DIM-WAM: World-Action Modeling with Diverse Historical Event Memory
1. Paper overview
In one sentence: DiM-WAM improves temporally dependent manipulation by conditioning joint video/action denoising on bounded observation memory, while relying on coarse progress supervision whose transfer beyond the tested tasks remains unresolved. e01e02e03e04e05e11e12e13e15
| At a glance | What to know |
|---|---|
| Research problem | Source description Similar local observations can require different actions after different histories: a switch may need turning on or off, and a displaced block must return to its earlier location. A short WAM cache loses that evidence. The paper seeks compact persistent memory without continually enlarging local context. e02e03e10 |
| Core mechanism | |
| A key reported result | RMBench: nine-task average: DiM-WAM 69.8 overall; M(1) 80.6; M(n) 56.3. Full-task success rate (%). 50 shared demonstrations and 100 evaluation rollouts per task; LingBot-VA/DiM-WAM share non-method optimization settings and 1,500 steps. Head + front cameras, 30 frames, 128×128 inputs; strides 1 for M(1), 2 for M(n). Training-matched LingBot-VA: 34.8 overall; M(1) 34.2; M(n) 35.5. Overall gain: 35.0 percentage points. Supports the combined DHEM/progress intervention. Mem-0 at 42.0% is benchmark-reported context; Fast-WAM at 4.0% is protocol matched but uses its own optimization recipe. e10e11e12 |
| Reading caution | Source description The audit covers only Put Back Block. Wrist-view success rises from 41% to 86% with the initial state outside the window, suggesting pose-dependent leakage. Other tasks may retain leakage. “Chance” is target-choice accuracy, not a full-task upper bound. e10 |
Core contributions
Figure 2(a–c). Observation memory supplies cross-stage context to video and action denoising. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow panel (a) from observed history through bank-specific queries, low-rank cross-attention and temporal-semantic merging. The colored bank tokens receive identity embeddings, are flattened, and gain slot-relative RoPE before conditioning the Wan-DiT alongside local KV context. Video noise and action noise enter the same depicted denoising block. Panel (b) separates clean conditioning from noisy prediction targets and illustrates progress supervision. Its progress-to-prediction arrow should be read as training influence: Section III-E explicitly excludes the progress head from inference. Panel (c) illustrates weighted fusion within middle history; the complete admission/discard rule and initialization cases are in Algorithm 1. The persistent memory contains observations, not imagined futures. e03e04e05e06e07e08
What it supports. DHEM separates bounded historical storage from current prediction. Banks can preserve different summaries of the same observations, and their readout conditions both predicted visual evolution and executable actions. This architecture supplies a route for earlier events to affect control without extending the persistent local KV cache with memory tokens.
Where the evidence stops. Panel (c) has a numerical inconsistency: D and E carry masses 4 and 2, but G is labeled mass 5. Equation (11) requires mass 6; the displayed time 32 is consistent. The crop is unchanged. Use Equation (11) and Algorithm 1 for the update.
2. Motivation
2.1 The problem and the proposed response
Similar local observations can require different actions after different histories: a switch may need turning on or off, and a displaced block must return to its earlier location. A short WAM cache loses that evidence. The paper seeks compact persistent memory without continually enlarging local context. e02e03e10
2.2 What this reading follows
A robot can arrive at nearly the same image after different histories and still need a different next action. DiM-WAM addresses this ambiguity by adding persistent event memory to LingBot-VA. Its banks summarize actual observations, retain relatively novel events and compress redundant history before returning that evidence to video/action denoising. This reading follows the supplied v2, whose results differ from the catalog abstract. The strongest comparison matches LingBot-VA and DiM-WAM in demonstrations and non-method training settings. The illustrations below explain the memory update, show why the evaluation protocol matters, and separate executed robot success from weaker evidence about bank specialization. e01e02e03e04e05e11e12e13e15
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 | Dual-system |
| Prediction paradigm | IDM |
| Quadrant | Q4 · Dual-system × IDM |
3.1 Evidence-based assessment
Conflicts with the recorded classification
WAM and memory categories are supported. Figure 2(a) depicts one Wan-DiT receiving video/action noise and DHEM context; Equation (2) jointly outputs video and actions. Figure 3 permits action queries to see same-segment video tokens, but does not establish a separate inverse-dynamics model. The depicted architecture favors joint prediction in one denoising framework over Dual-system × IDM. Inherited backbone internals are incompletely specified; the catalog snapshot remains unchanged. e03e07
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 Why the latest image is not enough
Put Back Block makes the information problem concrete. The robot first encounters a block at one of four locations, moves it to a common intermediate state, presses a button and must then return it to the original place. A local policy can lose the distinguishing observation by the time the return action is needed. DHEM writes summaries of actual observations into several banks so that this earlier evidence can survive. The initial-state anchor is protected, while the latest slot contains the latest retained event; a redundant new observation can be discarded, so that slot need not represent the most recent observation received. The middle slots compress history through adjacent-pair merging. Bank queries determine candidate content, while each bank's retained state determines its subsequent maintenance decisions. e03e04e05e06e10
5.2 Separate event time, readout order and supervised progress
Three temporal quantities serve different purposes. Stored timestamps enter the redundancy score and become mass-weighted representative times after fusion. Readout positions instead follow retained slot order: Equation (12) spaces their RoPE coordinates by two and gives corresponding slots in different banks the same coordinate. Bank identity embeddings preserve which bank supplied each token. Finally, the progress target comes from normalized position within a demonstration, discretized into bins and predicted from pooled memory. It does not require explicit semantic stage labels. This head shapes representations during training and is absent from the inference decision path. At inference, the readable event sequence conditions video/action denoising, and subsequent environment observations supply new memory writes. Consequently, predicted futures and the auxiliary progress estimate should not be described as stored observations or an explicit long-horizon planner. e04e06e07e08e09
5.3 Read the result as several levels of evidence
Reader analysis: the main table contains comparisons with different explanatory strength. LingBot-VA versus DiM-WAM is the controlled test of the combined intervention because demonstrations, non-method optimization settings and evaluation protocol match. Fast-WAM shares the protocol but keeps a separate optimization recipe; benchmark-reported VLA values provide context without the same control. The ablation then narrows the mechanism: one versus four banks at 32 total slots tests organization more directly, while increasing to eight 12-slot banks changes organization and capacity together. Real-robot results add evidence of executed behavior, but ten trials and the Press Twice ceiling limit precision and discrimination. Finally, the single-episode PCA plot illustrates a possible internal behavior rather than establishing its causal role. These levels support useful memory without resolving every architectural explanation. e11e12e13e14e15e16
5.4 Training and inference
During training
A mean-pooled memory readout predicts ten progress bins from normalized demonstration position: i/(T−1) for T>1, zero for a single segment. These are temporal proxies, not semantic stage annotations. Squared cosine similarity between normalized bank means discourages collapse. e08e09e11
Training uses AdamW at 10⁻⁵, batch size 1, gradient accumulation 10, 1,500 steps, eight 12-slot banks, clean-KV dropout 0.3 and four-segment truncated backpropagation. Auxiliary weights are λ_div=10⁻³ and λ_prog=10⁻². Each task-specific model uses eight NVIDIA H800 GPUs; comparisons are not compute matched. e11
During inference
Instruction, local context and memory condition future-latent/action prediction. Actions control the robot; subsequent actual observations update memory. The progress head is excluded from inference decisions. Executed action-chunk length and denoising-step count are not specified in this PDF. e03e04e07e08e17
5.5 Implementation flow
- Represent event history
Each of K banks has N slots storing an event vector, timestamp, accumulated mass and validity. The first slot anchors the initial observation; the last holds the latest retained event. Storage is O(KN), independent of trajectory length. e03e05e06
- Write observation-grounded candidates
After a segment, each bank's learnable query attends normalized visual features through shared projections; an FFN residual produces a candidate. Banks see the same actual observation but can summarize different information. Generated futures are not memory writes. e04
- Retain novelty and compress redundancy
When full, compare incoming-to-latest redundancy with the highest redundancy among adjacent middle-history pairs. Discard the incoming candidate at greater-or-equal score; otherwise merge the historical pair and admit it. Fusion weights vectors and timestamps by accumulated mass. Initialization duplicates are removed before ordinary compression. e05e06
- Read memory into denoising
Flatten valid tokens in slot-major, bank-minor order, add bank identity and apply RoPE with two-coordinate slot spacing. RoPE follows retained slot order, not mass-weighted timestamps. Readout conditions video/action prediction without appending event tokens to the persistent KV cache. e07
6. Experiments & results
DiM-WAM augments LingBot-VA with bounded, observation-grounded event memory that conditions future-video and action denoising. Bank-specific summaries preserve cross-stage evidence, while auxiliary trajectory-progress supervision shapes memory. The inspected v2 reports RMBench success of 69.8% versus 34.8% in a training-matched comparison, and 90.0% versus 52.5% full-task success on four Franka tasks. The combined design is supported; stable semantic bank specialization remains unestablished.
6.1 Read the original evidence
Table I. A nominal memory task can expose useful history through its observation protocol. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the first two rows, which hold the 30-frame window and stride 1 fixed while changing the camera set. Both mark the initial state as outside the local window, yet success differs sharply. Then read down the head-plus-front rows: expanding temporal coverage through a longer window or wider frame subsampling can make the initial state visible again. Window counts input frames; stride is the input-frame subsampling interval, not an action horizon. The source caption defines “Chance” as ideal target-choice accuracy from locally visible target information. It is not an upper bound on full-task success, so the Chance and Success columns measure different things. e10e11
What it supports. Adding the wrist view raises success from 41% to 86% even when the initial state is outside the window. The authors interpret this as possible pose-dependent leakage and choose head-plus-front, window 30 and stride 1 for the M(1) controlled runs. Thus observation design affects how convincingly success tests historical memory.
Where the evidence stops. This audit uses LingBot-VA's official training-parameter configuration and covers only Put Back Block. It motivates the selected protocol but cannot establish that every remaining benchmark task is free of local cues or other leakage.
Table II. The controlled gain is between LingBot-VA and DiM-WAM, within the dagger-marked protocol. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the last two columns first: LingBot-VA and DiM-WAM share the audit-guided camera/window protocol, demonstrations and non-method optimization settings. The dagger also marks Fast-WAM as protocol matched, but that model retains its own optimization recipe. DP, ACT, π0.5, X-VLA and Mem-0 are benchmark-reported reference values. The TMC column divides tasks into M(1), requiring a few key past observations, and M(n), requiring accumulated interaction history. Group averages summarize those subsets; the final row averages all nine tasks. Check individual rows before treating the overall average as uniform competence: Observe and Pick Up remains difficult, and Mem-0 has the highest Cover Blocks entry. e10e11e12
What it supports. The matched WAM comparison rises from 34.8% to 69.8% overall, a 35.0-percentage-point gain. M(1) improves from 34.2% to 80.6%, and M(n) from 35.5% to 56.3%. These are reported task-success rates supporting the combined memory/progress design under this protocol, rather than measurements of generated-video quality.
Where the evidence stops. The 69.8% versus Mem-0's 42.0% difference is contextual, not an isolated memory-design effect under matched training. The comparisons are not compute matched, and the table supplies no repeated-seed uncertainty estimates.
Table III. The robot evaluation counts completed stages and fully completed trials separately. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each task has paired SSR and SR columns. SSR counts completed stages out of evaluated stage opportunities; SR counts trials that complete the whole task. The denominators are 40 stages for Find Blue Block, 60 for each swap task and 30 for Press Twice, with ten full-task trials everywhere. Only the final average columns are percentages. Compare DiM-WAM and LingBot-VA within the same task before using the averages. All methods use the same 15–25 demonstrations per task and a single third-person camera, but other baselines have different optimization recipes. These counts concern physical Franka Panda executions; partial stage completion does not equal full-task success. e11e13
What it supports. DiM-WAM reports 93.5% average SSR and 90.0% average SR, versus 70.6% and 52.5% for LingBot-VA. Find Blue Block improves from 1/10 to 9/10 completed trials and Triangle Swap from 4/10 to 8/10. Press Twice is tied at 10/10 and provides no observed memory-specific advantage.
Where the evidence stops. Ten trials per task yield point estimates. Figure 5's narrative step counts differ from the scored-stage denominators for Find Blue Block and Press Twice; their mapping is unspecified. The local window covers about 85% of Press Twice demonstrations, explaining its limited discrimination.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RMBench: nine-task average 50 shared demonstrations and 100 evaluation rollouts per task; LingBot-VA/DiM-WAM share non-method optimization settings and 1,500 steps. Head + front cameras, 30 frames, 128×128 inputs; strides 1 for M(1), 2 for M(n). | DiM-WAM 69.8 overall; M(1) 80.6; M(n) 56.3. Full-task success rate (%) | Training-matched LingBot-VA: 34.8 overall; M(1) 34.2; M(n) 35.5. Overall gain: 35.0 percentage points. Supports the combined DHEM/progress intervention. Mem-0 at 42.0% is benchmark-reported context; Fast-WAM at 4.0% is protocol matched but uses its own optimization recipe. e10e11e12 |
| Franka Panda: four-task execution Same 15–25 demonstrations per task across methods; ten full-task trials each; one third-person 224×224 camera. | DiM-WAM SSR 93.5%, SR 90.0%; task SR counts: Find Blue Block 9/10, Line Swap 9/10, Triangle Swap 8/10, Press Twice 10/10. Average stage success ratio (SSR) and full-task success rate (SR) | LingBot-VA SSR 70.6%, SR 52.5%; corresponding SR counts 1/10, 6/10, 4/10, 10/10. Executed robot outcomes: SSR counts completed stages; SR requires completing a trial. Ten trials yield point estimates. Press Twice has a ceiling effect: the local window covers approximately 85% of demonstrations. e11e13 |
| Swap T and Swap Blocks: memory ablation Table IV, two RMBench tasks; variants change bank structure/capacity or remove progress supervision. | 1×32: 75.5; 4×8: 90.0; 8×12: 96.5; 8×12 without progress head: 90.5. Mean success rate (%) | At 32 total slots, multiple banks gain 14.5 percentage points. Removing progress supervision from 8×12 loses 6.0 points. The 4×8 versus 8×12 comparison changes bank count and capacity together. Maintenance subcomponents are not separately ablated. e14e16 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table IV. The fixed-capacity comparison supports multiple banks; the larger-memory comparison mixes two changes. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Interpret each layout as number of banks multiplied by slots per bank. Start with 1×32 and 4×8: both retain 32 slots in total, making this the most direct comparison of how storage is organized. Next compare 4×8 with 8×12, remembering that the latter has 96 slots and twice as many banks. Finally compare the last two rows, which retain the same 8×12 memory configuration while removing the progress head and its auxiliary supervision. Swap T and Swap Blocks are separate task columns, and Avg. is their mean. This table tests selected design choices, not all admission, merging and readout mechanisms independently. e08e09e14e16
What it supports. At fixed 32-slot capacity, the reported mean rises from 75.5% to 90.0% with four banks. The full 8×12 model reaches 96.5%; removing progress supervision reduces it to 90.5%. Both tasks move in the same direction for these two comparisons, supporting bank organization and auxiliary progress cues on the tested pair.
Where the evidence stops. Scaling from 4×8 to 8×12 cannot isolate capacity from bank count. Only two tasks are ablated, without a displayed uncertainty analysis or separate tests of novelty selection, mass weighting and the diversity penalty.
Figure 6(a–b). One successful episode illustrates differentiated retention without proving stable semantic specialization. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. In panel (a), follow one horizontal bank trace along normalized task progress, then compare retained event positions across traces. The colors identify banks; differing positions show that their retained histories need not coincide. Panel (b) uses the corresponding bank colors and marker shapes to display a two-dimensional PCA projection of event tokens. Its axes report 37.3% and 16.4% of variance for PC1 and PC2. Compare both separation and overlap: some bank groups are compact and distant, while others occupy overlapping regions. Section IV-F identifies this as one successful Put Back Block episode. Read it as a diagnostic of that retained state, not a task-level performance comparison. e15e16
What it supports. The plotted banks preserve different temporal samples and exhibit distinguishable token distributions in this episode. This is consistent with the intended complementary memory behavior. It does not identify what semantic concept any bank represents, demonstrate that its role recurs in another episode, or establish that the separation causes higher control success.
Where the evidence stops. The source explicitly limits this observation to one successful episode. Neither the PCA clusters nor the displayed ellipses establish stable semantic roles or confidence intervals. Failed episodes, cross-task consistency and intervention-based bank importance are not demonstrated here.
7. Analysis & limitations
7.1 What the evidence leaves open
The audit covers only Put Back Block. Wrist-view success rises from 41% to 86% with the initial state outside the window, suggesting pose-dependent leakage. Other tasks may retain leakage. “Chance” is target-choice accuracy, not a full-task upper bound. e10
Broader task, embodiment and observation generalization remains unestablished. Trajectory-position supervision may misrepresent semantic progress when stage durations vary. Figure 6 shows one successful episode, insufficient to establish stable semantic bank roles. e15e16
Figure 2(c) labels the D/E fused event with mass 5 despite input masses 4 and 2. Equation (11) requires 6; the illustrated timestamp 32 is consistent. The original graphic is retained, and the equation governs this explanation. e06
Figure 5 lists three narrative steps for Find Blue Block and four for Press Twice, whereas Table III uses 40 and 30 stage opportunities across ten trials. Their relation to the scoring denominators is unexplained. e13
7.2 Questions for discussion
- Would bank diversity remain helpful at equal capacity and compute across more than two tasks?
- Does progress supervision encode transferable event completion or sensitivity to demonstration timing?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires LingBot-VA, RoboTwin 2.0/RMBench, matching demonstrations and camera windows, DHEM updates and joint losses. Robot evaluation additionally requires Franka tasks and scoring. Exact checkpoint, frozen-module policy, numeric horizons, denoising schedule, seeds and detailed stage-scoring rules are not fully specified. e10e11e13e17
Proposed checks: compare bank layouts at matched capacity across repeated seeds, and vary demonstration stage durations while fixing memory and data splits to test progress supervision under timing changes. Neither experiment was run. e08e14e16
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 bank organization help when capacity is controlled?
Reader-proposed experiment: repeat 1×32 versus 4×8 on Swap T and Swap Blocks with identical demonstrations, camera protocol, backbone initialization and optimizer settings, using at least five paired seeds and 100 evaluation rollouts per task per seed. Add 1×96, 4×24 and 8×12 to separate bank count from total slots at the larger capacity. Specify how the single-bank case handles the diversity loss's empty bank-pair set, and record runtime and memory alongside success. Report paired differences and uncertainty. If multi-bank gains disappear across seeds at each fixed capacity, the organization-specific interpretation weakens even if the largest configuration remains best. e09e11e14e16
Check 2: Does progress supervision survive altered stage duration?
Reader-proposed experiment: train paired 8×12 models with and without progress supervision on the same demonstrations and training budget. Evaluate both on the same held-out tasks with variable pauses inserted at fixed semantic states, preserving event order and target identity. Use the same perturbed observations for both models, repeat across seeds, and measure full-task success plus the first action chosen after each pause. In a second matched training condition, redistribute stage durations and recompute the source's trajectory-position labels. If progress supervision loses or reverses its benefit under timing changes while the no-head control remains stable, that would challenge the temporal proxy's robustness. No such test was run here. e08e11e13e14e16
8.3 Reading coverage
Visual audit: All eight original PDF pages were rendered and actually viewed. The pass covered the title/authors/version/affiliations, Figures 1–6, all method equations and Algorithm 1, the visibility matrix, Tables I–IV, implementation/hardware details, real-task illustrations, limitations and references. All six final original crops were separately viewed; labels, table denominators, axes and legends are readable. Figure 2(c)'s fused-mass inconsistency and the Figure 5/Table III scoring-label ambiguity are disclosed. No separate supplement, code or external project material was inspected, and no experiment was reproduced.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. Appendix coverage: not present.
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
- Abstract and I. Introduction (p. 1)
- II. Related Work, A–C (p. 2)
- III. Method, A–E; Equations (1)–(19) and Algorithm 1 (pp. 2–5)
- IV. Experiments, A–F: setup, protocol audit, simulation, robot evaluation, ablations and bank analysis (pp. 5–7)
- V. Conclusion, limitations and future work (p. 7)
- Figure 6 and complete references [1]–[38] (p. 8)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Identity/version: the title page prints “DiM-WAM: World Action Modeling with Diverse Historical Event Memory,” with capitalization/hyphenation differing from the supplied observedTitle. All ten authors match the catalog in order. The margin identifies arXiv:2606.27677v2 [cs.RO], 12 Jul 2026 (e01). This report reads that supplied revision; v1 and an independent revision history were not supplied.
- The catalog abstract differs numerically from the inspected v2: it gives LingBot-VA RMBench success as 28.4%, robot stage success as 70.7% to 91.5%, and robot full-task success as 52.5% to 80.0%. The inspected source instead gives 34.8%, 70.6% to 93.5%, and 52.5% to 90.0%, respectively (e01, e12, e13). No equivalence between editions is assumed.
- 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 all eight PDF pages, Figures 1–6, Tables I–IV and every final crop.
- Separate supplemental material availability has not been fully verified. No separate supplement was supplied; the eight-page PDF contains no appendix.
- All four supplied text chunks were read completely. Code and the project page were not inspected, and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author/affiliation block, arXiv margin and abstract
The title is DiM-WAM: World Action Modeling with Diverse Historical Event Memory. Authors are Kai Wang, Zhaopeng Gu, Yixiang Chen, Yuan Xu, Qisen Ma, Jiabing Yang, Zhaowen Li, Yan Huang, Liang Wang and Peng Su. Affiliations are CASIA, Shenzhen Yinwang Intelligent Technology Co., Ltd., and FiveAges. The margin says arXiv:2606.27677v2 [cs.RO], 12 Jul 2026. Abstract results are 34.8% to 69.8%, 70.6% to 93.5%, and 52.5% to 90.0%.
Go to primary source ↓e02PDF p. 1, I. Introduction and Figure 1
Similar switch observations require different actions depending on interaction history. The introduction motivates complementary historical events and discusses the cost of expanding local context.
Go to primary source ↓e03PDF p. 2, III-A–B, Equations (1)–(3); p. 3, Figure 2(a)
Instruction, local KV context and persistent memory condition horizon-H visual/action prediction. Bounded event tuples store vector, time, mass and validity. Figure 2(a) depicts both noise modalities and context entering one Wan-DiT block.
Go to primary source ↓e04PDF p. 3, III-C(a), Equations (4)–(5)
Bank-specific normalized queries attend normalized observation features with shared projections; a residual FFN produces candidates. Stored tokens derive from real observations rather than generated predictions.
Go to primary source ↓e05PDF p. 3, III-C(b), Equations (6)–(10); p. 4, Algorithm 1, lines 4–31
Adjacent middle pairs are ranked by temporally decayed cosine redundancy. Incoming candidates are discarded at greater-or-equal redundancy. Initialization duplicates are removed first; otherwise a historical pair is merged, history compacted and the candidate admitted, preserving anchor and latest-retained slots.
Go to primary source ↓e06PDF p. 4, Equation (11) and surrounding III-C text; p. 3, Figure 2(c)
Fusion adds masses and averages tokens/timestamps by mass; τ=N−1. Figure 2(c) shows A/B masses 3 and 2 becoming 5 at time 13.2. D/E masses 4 and 2 lead to a label of mass 5 at time 32, inconsistent with additive mass in Equation (11).
Go to primary source ↓e07PDF p. 4, III-D, Equations (12)–(13); p. 5, Figure 3
Readout adds bank identity and slot-relative RoPE with two-coordinate spacing and slot-major/bank-minor flattening. Readout leaves stored content unchanged; event tokens are not appended to the persistent KV cache. Figure 3 has Q rows and K columns, with action rows including same-segment video columns.
Go to primary source ↓e08PDF p. 4, III-E, Equations (14)–(15); p. 5, Equation (16) and following paragraph; p. 3, Figure 2(b)
A head mean-pools valid readout tokens to predict bins of normalized trajectory position. The target handles T=1 and clips the endpoint to B−1. Cross-entropy is auxiliary; the head does not participate in inference decisions.
Go to primary source ↓e09PDF p. 5, III-E joint objective, Equations (17)–(19)
Diversity loss averages squared cosine similarity between normalized bank means. Video and masked-action diffusion denoising MSE have unit weights; diversity and progress terms have separate coefficients.
Go to primary source ↓e10PDF p. 5, Table I, IV-A and IV-B; p. 6, Figure 4
RMBench uses RoboTwin 2.0 with five M(1) and four M(n) tasks. Put Back Block requires returning to one of four initial positions. The official-configuration LingBot-VA audit varies cameras/window/stride; success is 86 versus 41 with/without wrist view at window 30, stride 1. Longer coverage exposes initial state. Chance is ideal target-choice accuracy, not a full-task bound. Only one task is audited.
Go to primary source ↓e11PDF pp. 5–6, IV-B implementation details and IV-C comparison setup
In-house WAMs share audit-guided views/window/resolution/strides. LingBot-VA and DiM-WAM additionally share 50 demonstrations, 1,500 steps and non-method settings, with 100 rollouts per task. Page 6 specifies AdamW 10⁻⁵, batch 1, accumulation 10, eight 12-slot banks, weights 10⁻³/10⁻², B=10, dropout 0.3, four-segment truncation and eight H800 GPUs. Robot inputs are 224×224; outputs are seven absolute joint positions and binary gripper. Fast-WAM and π0.5 have distinct recommended training recipes.
Go to primary source ↓e12PDF p. 7, Table II, all task and average rows; p. 6, IV-C
DiM-WAM/LingBot-VA averages are 69.8/34.8 overall, 80.6/34.2 for M(1), and 56.3/35.5 for M(n). Mem-0 42.0 is benchmark reported; Fast-WAM 4.0 is in-house protocol matched. DiM-WAM leads eight tasks; Cover Blocks is 56.0 versus Mem-0 68.0. Observe and Pick Up is 13.0 for DiM-WAM.
Go to primary source ↓e13PDF p. 6, IV-D and Figure 5; p. 7, Table III and continuation of IV-D
Each real task uses 15–25 shared demonstrations and ten trials. SSR denominators are 40/60/60/30; each SR denominator is 10. DiM-WAM SR successes are 9/9/8/10; LingBot-VA 1/6/4/10. Averages are SSR 93.5/70.6 and SR 90.0/52.5. Press Twice local coverage is approximately 85%, with both models at 10/10. Figure 5 lists three steps for Find Blue Block and four for Press Twice without explaining their relation to scored stage counts.
Go to primary source ↓e14PDF p. 7, Table IV and IV-E
Swap T/Swap Blocks rates are 71/80 for 1×32, 88/92 for 4×8, 97/96 for 8×12, and 92/89 without the progress head. Means are 75.5, 90.0, 96.5 and 90.5. The text identifies the capacity confound in scaling to 8×12.
Go to primary source ↓e15PDF p. 7, IV-F; p. 8, Figure 6(a–b) and caption
One successful Put Back Block episode illustrates different retained event positions and distinct but overlapping bank PCA distributions. PC1/PC2 labels report 37.3%/16.4% variance. The text does not infer a general trend or stable semantic specialization.
Go to primary source ↓e16PDF p. 7, V. Conclusion, Limitations and future work
The authors acknowledge limited task/embodiment/observation scope, incomplete maintenance-component isolation, coarse trajectory-position supervision and only episode-level evidence of differentiated retention.
Go to primary source ↓e17PDF pp. 2–5, III-A–E; pp. 5–6, IV-A–D
The PDF supplies a DHEM algorithm, symbolic horizon H and a partial optimizer/hardware recipe. It does not identify an exact base checkpoint, full frozen/trainable-module policy, numeric prediction/execution horizons, inference denoising schedule or seeds. Robot stage scoring is reported as aggregate counts without complete operational rules.
Go to primary source ↓8.5 Primary sources
DIM-WAM: World-Action Modeling with Diverse Historical Event Memory ↗
PDF · 5,253 extracted words
Source fingerprint
40e62192c505e65ea4d608121aed07225f9317d876b6c2a358f793c1a737af16