AIM-CoT: Turning Multimodal Reasoning from Attention-Map Watching into Active Information Foraging

Core positioning: AIM-CoT is a training-free framework for interleaved multimodal chain-of-thought. It selects image regions by predictive entropy reduction rather than equating high attention with valuable evidence, uses changes in text-to-visual attention to time insertions, and first repairs the text–vision granularity mismatch with query-conditioned captions.

Core paper：Xiping Li, Jianghong Ma. AIM‑CoT: Active Information-driven Multimodal Chain-of-Thought for Vision-Language Reasoning. ACL 2026 Main, Long Paper, pp.26656–26681. Tables were checked against the official ACL PDF and arXiv v4; retrieval and link verification date: 2026-10-06.

0. Link verification table

Note: no standalone AIM-CoT entry was found on Hugging Face Daily Papers. Its Papers with Code path currently redirects to Hugging Face trending rather than a valid paper card. Semantic Scholar can be located through arXiv's References & & Citations links, but direct API requests did not return reliably during this check. These aggregators are therefore not treated as primary factual sources.

1. Candidate pool, deduplication, and random sampling

The run date was a Tuesday, assigning the domain toVision and Multimodality. First checked ml-report-index.mdunder the selection rule. Papers deeply reviewed in the preceding 14 days—Catch Your Emotion, Linear Mechanisms for Spatiotemporal Reasoning, BridgeVLA, PointMapPolicy, and closely overlapping topics—were excluded. Candidates were cross-searched through ACL Anthology, arXiv, PMLR/top-conference proceedings, OpenReview, Hugging Face Daily Papers, Papers with Code, and Semantic Scholar. Official venues and numerical results rely only on papers and conference pages.

The candidate pool contained 25 papers:

1. LV‑XAttn（ICML 2025 / PMLR）
2. Does RLVR Extend Reasoning Boundaries? Investigating Capability Expansion in Vision‑Language Models（ACL 2026）
3. Visual Graph Arena（ICML 2025 / PMLR）
4. AIM‑CoT: Active Information‑driven Multimodal Chain‑of‑Thought for Vision‑Language Reasoning（ACL 2026 Main）
5. Parrot: Multilingual Visual Instruction Tuning（ICML 2025 / PMLR）
6. Vision‑Language Model Selection and Reuse for Downstream Adaptation（ICML 2025 / PMLR）
7. VL‑Calibration（ACL 2026）
8. Do Vision‑Language Models Really Understand Visual Language?（ICML 2025 / PMLR）
9. Forest Before Trees（ACL 2026）
10. From Local Details to Global Context（ICML 2025 / PMLR）
11. Bring Reason to Vision（ICML 2025 / PMLR）
12. Textural or Textual（ICML 2025 / PMLR）
13. Re‑ranking Reasoning Context with Tree Search（ICML 2025 / PMLR）
14. Benchmarking Deflection and Hallucination（ACL 2026）
15. Interleaved Latent Visual Reasoning with Selective Perceptual Modeling（ACL 2026）
16. Kernel‑based Unsupervised Embedding Alignment（ICML 2025 / PMLR）
17. Explanatory Instructions（ICML 2025 / PMLR）
18. Look Less, Reason More（ACL 2026）
19. Improving Zero‑Shot Adversarial Robustness by Path Simplices（ICML 2025 / PMLR）
20. Beyond the Panorama（ACL 2026）
21. The Hidden Life of Tokens（ICML 2025 / PMLR）
22. When Seeing Overrides Knowing（ACL 2026）
23. Learn from Downstream and Be Yourself（ICML 2025 / PMLR）
24. Retrieval‑Augmented Perception（ICML 2025 / PMLR）
25. R2‑T2（ICML 2025 / PMLR）

The randomization command was shuf -i 1-25 -n 1returned 4. Candidate 4 had an official ACL long paper, complete main tables, component ablations, threshold sensitivity, significance tests, and appendix robustness experiments. With sufficient material for a deep reading,no redraw occurred (0 times)。

2. Motivation: The Problem Is Looking at the Wrong Evidence, Too Late, or Too Often

Interleaved multimodal chain-of-thought (I-MCoT) inserts local image regions during autoregressive reasoning so textual reasoning can revisit evidence. Success depends on two decisions:What to see—which image region is most useful;When to see—when to insert it without interrupting reasoning. ICoT chooses the highest cross-attention patch and inserts at every generated newline. AIM-CoT argues that both rules are overly passive.

The paper presents three sets of quantitative evidence.

• High attention is not sufficient: masking ICoT's Top-K_mask high-attention regions across four backbones, atK_mask=50 the average drops on M3CoT and ScienceQA are only 0.73% with 0.55%(Table 8, PDF p.14). If the highest-attention regions were essential evidence, masking 50 regions should not have such a small effect.
• High attention is not necessary either: across 1,000 examples from M3CoT and ScienceQA, more than 75% of examples have IoU below 0.5 between the highest-attention region and the actual critical region (§3.1, PDF pp.2–3). Attention peaks often fall on high-contrast backgrounds, label text, or generally salient objects rather than the entities needed to answer.
• Short text cannot constrain dense visual information: M3CoT queries average fewer than 15 tokens despite complex images. Expanding them into query-conditioned captions of 48/96/144 tokens raises attention-to-target Mean IoU from 4.5% to 12.2% / 18.6% / 24.1%, an absolute gain of 7.7 / 14.1 / 19.6 pp, all with p<0.001(Table 9, PDF p.15). This supports insufficient semantic anchors, rather than simply insufficient model capacity, as the diagnosis.

The paper thus frames reasoning asactive information foraging: build a searchable semantic map with query-conditioned captions, probe candidate regions when visual evidence is needed, and value a region by how much it reduces uncertainty in the next-token distribution.

3. Formalization: The CAG → AVP → DAT Loop

3.1 CAG: Repairing the text–vision granularity gap

Given image (I), question (x), and caption prompt (P_{CAG}), the VLM generates a query-conditioned caption strictly grounded in the image:

[ D_{CAG}=\operatorname{VLM}(I,x,P_{CAG}),\qquad x'=\operatorname{concat}(x,D_{CAG}). ]

Cross-attention is then computed from augmented textual states (H_T) and visual states (H_V):

[ A'=\operatorname{softmax}\left(\frac{(H_TW^Q)(H_VW^K)^\top}{\sqrt{d_K}}\right). ]

The caption is not the final rationale; it supplies semantic anchors for attention-based candidates. CAG also includes a negative constraint against inventing invisible details, reducing upstream caption hallucinations (Equations 1–2 and 6; Appendix A.3, PDF pp.12–13).

3.2 AVP: Selecting regions by entropy reduction instead of attention magnitude

The image is partitioned into an (s_g\times s_g) grid, with regions formed by contiguous (s_r\times s_r) blocks. The candidate set combines two routes:

[ C=C_{attn}\cup C_{exp}, ]

Here (C_{attn}) contains the Top-(N) regions from the augmented attention map, while (C_{exp}) contains (M) candidates sampled uniformly from the remainder to break attention tunnel vision. Base uncertainty about the next token (Y) and conditional uncertainty after inserting candidate (R_i) are:

[ U_B=H(Y\mid I,x,y_{<t}), ]

[ U_{C,i}=H(Y\mid I,x,y_{<t},R_i), ]

[ IG(R_i)=U_B-U_{C,i}. ]

At each round, AVP chooses the region with maximum information gain, inserts it into context, and recomputes gains for the next round. The second region is therefore evaluated conditional on having already seen the first, avoiding redundant one-shot Top-K selection.

Input: candidate set C, number of selections K
S ← ∅
Repeat K times:
  Compute U_B = H(Y | current context, S)
  For each R_i ∈ C \ S:
    Compute U_C,i = H(Y | current context, S ∪ {R_i})
    IG_i ← U_B − U_C,i
  S ← S ∪ {argmax IG_i}
Return S

The paper does not rigorously prove that this information-gain function is necessarily submodular. Empirical tests find diminishing marginal returns in 61.09%–72.00%of M3CoT examples and 61.67%–68.33% of LLaVA-W examples, with a binomial test yielding p≤0.0462(Table 15, PDF pp.23–24). Thus, the claim that greedy selection approximates the global optimum is best understood asmotivation under empirically approximate submodularity, not a general theoretical guarantee.

3.3 DAT: Inserting images when visual demand changes

Let (A_{vision}(t)) be the current token's total attention to all visual tokens. Then:

[ \Delta A_{vision}(t)=A_{vision}(t)-A_{vision}(t-1). ]

The fixed-threshold variant triggers AVP only when (Delta A_{vision}(t)>\delta). Inserted regions come with instructions to treat them as supplementary references and cross-check them against textual context, reducing domination by incorrect visual evidence (Equation 5; §4.4, PDF pp.6–7).

The paper also offers a direction requiring no threshold tuning: over a sliding window of the previous (W=5) tokens, compute

[ Z_t=\frac{\Delta A_{vision}(t)-\mu_{t-W:t-1}}{\sigma_{t-W:t-1}+\epsilon}, ]

Trigger when (Z_t>\lambda), with default (lambda=3) (Equation 11; Appendix G, PDF pp.18–20).

3.4 Key mechanisms and engineering details

4. Datasets, settings, and metrics

The four backbones span early fusion with fixed resolution—Chameleon-7B and Janus-Pro-7B—and late fusion/dynamic resolution—Qwen2-VL-7B and Qwen2.5-VL-32B. Decoding uses temperature 0.7, top-p 0.9, and repetition penalty 1.2; at least 32 new tokens, up to 512 on M3CoT and 1024 elsewhere;do_sample=True(Table 7). Variation across repeated samples should therefore be reported, but the main tables lack means ± standard deviations.

5. Complete main-results matrix

The paper reports only 3 benchmarks, so five cannot be invented. The following fully transcribes 7 methods × 6 settings for four backbones. M3CoT/ScienceQA use Accuracy; LLaVA-W uses ROUGE-L. Bold indicates the column best. All values come from Table 1 (PDF p.8; ACL pp.26662–26663).

5.1 Chameleon‑7B

5.2 Janus‑Pro‑7B

5.3 Qwen2‑VL‑7B

5.4 Qwen2.5‑VL‑32B

The simple mean of the six column-wise relative improvements in Table 1, calculated for this report, is: Chameleon-7B 7.88%、Janus‑Pro‑7B 6.02%、Qwen2‑VL‑7B 2.63%、Qwen2.5‑VL‑32B 4.78%. Fixed-resolution early-fusion models benefit most, supporting the authors' perceptual-repair interpretation. Yet the 32B model also outperforms its 7B relative in gains, suggesting stronger models better exploit inserted evidence fragments; benefits cannot simply be attributed to weaker backbones needing more help.

6. Layer-by-Layer Ablations: Which Component Contributes Most?

6.1 Component ablations

AVP removal causes the largest drop in 8 of 9 expanded settings, making it the core component. CAG has the largest effect on M3CoT/Chameleon, consistent with the joint granularity bottleneck of short queries and a fixed-resolution backbone. Starting from BM=ICoT, adding CAG / AVP alone on Chameleon 0-shot gives M3CoT only +0.3 / +0.4, whereas combining them gives +1.0; ScienceQA gives +0.5 / +0.9, combined +1.7; LLaVA-W gives +0.6 / +1.2, combined +2.1(Table 3). CAG thus appears primarily to improve the candidate distribution searchable by AVP rather than directly improving answers.

6.2 Candidate pools and noise rejection

• Full C_attn ∪ C_rand On M3CoT/LLaVA-W, scores are 31.4/29.8; attention-only candidates fall to 30.8/28.9; random-only candidates score 30.4/28.6; FastSAM alone is worst at 29.9/27.7(Table 12, PDF p.22). Simple random exploration beats more elaborate proposals, suggesting candidate diversity matters more than object-like appearance.
• Across three repetitions, the proportion of salient regions selected by AVP from exploration candidates is 17.25%/20.44%/27.27%on M3CoT and 31.33%/25.77%/26.67%on LLaVA-W (Table 13). The exploration branch is not decorative redundancy.
• When noise patches are forced to have the highest attention, Top-K has a Noise Rejection Rate of 0%, and Chameleon's three benchmark scores drop by 4.2/3.8/5.1; AVP achieves an NRR of 96.4%/98.1%/95.5%, with drops of only 0.2/0.1/0.4. Qwen2-VL-7B likewise reaches 97.8%/99.2%/96.7%, with drops no greater than 0.2(Table 14, PDF p.23). This is the strongest causal-style support for using attention to propose candidates and information gain to decide.

6.3 Semantic hits and proxies for human intuition

Averaged over four backbones and 500 examples each from M3CoT and ScienceQA, Semantic Hit Rate uses SAM-mask IoU>0.5 as its criterion. ICoT Top-K achieves 18.7%/24.3%, while AIM-CoT achieves 65.2%/71.8%, an absolute gain of 46.5/47.5 pp(Table 4, PDF p.9). In GPT-4v's blind paired evaluation of 500 examples, AIM-CoT versus ICoT has Win/Tie/Loss rates of, on M3CoT, 76.4/16.2/7.4%，LLaVA‑W 81.2/13.6/5.2%(Table 6). Both support evidence resembling complete entities, but SAM and GPT-4v are proxy judges rather than real-user annotations.

6.4 Safety mechanisms

Without CAG's negative constraints, Chameleon's caption hallucination rates on M3CoT/ScienceQA/LLaVA-W are 5.8/4.5/8.4%; the full method reduces them to 1.9/1.2/2.7%, with corresponding performance rising from 30.1/52.5/27.9 to 31.4/53.1/29.8. Qwen2-VL-7B also moves from 3.5/2.8/5.1% to 0.8/0.5/1.5%(Table 18, PDF p.26).

In the extreme stress test that randomly inserts patches from other examples, Chameleon's three benchmark scores drop without safety prompts by 6.9/6.9/7.7, versus only 1.5/1.3/1.4with them; Qwen2-VL-7B goes from 6.5/6.5/5.8 to 1.2/1.3/1.2(Table 19, PDF p.26). Prompts can reduce noise domination, but this does not establish adversarial robustness or formal safety.

7. Hyperparameter Sensitivity, Phase Transitions, and Statistical Significance

7.1 Phase transitions in δ

On Chameleon-7B, LLaVA-W 0-shot, the paper sweeps (delta∈{0.1,0.125,0.15,0.175,0.2,0.225}). Performance first rises and then falls with the threshold, reaching ROUGE-L at δ=0.2 of 0.2983. Lower thresholds trigger more often, with excessive images interrupting linguistic reasoning; higher thresholds delay necessary evidence (Figure 5, PDF p.18). Other points lack exact values and error bars, so the full curve cannot be reliably reconstructed or a sharp transition asserted; only a unimodal trend is supported.

The gap between the optimal fixed threshold and adaptive Z-score is small. Across three backbones, M3CoT gives 31.4→31.1、39.7→39.5、58.7→58.6; ScienceQA gives 53.1→52.9、56.9→56.7、76.8→76.8; LLaVA-W gives 29.8→29.5、35.5→35.4、46.5→46.3(Table 11, PDF p.20). The deployable variant thus loses only 0–0.3 points, though the table omits Qwen2-VL-7B and reports no sweep over λ or W.

7.2 Significance

Table 10 compares each backbone/setting against the runner-up with one-sided tests: McNemar for M3CoT and ScienceQA, Wilcoxon signed-rank for LLaVA-W. All 24 comparisons have p<0.05, including 16 with p<0.01. Borderline cases include Chameleon M3CoT 1-shot at p=0.0228、ScienceQA 1-shot p=0.0246，Qwen2‑VL‑7B LLaVA‑W 1-shot p=0.03024; the strongest include several at p=0.0000 or <10^-5(Table 10, PDF p.17).

However,the paper reports no multiple-comparison correction, effect-size confidence intervals, independent seed count, or main-table means ± standard deviations. Because do_sample=True, a single significance test cannot replace variance across randomized runs. The paper also reports a Pearson correlation between DAT's synchronized insertion ratio and LLaVA-W ROUGE-L of r=0.2166, p=0.048, a weak association barely passing 0.05. Synchronization is 0.8889in the high-score group and 0.5000in the low-score group, with t-test p=0.0019(Appendix F, PDF pp.17–18). The correlation supports timing being related to performance, not attention changes causally improving answers.

8. Related-Work Lineage: Origins of the Problem and Transmission of the Design

8.1 Multimodal‑CoT（TMLR 2024）

Problem addressed: fuses textual and image features in two stages, generating a rationale before answering from it, showing that models below 1B can benefit on ScienceQA.Remaining gap: visual features are fused at input, but reasoning remains primarily textual and cannot precisely revisit local evidence mid-chain.AIM-CoT's response: insert local visual tokens directly into the autoregressive context.Design transfer: retains the idea of an explanatory intermediate state, but CAG only anchors attention while AVP dynamically selects the final evidence.

8.2 DDCoT（NeurIPS 2023）

Problem addressed: separates reasoning and recognition responsibilities, uses negative-space prompting to diagnose missing visual information, then reasons jointly.Remaining gap: depends on external recognition models and task decomposition; visual-tool timing remains a workflow rule.AIM-CoT's response: reads internal VLM states directly as trigger signals.Design transfer: DDCoT's gap diagnosis and division of responsibilities evolve into DAT diagnosing visual demand and AVP supplying evidence.

8.3 SCAFFOLD（COLING 2025）

Problem addressed: overlays a dot grid on the image and uses coordinate references in text to reduce ambiguous spatial references.Remaining gap: the static grid exposes all regions equally, without utility ranking or dynamic insertion timing.AIM-CoT's response: likewise recognizes that text–vision alignment needs an intermediary, but inserts only the most valuable regions when needed.Design transfer: SCAFFOLD's explicit spatial anchors are absorbed into CAG's semantic anchors and AVP's local-region representations.

8.4 ICoT（CVPR 2025）

Problem addressed: first interleaves paired image patches and textual rationales; ADS selects regions by attention Top-K and triggers on newlines, enabling training-free I-MCoT.Remaining gap: attention peaks are not task utility, and newlines indicate formatting rather than visual demand.AIM-CoT's response: attention still proposes candidates, but entropy reduction determines ranking; text-to-visual attention changes determine timing.Design transfer: ICoT supplies the overall paradigm and engineering interface; AIM-CoT replaces the decision core of what to select and when.

8.5 M3CoT（ACL 2024）

Problem addressed: uses 11,459 multistep examples across science, mathematics, and commonsense to expose limitations of single-step visual-reasoning evaluation.Remaining gap: short questions and complex images create a severe granularity gap; fluent high-level linguistic reasoning can coexist with failure to locate key visual evidence.AIM-CoT's response: CAG explicitly targets this structural bias, and M3CoT becomes the main diagnostic setting for attention misalignment.Design transfer: the benchmark defines the failure mode and directly shapes CAG's query-conditioned caption design.

8.6 Nonlinear technical evolution tree

Text CoT
├─ Input-level multimodal fusion → Multimodal-CoT
│  └─ Explanation first, but intermediate steps remain mainly textual
├─ Responsibility/tool decomposition → DDCoT
│  └─ Diagnose visual gaps, then invoke recognition
├─ Explicit spatial anchoring → SCAFFOLD
│  └─ Image coordinates ↔ textual references
└─ Visual insertion within reasoning → ICoT
   ├─ Attention Top-K patch selection
   ├─ Static newline triggers
   └─ AIM-CoT
      ├─ CAG: semantic anchors repair granularity gaps
      ├─ AVP: information gain + random exploration
      └─ DAT: attention changes + safe integration
         ├─ Adaptive Z-score avoids dataset-specific threshold tuning
         └─ Next: cost constraints, black-box proxies, video/3D

Three shifts drive this evolution: visual information moves from input features to intermediate reasoning evidence; attention becomes a cheap candidate generator rather than an explanation, while task uncertainty drives decisions; and fixed formatting triggers give way to internal-state triggers. AIM-CoT's main value is separating selection and timing into measurable, replaceable control problems.

8.7 Cross-paper comparison

9. Contributions, reusable designs, and transfer potential

1. Decision hierarchy matters more than any single scorer. AIM-CoT retains attention as a retriever and uses information gain for reranking. Table 14 shows 95.5%–99.2% NRR even when noise attention is maliciously inflated. This two-stage retrieval pattern can transfer to long-video frames, webpage elements, medical-image slices, or robot camera-view selection.
2. Express tool timing as state-change detection. DAT monitors cross-modal attention changes rather than newlines or periods. The pattern can transfer to RAG, triggering retrieval when internal state shifts from parametric knowledge toward external factual needs, or to agents' visual reinspection, code execution, and sensor sampling.
3. Exploration is a cheap, effective defense against blind spots. Random candidates outperform Selective Search/FastSAM in Table 12, and roughly 17%–31% of final salient regions come from exploration in Table 13. The reusable principle is to let interpretable heuristics handle most recall while reserving a small budget for unbiased exploration.
4. Bind safety prompts to evidence insertion. Table 19's random incorrect-patch stress test shows simple consistency instructions shrinking losses from 5.8–7.7 points to 1.2–1.5. Though not a formal guarantee, this is a low-cost default semantic safeguard for retrieval- or tool-augmented pipelines.
5. Adaptive thresholds transfer better than dataset-specific tuning. Z-score trails by only 0–0.3 points across 9 reported settings, suggesting a tiny sacrifice in peak performance can buy cross-task stability in deployment.

10. Author-Acknowledged Limitations and Independent Critique

10.1 Author-acknowledged limitations

The official ACL version hasno standalone Limitations section, so an explicit author-provided list cannot be invented. Three constraints are nevertheless acknowledged in the text/appendix:

1. Thresholds depend on task granularity: the authors acknowledge that the optimal fixed (delta) is task-dependent: 0.5 for M3CoT and 0.2 for ScienceQA/LLaVA-W, mitigated with Z-score (Appendix G, PDF pp.18–20). Fixed thresholds are therefore not plug-and-play across domains.
2. Performance depends on backbone architecture and capability: differences among Chameleon, Qwen2-VL, and Qwen2.5-VL are attributed to resolution mechanisms, caption quality, and long-context integration (Appendix K, PDF pp.24–25). AIM-CoT's gains are not independent of backbone capability.
3. Visual insertion itself introduces noise: the authors identify interference from inserted evidence as a potential challenge, motivating safety instructions and random incorrect-patch stress tests (§4.4 and Appendix L, PDF pp.7, 25–26). Reliability depends on selection and prompting working together.

10.2 Independent critique

1. [Efficiency/engineering] Each AVP trigger repeatedly evaluates next-token entropy for N+M candidates, and when K=3 it requires multiple conditional forward rounds. Yet the paper reports no wall-clock time, FLOPs, memory, token overhead, or average trigger count. Training-free does not mean cheap inference; a single decoding step could become dozens of candidate evaluations in a service.
2. [Reproducibility] Although the paper claims code availability, the anonymous endpoint returned HTTP 401 on 2026-10-06, and no deanonymized GitHub repository was found. Despite substantial method detail, end-to-end replication lacks key implementation choices, especially attention-layer selection and head aggregation across architectures, which materially affect DAT/AVP.
3. [Statistics] Decoding uses do_sample=True, temperature=0.7, but main tables report single values without seed counts or means ± standard deviations. Small advantages of 0.5–0.9 pp may be comparable to sampling variance. Sample-level McNemar/Wilcoxon tests do not replace uncertainty across runs.
4. [Multiple testing] Table 10 performs 24 one-sided comparisons without reporting Holm or Benjamini–Hochberg corrections. Under strict family-wise error control, findings near 0.05 such as p=0.03024 become more fragile. Most p-values are small, but claiming universal robustness is unwarranted.
5. [Causal interpretation] The correlation between DAT synchronization and ROUGE-L is only r=0.2166, p=0.048, a weak effect potentially confounded by answer length and question difficulty. Attention changes reflecting cognitive demand remain a plausible hypothesis rather than an identified causal mechanism; controlled randomization of trigger times is needed.
6. [Proxy evaluation] Semantic Hit Rate treats SAM IoU>0.5 as a proxy for semantic correctness, while GPT-4v preference uses a model to judge models. Tables 4/6 may favor complete objects and common visual semantics. For text, relationships, counting, or tiny local cues, complete objects may not be the best evidence.
7. [Accessibility] DAT and AVP require token-level internal attention and distribution entropy after candidate insertion, unavailable through closed multimodal APIs. Even with open weights, some efficient attention kernels do not return full maps. Plug-and-play use mainly applies to modifiable white-box research stacks.
8. [Benchmark extrapolation] The main experiments cover only 3 static-image VQA benchmarks, with LLaVA-W the sole open-ended task. They omit video, OCR documents, charts, medical imaging, and real interaction. Effectiveness for long temporal sequences, cross-frame evidence, or high-stakes domains is unestablished; Table 1's breadth comes from backbones and shot settings rather than task types.

11. Open Challenges as Executable Head-to-Head Experiments

Theory

• Is information gain truly submodular?: with the same backbone and candidate pool, compare greedy AVP, beam search, MCTS, and exhaustive optimum for small pools. Report optimality gaps, forward-pass counts, and submodularity violation rates, stratifying by Table 15's K_small=2..5 rather than reporting only a majority proportion.
• Causality of trigger signals: fix insertion counts and selected regions, randomize timing, and compare DAT, newline triggers, uniform intervals, maximum predictive entropy, and random triggers. Only superiority over all budget-matched controls isolates timing from insertion count.
• Alignment of proxy objectives: on the same candidate set, compare Δ next-token entropy, answer-token margin, final-answer log-probability, mutual-information estimates, and a trained value model. Measure Kendall τ between local proxy rankings and final accuracy to determine whether entropy reduction reliably represents task utility.

Engineering and applications

• Quality–latency Pareto frontier: on A100/H100 and consumer GPUs, fix batch size/resolution for No-CoT, ICoT, and AIM-CoT. Report time to first token, per-token latency, P50/P95/P99, peak memory, total forward passes, and energy. Cap AVP at 1/2/3 regions and plot Accuracy/ROUGE-L against latency.
• Removing white-box dependence: train a black-box trigger from observable signals such as output entropy, hidden-state drift, or missing visual entities in generated text, then compare against DAT under equal budgets. The goal is approximate deployment with closed APIs.
• Evidence utility for real users: ask humans to answer using only candidate crops, without the full image. Compare AVP, ICoT Top-K, and SAM proposals on accuracy, response time, and missing-evidence rate, replacing sole reliance on GPT-4v.

Emerging directions

• Video I-MCoT: compare frame-level DAT + within-frame AVP against joint spatiotemporal retrieval for long-video QA. Fix total visual tokens and stratify by cross-frame dependency length to test whether local entropy reduction overfavors nearby frames.
• Embodied/robotic active vision: expand candidates into executable viewpoint actions and compare crop-only AIM-CoT, camera motion, and mixed actions on the same task. Report success, action count, and safety violations to test transfer from viewing patches to actively observing the world.
• Retrieval-augmented multimodal reasoning: include external image–text collections as well as the current image. Compare attention recall + IG ranking, vector retrieval + reranking, and end-to-end retrieval policies under equal retrieval-call counts, measuring factual accuracy and citation completeness.

12. Recent-Work Radar

1. Look Light, Think Heavy: What Multimodal Chain‑of‑Thought Reasoning Can and Cannot Do(ACL 2026): tests MCoT's capability boundaries, a useful counterpoint to AIM-CoT that cautions against equating more or longer visual steps with deeper reasoning.
2. Let’s Think with Images Efficiently! / DaP‑ICoT(2026): also pursues dynamic, precise visual thinking and reports 72.6% lower visual-token consumption. It should be compared with AIM-CoT under a shared budget rather than separately against old ICoT.
3. Look Less, Reason More(ACL 2026): studies the synergy between fewer visual operations and better reasoning. At 73.4% on HR-Bench 4K, tool use is only 20.1%, 66.5% below prior methods; this directly relates to AVP's selective observation.
4. Interleaved Latent Visual Reasoning with Selective Perceptual Modeling(ACL 2026): brings selective visual perception into latent reasoning, potentially reducing repeated encoding of dense images; interpretability and causal verification remain challenges.
5. Retrieval‑Augmented Perception(ICML 2025): dynamically retrieves high-resolution crops through visual RAG, reporting 43%/19% gains for LLaVA-v1.5-13B on V*-Bench/HR-Bench. It complements AIM-CoT's candidate space but raises retrieval-noise and latency concerns.
6. R2‑T2(ICML 2025): locally reoptimizes multimodal MoE routing at test time without changing backbone parameters. As a compute-matched baseline, it can test whether extra routing computation or visual foraging is more effective.
7. Improve Vision Language Model Chain‑of‑Thought Reasoning(ACL 2025): expands rationales with GPT-4o, then uses short answers for outcome rewards/DPO. This training-based approach provides a quality–cost comparison with training-free AIM-CoT.

13. Final Assessment

AIM-CoT is valuable because it turns what to look at and when to look—two vague engineering questions in interleaved multimodal reasoning—into measurable region utility and state changes, supported by comprehensive ablations, noise stress tests, and significance tests. The key result is Chameleon-7B/LLaVA-W 0-shot improving from ICoT's 25.2 to 29.8 (+4.6 points, +18.25%)while still rejecting 95.5% of noise regions when their attention is artificially inflated.

However, its mechanism validation is stronger than its systems evaluation: service overhead, cross-seed variance, and multiple-testing corrections are absent, and the code endpoint is currently unavailable. If future work establishes AVP's gains under equal latency budgets and replaces white-box attention dependence with observable triggers, the candidate-retrieval → information-gain reranking → state-change triggering design may transfer more broadly than AIM-CoT itself.

Data-source statement: all experimental numbers come from Tables 1–19, Figure 5, and Appendices A–L of the official ACL version. Relative gains were recalculated as (AIM-CoT − runner-up) / runner-up , and absolute gains as score differences. Unreported standard deviations, seeds, other sensitivity-curve points, latency, and FLOPs were not fabricated.

Index status: the entry “2026-10-06 | Vision and multimodality | AIM-CoT | ACL 2026 Main | Selects visual evidence through information gain and triggers insertion through attention changes; wins in 24/24 settings, but code availability and inference overhead remain replication gaps” has been appended to ml-report-index.md。