{"id":"b360ef46-4aa0-4911-b85e-62e3c0461104","arxiv_id":"2607.21125","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A multi-agent image-restoration framework with a self-updating causal memory graph reports 0.1–0.4 dB PSNR gains over prior restoration agents and lower rollback rates.","lead":"An image-restoration agent is paired with a 'causal memory graph' that records which tools improve which image problems, and the memory is updated as the agent works. On standard and mixed-degradation benchmarks, the system reports small but consistent gains over prior restoration agents and better user-preference alignment.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Self-evolving memory is updated during evaluation (Eq. 21) with no stated reset/frozen protocol, so the reported gains over IAMAgent may reflect test-time adaptation or test-set leakage rather than transferable causal memory.","rationale":"The reader's weakest assumption is exactly the load-bearing concern: memory evolution during evaluation is not bounded by a stated protocol, and the paper's 'zero-shot' claim is not checkable without knowing whether memory is frozen, reset per image, or persistent across the test set. My read of the paper confirms this: Section III-E explicitly updates memory after every action (Eq. 21), and Section IV-E labels real-world results as zero-shot without specifying memory state. Since the reported gains over IAMAgent are modest (0.31 dB all-in-one average; 0.13–0.18 dB on real-world benchmarks), the missing protocol is not a minor detail — it determines whether the causal memory contributes transferable knowledge or merely serves as an online adaptation buffer. I do not think the result is false; the paper's design is coherent and the shared tool library / frozen tool parameters are good methodological choices. But the evidence as written cannot settle the attribution. This supports the reader's CONDITIONAL verdict rather than a rejection. A concrete frozen-memory or per-image-reset rerun would resolve the concern, so no change to the reader's verdict is needed.","tokens_in":20403,"tokens_out":3601,"duration_ms":35186,"concrete_test":"Re-run Table III and Table IX under two protocols: (1) frozen memory — initialize G once from training experience only, then disable all updates (force π_m to 'ignore' or freeze Eq. 21) during evaluation; (2) per-image reset — reset G to empty at the start of each test image, allowing only within-trajectory updates. If either protocol causes Causal-AgentIR's average PSNR to drop from 35.55 dB to within 0.1 dB of IAMAgent's 35.24 dB (or real-world metrics to regress toward MAIR/HybridAgent), the headline gains are attributable to test-time memory evolution or order leakage, not to transferable causal knowledge. The authors should also state explicitly whether G is initialized from training splits of the same benchmark datasets and whether results depend on test-set order.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the self-evolving causal memory graph produces transferable restoration gains (Table III: 35.55 dB vs IAMAgent 35.24 dB; Table XII: +0.40 dB from memory). For this claim to hold, memory evolution must not be allowed to fit the test set. Section III-E updates memory after every executed action via Eq. (21), with operations add/update/merge/reinforce/ignore/discard. Section IV-E calls the real-world benchmark results 'zero-shot,' but the paper never states whether G is reset per image, frozen after a training phase, or carried across the benchmark; nor whether G is initialized from prior experience on the same degradation types used in evaluation. If G persists across test images, later images inherit edges learned from earlier test images — an evaluation-order artifact. If G is reset per image, the method is performing per-image online adaptation that the baselines (MAIR, IAMAgent) do not, so the comparison does not isolate transferable memory. The reported 0.1–0.4 dB gaps are plausible within the range such adaptation could explain. No code, data, hyperparameters, or memory-reset protocol is provided, so the text cannot currently distinguish causal-memory transfer from test-time fitting. This is an omitted-protocol issue, not an internal inconsistency, and is addressable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Causal-AgentIR, a hierarchical multi-agent framework for image restoration in which restoration experience is organized as a self-evolving causal memory graph. Nodes represent degradations, regions, tools, actions, quality changes, costs, and user preferences; edges carry contextual conditions, causal-effect weights, and confidence scores. A memory-curator agent adds, updates, merges, reinforces, ignores, or discards edges after each executed action, while a planner, degradation agent, tool agents, causal-memory agent, and critic collaborate to select restoration actions. Experiments cover task-specific, all-in-one, real-world, and mixed-degradation settings. The paper reports best or near-best PSNR/SSIM in many benchmarks and attributes the improvement to multi-agent collaboration (+0.26 dB) and self-evolving causal memory (+0.40 dB) in Table XII.","tokens_in":20951,"tokens_out":4429,"duration_ms":44772,"significance":"If the reported results are reproducible and the memory protocol is valid, the structured causal memory graph is a meaningful step beyond static tool descriptions and unstructured textual summaries in agentic image restoration. The explicit modeling of confidence, effect magnitude, and selective forgetting is a promising design. The paper also provides extensive benchmark comparisons, ablations, and a user study. However, the manuscript currently omits the training procedure for the learnable memory policy, the values of more than a dozen hyperparameters, and any statement about whether the memory graph is reset or frozen during evaluation. These omissions are load-bearing for the central claim that the gains come from transferable causal memory rather than test-time adaptation or reward self-consistency.","major_comments":[{"comment":"The real-world results are called 'zero-shot' in §IV-E, but the paper never states whether the memory graph G is reset per image, frozen after a training phase, or carried across test images. Since §III-E updates memory after every executed action, a persistent graph lets later test images benefit from edges learned on earlier test images, while a per-image reset turns the method into online adaptation that the baselines do not perform. Either way, the +0.40 dB memory gain in Table XII does not isolate transferable causal knowledge unless the protocol is specified and a no-adaptation variant is reported.","section":"§IV-E / §III-E, Eq. (21)"},{"comment":"The memory reward R_mem_t = U(Π_{t+1}|G_{t+1}) − U(Π_{t+1}|G_t) is computed from the same utility U that uses edge weights w and confidences γ that the curator is updating. The update rules in Eqs. (24)–(25) push w and γ toward the currently observed ΔQ_t, so R_mem_t can be positive by construction even if future restoration quality is unchanged. This reward therefore measures self-consistency of the memory update, not an independent improvement in future planning. Please redefine R_mem using held-out planner outcomes or a separate validation reward.","section":"§III-E, Eq. (31)"},{"comment":"The memory-operation policy π_m is called 'learnable,' but the manuscript provides no training objective, dataset, optimization procedure, or loss function for it; the equations describe only the operation taxonomy and reward. Furthermore, the values of α, ρ, β1–β5, τ_Q, τ_δ, τ_γ, τ_w, λ_s, λ_γ, λ_w, λ_C, γ0, and η1 are not reported. Without these details, the experiments are not reproducible, and it is unclear whether the policy is actually trained or is a hand-crafted rule. Please provide the full training protocol and all hyperparameter values.","section":"§III-E, Eqs. (21)–(31)"},{"comment":"The retrieval relevance in Eq. (14) depends on a similarity function Sim(ψ_t, z_e) that is never defined, and on edge confidence γ_e and effect |w_e| without specifying how these are normalized. Similarly, Eq. (16) aggregates over causal paths Ω(Π_t, g_t) without defining how paths are enumerated, their maximum length, or how the product-of-γ and sum-of-w in Eq. (17) are normalized across plans with different path counts. These details are necessary to determine whether the planner genuinely uses multi-hop causal reasoning or effectively uses a single-edge heuristic.","section":"§III-D, Eqs. (14) and (16)"}],"minor_comments":[{"comment":"The section heading has a stray space: 'F . Agent Efficiency...' Please fix the formatting.","section":"§IV-F"},{"comment":"The abbreviations R+S+H+B+N and H+R+N are used without being spelled out. Please define them at first use.","section":"§IV-A"},{"comment":"The components of ΔQ_t (fidelity, perceptual, objective, artifact, cost, preference) have different units and scales, but no normalization protocol is given. Without this, the weights β1–β5 and the threshold τ_Q are not meaningful.","section":"§III-E, Eq. (19)"},{"comment":"Figure 6 refers to 'ChatGPT 4o' and 'the all-in-one method' without specifying which all-in-one method was used. The user study is described as small-scale; please report the number of comparisons per condition and any variability/confidence measure.","section":"§IV-I, Fig. 6"},{"comment":"Releasing code, trained memory policies, and the exact memory-initialization/reset script would substantially increase the credibility of the empirical claims.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the experiments are extensive, but the missing training details, unspecified hyperparameters, and the circular memory reward are substantial issues. I would be willing to review a revised version that addresses these points, especially the memory-reset protocol and a non-circular definition of the memory reward."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the causal memory graph is a real idea and the paper is coherent, but the current evidence doesn't show transferable gains because the memory-update protocol during evaluation is unspecified and the memory-policy reward is circular.\n\nWhat's new: the structured memory graph with typed nodes (degradation, region, tool, action, quality, cost, preference) and edges with confidence/effect weights, plus the add/update/merge/reinforce/ignore/discard operations, is not in MAIR or AgenticIR. The hierarchical multi-agent setup with a dedicated memory curator and causal-memory agent is a useful organizational contribution. Benchmarks are consistent, and the Table XII ablation gives memory +0.40 dB and multi-agent +0.26 dB, which is plausible.\n\nWhere it's soft: the biggest problem is that Section III-E updates memory after every executed action (Eq. 21) and Section IV-E calls real-world results \"zero-shot,\" but the paper never says whether G is reset per image, frozen after a training phase, or seeded from the same test distributions. If G carries across test images, later images inherit edges learned from earlier test images—an evaluation-order artifact. If reset per image, the method is doing per-image online adaptation that the baselines don't, so the comparison doesn't isolate transferable memory. That's an omitted protocol, not a proven falsehood, but it's load-bearing.\n\nSecond, the memory-policy reward R_mem (Eq. 31) is the difference between two values of U computed from the very edge weights and confidences the curator is updating. The policy is rewarded for increasing its own belief, not for demonstrated improvement on future plans. That's circular.\n\nThird, the \"learnable\" memory policy has no training procedure, no hyperparameter values, no code or data. The paper lists a dozen parameters (α, ρ, β's, τ's, λ's) without values, and Eq. 19 mixes fidelity, perceptual, cost, and preference signals without saying how they're normalized. So the empirical core is not independently checkable from the text.\n\nThe \"causal\" language is also stronger than what's supported—these are observed correlations from executed trajectories, not interventional estimates.\n\nWho it's for: people building LLM-based restoration agents and memory-augmented planner systems. They'll get a clear framework and a good comparison table, but they should treat the quantitative gains as provisional.\n\nRecommendation: send it to peer review. The idea is novel, the paper is coherent, and the missing protocol details are fixable in a revision. A serious referee should ask for the memory reset/initialization specification, a non-circular reward or separate validation of memory quality, and code or detailed hyperparameters.","headline":"Novel memory-graph idea for restoration agents, but the empirical claims are conditional on a missing evaluation protocol and a circular memory reward.","tokens_in":21388,"tokens_out":3143,"would_cite":false,"duration_ms":29966,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A self-evolving causal memory graph lets image-restoration agents learn from every action and reach 35.55 dB average PSNR on combined rain, snow, haze, blur, and noise.","keywords":["image restoration","multi-agent system","causal memory graph","self-evolving memory","degradation-aware planning","tool selection","execution ordering","user preference"],"falsifier":"Run the all-in-one benchmark twice with identical agents and tool library—once with the memory graph frozen at its initial state and once with it evolving during evaluation. If the evolving version keeps its reported ~0.40 dB advantage, self-evolution is doing real work; if the advantage disappears, the gains come from adapting to test images.","tokens_in":20342,"feed_emoji":"🧠","tokens_out":7259,"duration_ms":67112,"temperature":0.7,"pith_summary":"The paper proposes Causal-AgentIR, a hierarchical multi-agent system for image restoration that records every executed restoration step—degradation, region, tool, action, quality change, cost, user preference—in a structured causal memory graph and updates that graph after each action. Its central claim is that this self-evolving memory, rather than any new restoration network, is what lets the system choose better tools and better execution orders over time. A sympathetic reader would take the paper as establishing that image restoration can be framed as a lifelong learning problem: with the same frozen tool library, the agent improves because it remembers which operations help or harm under which conditions. The reported results put the all-in-one average at 35.55 dB PSNR, ahead of every compared agent and model, with the ablation attributing 0.40 dB specifically to memory evolution and 0.26 dB to multi-agent collaboration.","feed_headline":"Self-evolving memory lifts restoration agents to 35.55 dB PSNR","feed_subtitle":"The same tool library improves as the agent remembers which operations help or harm under each degradation.","key_machinery":"The self-evolving causal memory graph: a directed graph whose nodes encode degradation patterns, image regions, restoration tools, actions, quality changes, computational costs, and user preferences, and whose edges store a contextual condition, an estimated effect on restoration quality, and a confidence score. The graph carries the argument because it makes action–outcome relations retrievable, combinable, and updatable: a query over the current image state and user instruction retrieves relevant subgraphs, path-level causal scores turn those subgraphs into expected utility for candidate plans, and a memory curator revises edge weights and confidences after each executed action. The paper","core_discovery":"Causal-AgentIR claims that restoration experience is better stored as a graph than as text or weights. Nodes stand for degradations, image regions, restoration tools, actions, quality-change measurements, computational cost, and user preferences; edges carry a contextual condition, an estimated effect on restoration quality, and a confidence score. During inference the causal-memory agent retrieves relevant subgraphs and scores candidate plans by combining edge confidences and effects across causal paths, so the planner can weigh whether to apply a tool globally or regionally and in what order. After each action, a memory curator evaluates the observed quality change and decides to add, upda","pith_inferences":["Editorial inference: because the graph format is tool-agnostic, the same self-evolving memory mechanism could coordinate tools in other vision agent tasks—video restoration, enhancement, or editing—without changing the graph design.","Editorial inference: the paper's 'causal' edges come from observed trajectories, not controlled interventions; a stricter causal test would randomize or counterfactually reorder tool sequences to check whether the learned effect estimates hold up.","Editorial inference: the reported zero-shot results would be cleaner if memory were reset or frozen across benchmarks; without that control, gains could partly reflect test-time adaptation rather than transferable knowledge."],"forward_implications":["With frozen restoration tools, the reported coordination gains (0.73 dB over the no-memory, no-collaboration baseline) come entirely from planning, memory-guided ordering, and regional application.","Self-evolving causal memory alone contributes 0.40 dB and cuts rollback rate from 0.38 to 0.22, implying the agent avoids harmful operations before executing them.","The best all-in-one restoration result reported, 35.55 dB average PSNR, surpasses every compared model and agent on the combined rain, snow, haze, blur, and noise benchmark.","Real-world zero-shot generalization improves on no-reference metrics for rain, haze, and noise, suggesting memory-based planning transfers to degradations outside the training distribution.","Recorded user preferences raise preference-alignment rate to 0.90, so the same memory mechanism adapts restoration behavior to individual users."],"fun_headline_variants":["Graph memory beats text for image restoration agents","Causal memory graph lifts restoration to 35.55 dB","Agents self-evolve memory as causal graph for sharper images","Why restoration agents remember better as graphs, not text","Causal-AgentIR: adaptive restoration via evolving causal memory"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported gains stand only if the memory graph is populated from prior experience and not from the test images being scored; the paper updates memory after every executed action but never states whether memory is reset or frozen across the benchmarks it calls zero-shot.","fun_headline_variants_meta":{"raw":{"variants":["Graph memory beats text for image restoration agents","Causal memory graph lifts restoration to 35.55 dB","Agents self-evolve memory as causal graph for sharper images","Why restoration agents remember better as graphs, not text","Causal-AgentIR: adaptive restoration via evolving causal memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000134,"raw_usage":{"total_tokens":991,"prompt_tokens":773,"completion_tokens":218,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":137}},"tokens_in":517,"tokens_out":218,"duration_ms":3294,"temperature":1.0,"reasoning_tokens":137,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T08:22:28.083069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the all-in-one benchmark twice with identical agents and tool library—once with the memory graph frozen at its initial state and once with it evolving during evaluation. If the evolving version keeps its reported ~0.40 dB advantage, self-evolution is doing real work; if the advantage disappears, the gains come from adapting to test images.","supporting_citations":[],"review_version":1}