their internal representations differ. Its answer is to turn the teacher's current-observation features into conditioning tokens that the student can already consume. Learnable queries compress those features; a general branch and selected specialists transform them; the student receives the result alongside its language embeddings. The experiments connect this interface to stronger out-of-distribution manipulation and physical task completion. Read the efficiency claim carefully: 1.17% refers to trainable parameters relative to both backbones combined. The teacher still runs, and the source's main equations and appendix disagree about the exact adapter implementation. e-probleme-teachere-compressione-routinge-main-resultse-reale-budgete-layout-conflict
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded outside-quadrant classification fits the contribution: a transfer framework between existing WAMs, rather than a newly defined monolithic prediction/control architecture. Teacher–student multiplicity does not establish an imagine-then-inverse-dynamics pipeline. Action/video training losses likewise do not alone establish a One Model classification. Efficiency and generalization tags are supported, with deployment cost qualified by the retained teacher. e-taxonomye-interfacee-objectivee-depth
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Turn a teacher observation into conditioning tokens
Begin with the current scene and instruction. Instead of asking DreamZero to synthesize a future, CKT-WAM queries an intermediate transformer layer at the clean timestep. In the main mathematical description, a bottleneck maps those states into the student's embedding width, and two learnable query banks each produce a compact sequence. One sequence receives a generalized residual transformation. The other receives a weighted mixture of the selected specialized adapters. The router makes that selection from pooled teacher states, so different observations can activate different specialists. Concatenating both branches forms C_A. With the stated 32 queries in each bank, Equation (7) produces 64 context tokens before the separator. This explains the intended information flow; Appendix B.4's alternative per-adapter compression accounting remains an unresolved implementation difference. e-teachere-compressione-routinge-confige-layout-conflict
5.2 Understand how a frozen student can still learn from context
A frozen backbone can change its output when its input conditioning changes. Here, future-video and action reconstruction losses backpropagate through the frozen student's computation into the trainable CKT interface. The source therefore trains useful conditioning without updating the student's weights. At inference, context is computed once for the observation and cached through the denoising calls. Appendix B.5 places it only in cross-attention keys and values: visual patches supply queries and can retrieve different combinations of the same teacher tokens. Reader interpretation: this separates the cost of constructing a scene summary from the repeated cost of consuming it. The positional proposition protects the visual coordinate grid and rotation matrices; it does not imply unchanged hidden features, identical predictions or preserved task accuracy. e-objectivee-injectione-positionale-teacher
5.3 Separate useful transfer from proven specialization
The ablation table is the strongest direct evidence that the added interface matters: removing context reduces total success from 86.1% to 82.2%. Removing either branch also hurts, while disabling balancing retains the architecture and lowers success to 84.1%. Appendix Figure 4 adds an observational clue: preferred adapters differ across stages in 120 cube-catching episodes. Figure 7 shows the balancing objective falling toward its balanced reference while averaged reconstruction losses remain stable. Reader interpretation: these findings are consistent with coordinated, input-dependent transfer, but they do not establish that particular experts causally implement particular skills. Branch removals alter capacity, and stage correlations can reflect changing visual appearance. A matched-context intervention and a controlled implementation audit would provide more discriminating evidence than another unqualified aggregate score. e-ablatione-diagnosticse-layout-conflict
5.4 Training and inference
During training
Both backbones remain frozen while gradients through the student train the context interface. Action and video targets share one sampled log-normal noise level but use independent Gaussian perturbations. Masked, noise-weighted reconstruction losses supervise clean targets; a balancing loss discourages concentrated expert allocation. e-objective
Training uses the combined training sets of all four LIBERO suites. The student starts from Cosmos-Policy-LIBERO-Predict2-2B. Defaults include eight A800 GPUs, bfloat16 distributed training, AdamW, learning rate 3×10⁻⁴, 1,000 warmup steps and cosine decay; Table 5 gives the remaining stated settings. e-confige-protocol
During inference
Compute and cache teacher context once per observation, then reuse it at each student denoising call, including classifier-free-guidance branches. The student remains the action generator. Real-world evaluation uses continuous visual feedback, but the paper does not specify the complete action-chunk execution schedule or control frequency. e-teachere-positionale-real-protocole-repro-gap
Freezing parameters saves optimization state; it does not remove teacher weights or computation from deployment. Sparse execution reduces active specialists but all eight specialists remain stored. e-routinge-budgete-depth
5.5 Implementation flow
- Query the teacher once
DreamZero-14B is frozen. Extract layer-20 states from its 40-layer transformer at the fixed clean timestep t*=0, using current visual tokens and text conditioning. The teacher does not generate future video or actions; its context is reused throughout the student's denoising trajectory. e-teachere-config
- Compress and route
The main formulation projects teacher states through a shared bottleneck into student embedding space. Generalized and specialized learnable queries compress the sequence separately. Residual adapters transform these contexts; a router pools teacher states, selects the top two of eight specialists, and renormalizes their weights. Concatenating generalized and aggregated specialized outputs produces the context. e-compressione-routinge-config
- Inject through conditioning
Append the context and a learnable separator after the student's text embeddings. Student visual tokens query the enlarged conditioning keys and values through frozen cross-attention. The context is not appended to visual self-attention; visual patch coordinates and RoPE rotation matrices retain their original indexing. e-injectione-positional
6. Experiments & results
CKT-WAM trains a context-transfer module between frozen DreamZero-14B and Cosmos-Policy-2B backbones. One teacher observation pass supplies compressed, routed tokens to the student's text-conditioning pathway. The paper reports 86.1% LIBERO-Plus success and 83.3% average physical-task success. Its 187.4M trainable parameters equal 1.17% of the combined backbones, while teacher inference remains necessary. Main-text and appendix adapter specifications disagree (e-teacher, e-config, e-main-results, e-real, e-budget, e-layout-conflict).
6.1 Read the original evidence
Table 1. Aggregate robustness improves, but category-level leadership varies. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the final Total column first, then compare individual perturbation columns. All entries are success rates in percent. CKT-WAM reaches 86.1, compared with 82.2 for its Cosmos-Policy student baseline and 85.7 for π0.5. The seven columns test different shifts after training on the combined LIBERO training suites; zero-shot here means evaluating those shifts without their adaptation data, not absence of robot-task training. Camera, Light, Noise and Layout favor CKT-WAM among these rows. Robot and Background favor X-VLA. Inspect the Language numbers directly: VLA-JEPA's 88.1 exceeds CKT-WAM's bolded 86.7. The source's formatting and surrounding prose do not correctly identify that category's leader. e-protocole-main-resultse-statistics
What it supports. The table supports a 3.9-percentage-point aggregate improvement over Cosmos-Policy and a smaller 0.4-point advantage over π0.5. It provides evidence of stronger overall robustness within this evaluation. It does not show universal superiority across the seven perturbation types, even though the CKT-WAM row contains several bold entries.
Where the evidence stops. The paper claims best language performance despite the contrary values. No uncertainty accompanies these estimates, and the table does not establish matched pretraining data or compute across every listed model.
Table 2. Near-full-fine-tuning success with a smaller optimization footprint. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the left-side parameter column with the rightmost success column. CKT-WAM reports 1.17% trainable parameters and 86.1% success; full fine-tuning reports 100% and 86.3%. LoRA reports 2.59% and 82.8%, while GOAT reaches 84.8% with 1.77%. Section 4.3 is essential context: alternative methods inject teacher features through an MLP and adapt the selected teacher prefix plus the student. CKT-WAM freezes those backbones and learns its own compressed interface. Appendix Table 8 supplies the denominator behind CKT-WAM's headline percentage: approximately 16 billion combined teacher–student parameters. Its 187.4 million trainable parameters are also 9.37% of the student alone. e-pefte-budgete-layout-conflicte-statistics
What it supports. The reported success difference from full fine-tuning is 0.2 percentage points. CKT-WAM also exceeds the listed PEFT methods while reporting the smallest trainable ratio. This supports the method's adaptation-efficiency claim in this setup, especially when optimizer state and trainable weights are the resource of interest.
Where the evidence stops. Trainable ratio is not a runtime-memory or latency ratio. These methods change both where learning occurs and how features are compressed; the comparison does not isolate routing at matched capacity or establish statistical equivalence to full fine-tuning.
Table 4. Physical task completion improves on average, with folding still a relative weakness. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each task column reports the percentage of successful physical trials; the final column averages the four tasks. Section 5 specifies 45 trials per task, and Appendix B.3 requires completing the ordered subgoals under continuous visual feedback. The CKT-WAM row reads 73.3 for clothes folding, 88.9 for fruit sorting, 86.7 for cube catching and 84.4 for unmanned retail. Compare it with π0.5: that baseline is stronger on folding, at 77.8, but weaker in the other three columns. The visually inspected task figures show sequential manipulation stages, so these numbers concern executed robot episodes rather than generated-video quality. e-reale-real-protocole-statisticse-repro-gap
What it supports. CKT-WAM's average is 83.3%, compared with π0.5's 80.0%, Cosmos Policy's 70.0% and Fast-WAM's 70.6%. The evidence supports successful multi-stage execution in the authors' four tasks. It does not isolate an explicit planning algorithm, and the aggregate improvement coexists with a lower folding success rate.
Where the evidence stops. Trial counts are given without confidence intervals. Physical training-data and execution details remain incomplete; cube-catching stage descriptions also differ between Figure 3 and Appendix B.3. These gaps constrain reproduction and interpretation of the small average advantage.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO-Plus zero-shot robustness Training on all four LIBERO suites; evaluation under camera, robot, language, light, background, noise and layout shifts. | 86.1 Reported total success rate (%) | Cosmos-Policy 82.2; π0.5 85.7. Reader-computed gains: 3.9 and 0.4 percentage points. Highest listed aggregate, not dominance in every category. Language is 86.7 versus VLA-JEPA's 88.1, contradicting the surrounding claim of best language performance. e-protocole-main-results |
| LIBERO-Plus parameter-efficient adaptation Comparators inject MLP-projected teacher features and adapt the first 20 teacher layers plus the student using their respective PEFT methods. | 86.1 / 1.17% Total success (%) / reported trainable parameter ratio | LoRA 82.8 / 2.59%; GOAT 84.8 / 1.77%; full fine-tuning 86.3 / 100%. The full-fine-tuning gap is 0.2 percentage points. Different trainable locations and compression designs mean this is a system comparison, not an isolated test of routing. e-pefte-budget |
| LIBERO-Plus component ablation Same zero-shot protocol; remove one transfer component or its auxiliary objective. | Full model 86.1 Total success rate (%) | No context 82.2; no generalized adapter 83.0; no specialists 82.7; no auxiliary loss 84.1. Removing specialists costs 3.4 percentage points; dropping balancing costs 2.0. Branch removals also change capacity, so the results do not uniquely establish functional specialization. e-ablation |
| Teacher-depth latency–success trade-off LIBERO-Plus layer-selection sweep; latency is per action chunk. | Layer 20: 86.1 / 0.351 Success (%) / seconds per action chunk | Layer 0: 82.1 / 0.114; layer 40: 86.2 / 0.587. Intermediate extraction gives most of the observed success improvement. Layer 20 is not the fastest setting; the source does not give a complete latency measurement protocol. e-depthe-repro-gap |
| Real-world long-horizon manipulation 45 trials per task; all required subgoals must be completed in order under continuous visual feedback. | Clothes folding 73.3; fruit sorting 88.9; cube catching 86.7; unmanned retail 84.4; average 83.3. Task success and unweighted four-task average (%) | π0.5 average 80.0 and clothes folding 77.8; Cosmos Policy average 70.0; Fast-WAM 70.6. Physical execution evidence supports the reported task-level gains, but π0.5 wins folding. No confidence intervals establish the significance of the 3.3-point average gap. e-reale-real-protocole-statistics |
| Standard LIBERO in-domain evaluation Spatial, Object, Goal and Long suites using the stated LIBERO-initialized student. | 98.8 Average success rate (%) | Cosmos Policy 98.5; Long suite 97.8 versus 97.6. The saturated benchmark shows a small improvement; this is a separate protocol from LIBERO-Plus. e-libero |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. Both transfer branches and the balancing objective contribute in the reported setup. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the bottom row as the reference and read upward through one removal at a time. Without context, the total matches the Cosmos-Policy baseline at 82.2. Removing the generalized branch leaves 83.0; removing specialists leaves 82.7. Their respective gaps from 86.1 are 3.1 and 3.4 percentage points. Dropping the auxiliary balancing objective leaves 84.1 while retaining the architecture, making this the cleaner comparison for the regularizer. Scan across the perturbation columns to see that the full configuration also improves each listed category relative to these ablations. Section 4.4 explains which branch is retained in each removal. e-ablatione-main-resultse-statistics
What it supports. The full system performs better than every tested removal. The largest drop among individual component removals comes from removing specialists; removing all context is worse still. Balancing contributes a smaller but visible two-point difference while leaving the transfer architecture in place. These observations motivate testing both representation content and routing behavior.
Where the evidence stops. Removing a branch also removes capacity and conditioning structure. The table does not distinguish expert specialization from those changes, and it provides no run-to-run uncertainty for the observed differences.
Figure 2. Intermediate teacher features give a practical latency–success compromise. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read blue circles against the left success-rate axis and red squares against the right latency axis. Their vertical positions use different units and cannot be compared directly. The horizontal axis is the selected teacher layer, ℓ*. At layer zero the labels show 82.1% and 0.114 seconds per action chunk; at the dashed layer-20 reference they show 86.1% and 0.351 seconds. At layer 40 they show 86.2% and 0.587 seconds. Success varies little after the middle layers, while latency continues upward. This is consistent with using an intermediate teacher prefix to build context once and reusing it during the student's denoising. e-depthe-teachere-repro-gape-statistics
What it supports. Layer 20 captures almost all the improvement visible at layer 40 while avoiding much of its additional latency. However, context transfer is slower than the layer-zero point: the graph supports a cost–utility compromise, not a claim that adding the teacher accelerates the unaugmented student.
Where the evidence stops. No error bars are shown, and the detailed timing hardware, batch, denoising and action-horizon protocol is not provided. Seconds per action chunk cannot be converted into a low-level robot control frequency from this figure alone.
7. Analysis & limitations
7.1 What the evidence leaves open
Only the DreamZero–Cosmos pairing is evaluated. Broader heterogeneous-model transfer remains untested; the authors explicitly leave non-robotic language and vision-language applications to future work. e-confige-limit
Figure 1's performance inset says 85.4%, inconsistent with 86.1% in Tables 1–3. More substantially, the main equations describe shared compression plus residual specialists, whereas Appendix B.4 counts a projection/query/cross-attention module per adapter. The main settings specify 16 heads; the parameter appendix specifies eight. e-interfacee-main-resultse-layout-conflict
Routing plots show stage-dependent selection and a declining balancing loss, but neither proves causal skill decomposition. Tables report point estimates without uncertainty; the checklist cites trial counts as its justification for statistical reporting. e-diagnosticse-statistics
Cube-catching descriptions are not fully consistent: Figure 3 and Appendix B.1 include opening and closing a drawer, while Appendix B.3's prose emphasizes grasping and transporting a cube. Exact success conditions require reconciliation. e-real-protocol
7.2 Questions for discussion
- Would matched versus shuffled teacher context separate observation-specific transfer from extra conditioning capacity? (e-teacher, e-ablation)
- Can the shared-compressor equations reproduce the appendix's parameter count and latency, or do the reported experiments require independent per-adapter compressors? (e-layout-conflict, e-depth)
8. Reproducibility audit
8.1 Requirements and known gaps
Required assets include the specified teacher/student checkpoints, LIBERO training data, OOD evaluation configurations and the physical-task data/setup. Table 5 fixes bottleneck width 512, 32 queries per branch, eight specialists and top-2 routing. Tables 7–8 report 187,411,465 trainable parameters: 1.17% of approximately 16B combined backbone parameters, or 9.37% of the 2B student. e-confige-protocole-budgete-real-protocol
Resolve shared versus per-adapter compression, eight versus 16 attention heads, and the separator's mandatory versus optional status before claiming architectural reproduction. Batch size, exact training duration, software versions, denoising/action horizons, detailed latency hardware/procedure and real-world training-data scale are not sufficiently specified in the supplied PDF. e-layout-conflicte-injectione-positionale-repro-gap
The authors claim a code release, minimal example and environment instructions in Appendix A.2. These are availability claims in the PDF, not verified runnable artifacts. e-limit
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Does correctly matched teacher context carry useful scene information?
Reader-proposed experiment, not performed: after resolving the adapter specification, train identical interfaces with correctly paired teacher context versus context shuffled across observations within the same task, keeping the teacher/student checkpoints, context length, parameter budget and optimization schedule fixed. Include a learned input-independent context of the same length as a capacity control. Evaluate identical LIBERO-Plus episodes across several training seeds, reporting per-shift success and uncertainty as well as total success, latency and peak memory. A reproducible advantage for matched context over both controls would support observation-specific transfer. Similar performance after shuffling would weaken the claim that the gains require current-scene teacher knowledge. e-teachere-compressione-protocole-ablatione-budgete-statistics
Check 2: Can the published interface and parameter budget be satisfied together?
Reader-proposed audit, not performed: instantiate the main-text shared-compressor design and Appendix B.4's per-adapter design separately, recording tensor shapes, actual trainable counts and executed specialist branches before any benchmark comparison. Test the documented eight-head and 16-head settings at fixed feature width, and explicitly account for the separator. Compare counts against Table 7's 187,411,465; a mismatch identifies a specification that cannot reproduce the reported budget. For each implementation, hold observations and noise fixed, permute already embedded conditioning rows with their masks, and compare student predictions and visual RoPE indices in evaluation mode. Unchanged rotation matrices and numerically matching outputs are the stated invariance prediction; deviations would expose a positional, masking or injection mismatch. e-compressione-layout-conflicte-budgete-injectione-positional
8.3 Reading coverage
Visual audit: Read all nine text chunks, including all 31 pages. Actually rendered and visually inspected the declared pages: title/version/affiliations; architecture and method equations; all Tables 1–9; Figures 1–7; appendix settings, budget, injection proof and hook, routing/training diagnostics, limitations and the statistical-reporting checklist response. Six original 300-DPI crops were separately viewed and checked for complete labels, axes, legends and table rows. Figure 1's 85.4% inset, Table 1's erroneous language emphasis, and main/appendix implementation discrepancies are explicitly retained in the explanations. Reference pages and remaining checklist pages were read as text; separate supplements, code, external datasets and experiments were not inspected or reproduced.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 27. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title, authors, affiliations, version and abstract (p. 1)
- Sections 1–2: Introduction and Related Work (pp. 1–2)
- Sections 3.1–3.5: architecture, teacher extraction, compression, routing, injection and objectives (pp. 2–6)
- Sections 4.1–4.4: simulation settings, results, PEFT comparisons and ablations (pp. 6–8)
- Sections 5–6: real-world experiments and conclusion (pp. 8–9)
- References (pp. 10–13)
- Appendix A.1–A.4: disclosure, reproducibility, impact and limitations (p. 14)
- Appendix B.1–B.7: routing diagnostics, configurations, real-world protocol, parameter accounting, injection proof, training curves and LIBERO results (pp. 14–22)
- Appendix C: load-balancing implementation (pp. 22–23)
- Appendix D.1–D.3: detailed related work (pp. 23–24)
- NeurIPS Paper Checklist, items 1–16 (pp. 25–31)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- All nine supplied text chunks were individually read, covering all 31 PDF pages. The reviewed artifact is arXiv:2605.06247v1 [cs.RO], 7 May 2026. Its title and all ten authors match the catalog; the title page additionally lists LivsynRobotics, omitted from the catalog affiliation string. No other revision was supplied or compared.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This supplied extraction limitation was addressed by visually inspecting Figures 1–7, Tables 1–9 and the claim-supporting equation pages in the retained PDF.
- Separate supplemental material availability has not been fully verified.
- The linked code, checkpoints and external datasets were not inspected; no experiments were reproduced. The checklist and references were read as source content, not independently verified claims.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title, author/affiliation block and arXiv margin
Exact observed title matches the catalog. Authors: Yuhua Jiang, Yijun Guo, Hongbing Yang, Guojun Lei, Nuo Chen, Yinuo Zhang, Shaoqiang Yan, Bo Lin, Feifei Gao, Biqing Qi. Affiliations: Tsinghua University; LivsynRobotics; Shanghai AI Laboratory. Margin identifies 2605.06247v1, 7 May 2026.
Go to primary source ↓e-problemPDF p. 2, Introduction paragraphs 1–2 and Section 2
Motivates context transfer by heterogeneous latent/action interfaces, adaptation expense and rigid output or hidden-state matching.
Go to primary source ↓e-interfacePDF p. 3, Figure 1 and Section 3.1 continuation
Teacher states feed compression, a router, generalized and specialized adapters, then student conditioning. Frozen/trainable symbols distinguish backbones from the transfer module. Figure inset shows LoRA 82.8% to Ours 85.4%, unlike the result tables.
Go to primary source ↓e-teacherPDF pp. 3–4, Section 3.2, Equation (1)
Intermediate teacher extraction uses image/text tokens at t*=0 once; no teacher future-video/action rollout is used. Context is reused over student denoising.
Go to primary source ↓e-compressionPDF p. 4, Section 3.3, Equations (2)–(5)
Main formulation uses shared bottleneck projection, two query banks and learnable-query cross-attention, followed by generalized and specialized residual bottleneck adapters.
Go to primary source ↓e-routingPDF pp. 4–5, Section 3.3, router paragraph and Equations (6)–(7)
Mean teacher states drive softmax routing; top-k weights are renormalized and only selected specialist branches execute. Generalized/specialized contexts concatenate; trainable and active parameter counts differ.
Go to primary source ↓e-injectionPDF p. 5, Section 3.4, Equations (8)–(9)
Text embeddings, learnable separator and teacher context concatenate as student cross-attention conditioning. Context does not join the visual self-attention sequence.
Go to primary source ↓e-objectivePDF pp. 5–6, Section 3.5, Equations (10)–(12); pp. 22–23, Appendix C, Equations (23)–(26)
Frozen-backbone training optimizes masked noise-weighted clean video/action reconstruction using shared log-normal noise levels and modality-specific Gaussian noise. Balancing uses mean routing probabilities and top-k frequencies normalized by kB.
Go to primary source ↓e-configPDF p. 6, Section 4.1; p. 16, Table 5
DreamZero-14B teacher, Cosmos-Policy-2B student, layer 20/40; eight A800s, bfloat16, AdamW β=(0.9,0.95), weight decay 0.01, 1k warmup/cosine decay, adapter learning rate 3e-4, bottleneck 512, Kg=Ks=32, 16 heads, dropout 0.1, M=8/top-2, Pmean=1.39, Pstd=1.2, λvid=1, λbal=0.01.
Go to primary source ↓e-protocolPDF pp. 6–7, Section 4.2, evaluation setup
Student checkpoint is Cosmos-Policy-LIBERO-Predict2-2B; training combines four LIBERO suites. LIBERO-Plus zero-shot evaluation covers seven named OOD dimensions.
Go to primary source ↓e-main-resultsPDF p. 6, Table 1, CKT-WAM/Cosmos-Policy/π0.5/VLA-JEPA rows; p. 7, Section 4.2 interpretation
Total success is 86.1/82.2/85.7 for CKT/Cosmos/π0.5. CKT language 86.7 is below VLA-JEPA 88.1 despite bolding and the prose's best-language claim. CKT leads camera, light, noise, layout and total.
Go to primary source ↓e-peftPDF p. 7, Table 2 and Section 4.3
Success/trainable ratios: CKT 86.1/1.17%, LoRA 82.8/2.59%, GOAT 84.8/1.77%, full fine-tuning 86.3/100%. Baselines project teacher states with a two-layer MLP and adapt the first selected teacher layers and whole student.
Go to primary source ↓e-ablationPDF pp. 7–8, Section 4.4; p. 8, Table 3, Total column
Full 86.1; without context 82.2; without generalized adapter 83.0; without specialized adapters 82.7; without auxiliary loss 84.1. Auxiliary removal keeps the full architecture.
Go to primary source ↓e-depthPDF p. 8, Figure 2, labeled layer 0/20/40 points and Section 4.4
Success/latency is 82.1%/0.114 s at layer 0, 86.1%/0.351 s at layer 20, and 86.2%/0.587 s at layer 40. Axes specify seconds per action chunk, not per low-level control step.
Go to primary source ↓e-realPDF pp. 8–9, Section 5; p. 9, Table 4
45 trials per task. CKT success: folding 73.3, sorting 88.9, cube 86.7, retail 84.4, average 83.3. π0.5 averages 80.0 and folding 77.8; Cosmos Policy averages 70.0; Fast-WAM 70.6.
Go to primary source ↓e-real-protocolPDF p. 9, Figure 3; p. 14, Appendix B.1; pp. 15–17, Appendix B.3 and Figures 5–6
Success requires ordered subgoal completion with continuous visual feedback. Folding removes distractions, unfolds and folds; sorting sequentially places fruit; retail retrieves requested beverages. Drawer opening/closing appears in Figure 3 and B.1 cube stages, whereas B.3 prose describes grasp/transport/release.
Go to primary source ↓e-budgetPDF p. 18, Tables 6–8 and adjacent parameter-accounting text
Single adapter 20,531,712 parameters; eight specialists 164,253,696; router 2,626,057; total 187,411,465. Table 8 reports 9.37% relative to the 2B student, 1.34% to 14B teacher, and 1.17% to their approximately 16B combined size. All specialist weights count despite top-k selection.
Go to primary source ↓e-layout-conflictPDF p. 4, Equations (2)–(5); p. 6, Section 4.1; p. 16, Table 5 and Appendix B.4; p. 18, Tables 6–7
Main text places shared compression before residual bottleneck adapters and specifies 16 attention heads. B.4 describes projection/query/MHA inside each adapter, specifies eight heads, and counts nine separate full adapter branches. The PDF does not reconcile these descriptions.
Go to primary source ↓e-positionalPDF pp. 19–20, Appendix B.5.1–B.5.3, Equations (13)–(20), Proposition 1 and corollary
Only cross-attention K/V conditioning is changed. Visual-grid-based RoPE matrices retain their indexing; conditioning cross-attention has no positional encoding. A text-projection forward hook adds cached context per observation across denoising/CFG calls; appendix calls the separator optional.
Go to primary source ↓e-diagnosticsPDF pp. 14–15, Appendix B.1/Figure 4; pp. 21–22, Appendix B.6/Figure 7
Stage-dependent adapter selection is analyzed over 120 cube-catching episodes. Training curves show stable averaged losses and a balancing loss declining approximately 1.12 to 1.02. These are observational diagnostics rather than interventions on specialist roles.
Go to primary source ↓e-liberoPDF p. 22, Table 9 and Appendix B.7
Standard LIBERO average: CKT 98.8, Cosmos Policy 98.5. CKT Spatial/Object/Goal/Long: 99.0/99.8/98.6/97.8; Cosmos Long 97.6.
Go to primary source ↓e-limitPDF p. 14, Appendix A.2 and A.4
Authors claim open-source code, minimal example and environment instructions. Limitations explicitly leave transfer beyond robotics, including text modeling and VQA, untested.
Go to primary source ↓e-repro-gapPDF pp. 6–9, Sections 4–5; pp. 15–18, Appendices B.2–B.4; pp. 20–21, B.5.3 and Figure 7
Settings provide named hardware and many hyperparameters but do not fully specify batch size, exact training duration, software versions, action/video horizons, inference denoising schedule, latency measurement setup or physical training-data scale. The training-curve iteration axis does not establish every experiment's duration.
Go to primary source ↓e-statisticsPDF pp. 6–9, Tables 1–4; p. 22, Table 9; p. 27, checklist item 7
Result tables contain point estimates without error bars or confidence intervals. Checklist item 7 answers Yes on significance reporting with the justification that real-world trial counts are given.
Go to primary source ↓e-taxonomyPDF p. 23, Appendix D.1 final paragraph; p. 3, Figure 1; pp. 5–6, Section 3.5
Authors explicitly distinguish transfer between WAMs from proposing a new monolithic WAM. Diagram has frozen teacher and student connected by context; training supervises action and video outputs without introducing an inverse-dynamics action extractor.
Go to primary source ↓8.5 Primary sources
CKT-WAM: Parameter-Efficient Context Knowledge Transfer Between World Action Models ↗
PDF · 14,985 extracted words
Source fingerprint
bf270cb8e9926ac7f5f2868e3c3afbae675cd91b3a731cb3c45ba97e248cd722