ICML 2026 — visual examples

168 of 168 oral records have a writing-focused manuscript study. Official program snapshot: 2026-10-07. Selected rendered pages were inspected for each study; this is not a proof audit or reproduction of experiments.

The dated official program contains 168 distinct paper Oral events: 159 Main Track and 9 Position Paper Track. Their decision field says Accept (spotlight); the wider spotlight pool is not the oral corpus. One authorless workshop timetable event is excluded explicitly.

Every paper is matched to PMLR volume 306 by a concrete OpenReview identity. Published titles and author lists are used alongside preserved program metadata. Writing-focused scans and selected raster-page checks are distinct from download completion, exhaustive technical validation and scientific-library promotion.

The inspected sources are published proceedings PDFs, including attached supporting material where present. Physical page spans are not official content-page accounting. Every study records the source and actual inspection scope.

Inspect the complete inventory and source status · Official oral program

Choose a visual for its explanatory job

Use a diagram to explain a mechanism or assumption, a plot to expose a relationship, a table for exact comparison, and aligned qualitative panels for visible behavior. Captions should define the setting, symbols, units and supported conclusion. Match the display to the evidence, not to decoration.

Read the display, then its design lesson. Tables and numeric charts below reconstruct explicitly attributed factual subsets. Diagrams are original teaching layouts, not screenshots or reproductions of the papers' architecture. Selected examples are not a prevalence ranking of visual styles.

Execution trace: concurrent work and its dependency

ThreadWeaver: Adaptive Threading for Efficient Parallel Reasoning in Language Models · Figures 2–3, physical p. 3

Original teaching construction. A simplified fork–join construction motivated by the inspected request timeline. It omits token markers, trie training and reward assignment; it is not the complete ThreadWeaver implementation.

Observed in the source: The source pairs a concrete reasoning trace with a request/completion timeline and labels the stop conditions.

Prepare: Label shared context, independent branches, the join dependency and what remains sequential. Reuse those names in pseudocode and the timing definition.

Do not infer: More parallel branches do not by themselves imply lower latency: the longest branch, scheduling and sequential work matter.

Latency table: disclose the extra hardware

ThreadWeaver: Adaptive Threading for Efficient Parallel Reasoning in Language Models · Table 3 and Section 5.5, physical p. 7

Execution	GPU allocation	Average latency (s, lower)
Sequential	1 GPU	162.34
Parallel	4 GPUs	142.21

Source-reported numerical subset. Table 3 values on 50 MATH500 problems, eight runs per problem, batch size 1. GPU allocation comes from the adjacent experimental description. The source reports 1.14× speedup and no uncertainty interval for these cells.

Observed in the source: The paper separates this measured wall-clock result from its token-latency ratios and higher per-task headline speedups.

Prepare: State workload, batch size, hardware allocation, repetition and whether timing includes orchestration. Keep quality evidence adjacent in the paper.

Do not infer: This is not a fixed-GPU-budget speedup or evidence of 1.53× end-to-end acceleration.

Representation diagram: preserve labels across reordering

WeDLM: Reconciling Diffusion Language Models with Standard Causal Attention for Fast Inference · Figure 2 and Section 4.1, physical pp. 4–5

Original teaching construction. An original four-token example of the distinction between physical order and logical position. The source uses a larger example and a separate dual-stream attention diagram. This construction omits the full mask and objective.

Observed in the source: The source keeps position IDs visible while moving observed tokens to the prefix, then explains which context masked positions can access.

Prepare: Put semantic identity and storage position on separate label rows. Show the relevant attention relation separately if the reordering alone does not define it.

Do not infer: Reordering does not make this model equivalent to a fully bidirectional diffusion model; the paper explicitly rejects that equivalence.

Comparison table: adaptation and retention are separate

Midtraining Bridges Pretraining and Posttraining Distributions · Table 4, physical p. 6

Axes focus on the displayed range, not a zero baseline. Point numbers match the named configurations below the plot; exact values and comparison conditions remain in the table and caption.

70M model, GSM8K	SFT validation loss (lower)	C4 validation loss (lower)
Pretrain-only	1.384	6.353
Math midtraining (12%)	1.339	6.358
Continued pretraining (Math)	1.383	6.376

Source-reported numerical subset. Selected three-row, one-model/one-task subset of Table 4, averaged across five seeds. The source table does not supply error bars for these values. Mixture and continued-pretraining schedules are defined in Section 5.2.

Observed in the source: The mixed schedule improves both displayed losses against specialist-only continued pretraining, but the original pretrain-only model has the lowest C4 loss in this subset.

Prepare: Separate adaptation and retention columns, state lower/higher direction and name the comparison used by the claim.

Do not infer: Do not turn a mixed-versus-specialist result into a claim of beating every baseline on every metric, or call the tiny retention difference statistically established.

Ablation table: baseline, single changes and combination

Mind Your Margin and Boundary: Are Your Distilled Datasets Truly Robust? · Table 3, physical p. 8

Curriculum AAC	Contrastive CRL	CIFAR-10 robust accuracy (%)	CIFAR-100 robust accuracy (%)
Off	Off	24.81	9.4
On	Off	26.73	11.45
Off	On	26.04	10.98
On	On	28.41	12.3

Source-reported numerical subset. Two-dataset subset of the four-dataset component ablation; images per class = 10, accuracy averaged over the evaluated attacks. This table does not provide uncertainty for these cells.

Observed in the source: The design supplies both single-component controls rather than comparing only the full system against an unmodified baseline.

Prepare: Include the absent/absent control and both single changes. Hold the evaluation protocol constant and define the interaction claim before interpreting it.

Do not infer: The combined gain does not, by itself, prove the proposed robust-margin mechanism or certify robustness to all attacks.

Resource table: do not hide the cost of a speed gain

Mind Your Margin and Boundary: Are Your Distilled Datasets Truly Robust? · Table 4 and adjacent LS-PGD discussion, physical p. 8

Inner attacker	CIFAR-10 robust accuracy (%)	CIFAR-100 robust accuracy (%)	Relative speed
Multi-step PGD	28.63	12.54	1×
LS-PGD	28.41	12.3	1.4×

Source-reported numerical subset. Selected columns from Table 4 at images per class = 10; source accuracy is averaged over attacks. The source reports relative speed, not absolute elapsed time or uncertainty in this table.

Observed in the source: The displayed faster configuration has slightly lower accuracies. The nearby phrase suggesting slightly better accuracy should not override the numbers.

Prepare: Describe the measured quality–speed exchange and preserve the slower baseline's advantage. Add timing protocol and uncertainty when available.

Do not infer: A writing inconsistency does not establish that the scientific method is invalid; it shows why captions and conclusions need a numerical audit.

Keep the sweep, including collapse

Do We Need Adam? Surprisingly Strong and Sparse Reinforcement Learning with SGD in LLMs · Appendix F, Table 7, physical p. 18

Optimizer	Learning rate	Final reward
AdamW	1e-7	0.192
AdamW	1e-6	0.636
AdamW	1e-5	0.638
AdamW	1e-4	0
SGD	1e-3	0.111
SGD	1e-2	0.515
SGD	1e-1	0.635
SGD	1.0	0.644
SGD	10.0	0

Source-reported numerical subset. Final reward (source-reported scalar). Qwen3-8B optimizer learning-rate sweep; table has no uncertainty intervals.

Observed in the source: The sweep makes optimizer-specific useful scales and catastrophic settings visible; a best-number-only row would hide both.

Prepare: The sweep makes optimizer-specific useful scales and catastrophic settings visible; a best-number-only row would hide both.

Do not infer: Single source table and fixed reported setting, not optimizer equivalence across tasks. Do not share nominal learning-rate positions as if scales were matched; do not add error bars.

A mechanism control beside task quality

MuonSSM: Orthogonalizing State Space Models for Sequence Modeling · Appendix C.1, Table 12, physical pp. 21-22

Update operator	Validation accuracy (%)	Effective rank
Momentum only	72.04 ± 0.58	12.98 ± 0.81
Momentum + Frobenius normalization	72.53 ± 0.82	13.34 ± 0.43
Momentum + Newton-Schulz	74.67 ± 0.46	16.62 ± 0.57

Source-reported numerical subset. Validation accuracy (%) and effective rank. MMAct; matched backbone and training setup for three momentum-operator choices. Effective rank is (sum singular values)^2 / sum squared singular values.

Observed in the source: A controlled operator comparison lets a proposed representation mechanism and downstream quality be inspected together.

Prepare: A controlled operator comparison lets a proposed representation mechanism and downstream quality be inspected together.

Do not infer: The table and its local C.1 description report ± without defining its uncertainty semantics; preserve that ambiguity rather than inventing standard deviations or confidence intervals. Three aligned results alone are not a general causal law.

Hold outputs fixed; vary the evaluator

Position: Anthropomorphic Misalignment Research Needs Stronger Evidence · Appendix C.4, Table 1, physical p. 25

Aggregation	Inclusive (%)	Exclusive (%)	Delta (percentage points)
Argmax	42	26.87	15.13
Weighted	32.71	32.71	0

Source-reported numerical subset. Samples flagged as misaligned (%); delta in percentage points. Legal dataset, GPT-4o-mini judge, identical model generations. Inclusive requires coherence >50 and alignment ≤30; exclusive requires coherence >50 and alignment <30.

Observed in the source: A tiny classification-boundary choice can alter a headline rate; a factorial evaluator control makes that dependence visible.

Prepare: A tiny classification-boundary choice can alter a headline rate; a factorial evaluator control makes that dependence visible.

Do not infer: Two rows of a larger table, not a prevalence estimate of intent or an absolute safety metric. No uncertainty bars reported. Weighted aggregation reduces this boundary sensitivity in these rows, not all evaluator dependence.

Make privileged input part of the result

Towards Sub-Second Molecular Docking as a Structural Primitive: A Quantized Consistency Diffusion Framework · Appendix Table 4, physical p. 30

Information level	Prior description	Success count	Success rate (%)
Blind docking	Zero spatial prior	225 / 428	52.6
Noisy pocket	Center ±6 Å	253 / 428	59.1
Exact center	Exact pocket center	316 / 428	73.8
Interface-informed	Center + residue-level interface cue	353 / 428	82.48

Source-reported numerical subset. Docking success: ligand RMSD <2.0 Å. 428 test cases evaluated across four input-information regimes.

Observed in the source: An information-prior ladder makes differing task difficulty visible before comparing methods or speed claims.

Prepare: An information-prior ladder makes differing task difficulty visible before comparing methods or speed claims.

Do not infer: Preserves the table's reported precision. Do not attach the fastest timing to all accuracy rows: the paper's p. 9 GPU-region timing uses its own operating point and excludes preprocessing, confidence evaluation and compilation; this is not full end-to-end user latency.