{"id":"aee79fd3-54b1-467f-9a80-98842c5616c4","arxiv_id":"2611.16487","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"HeMeR is an inference framework that reconciles dynamically updated memory contexts with reusable KV states, reducing per-step inference time while maintaining or improving task success in embodied agents.","lead":"This paper presents HeMeR, a framework that reuses previously computed key-value cache states when memory contents change during robot decision-making, reducing computation time. It improves both task success and inference speed across simulated and real-world robotic tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Dependency annotation mechanism is underspecified; its correctness is critical for avoiding stale reuse or excessive recomputation, but the paper provides limited validation.","rationale":"The reader's weakest assumption was cross-adapter drift, which the paper acknowledges, controls via adapter-stable reconciliation, and bounds with ablation (SR from 71.43% to 64.85% when removed) and hidden-state drift analysis. That concern is adequately addressed. My concern targets a different, more foundational assumption: the dependency annotation mechanism that drives the reconciliation plan. The paper does not explain how dependencies are set, nor does it validate their correctness. This is load-bearing because incorrect dependencies would cause either stale KV reuse (hurting task success) or unnecessary recomputation (reducing efficiency gains). However, the paper's strong empirical results (Table 1, Table 2) and the sensitivity experiment (Table 20, varying dependency density from 0% to 100%) suggest that even under extreme densities the method performs well, partially mitigating the concern. Therefore I recommend no change to the verdict, but the concern deserves explicit future work or validation. I disagree with the reader's assessment of the weakest assumption because a more fundamental gap exists in the dependency specification.","tokens_in":34870,"tokens_out":3265,"duration_ms":28711,"concrete_test":"Select one ALFWorld evaluation episode and manually enumerate for each memory-content update the set of chunks whose KV states must be recomputed because their attention can be affected (i.e., any chunk appearing after a changed chunk in the causal sequence). Compare this ground-truth set with the set actually recomputed by HeMeR's dependency annotations. If the recomputed set omits any chunk that should be recomputed, the dependency annotations are incomplete; if it includes extra chunks, they are overly broad. Report the precision and recall of the dependency propagation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"HeMeR relies on declared dependency annotations to propagate recomputation from an updated memory chunk to its direct and transitive dependents (Section 4.3, paragraph 2). The paper does not describe how these dependencies are determined, whether they are manually specified, automatically inferred, or derived from the foundation model's attention pattern. If dependencies are incomplete, a dependent chunk might retain stale KV states after an update, degrading task performance. If they are overapproximate, unnecessary recomputation reduces efficiency. The sensitivity analysis in Table 20 varies dependency density but does not establish that the actual dependencies used in the experiments are correct or complete. Without this validation, the reconciliation plan's ability to localize updates while preserving correctness is undersupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper introduces HeMeR, an inference framework for foundation model-based embodied agents that reconciles dynamically updated memory contexts with reusable KV cache states. HeMeR represents memory contents as typed memory chunks with associated KV ranges, and uses two mechanisms: chunk-level KV reconciliation for content changes (insertions, evictions, reorderings) and adapter-stable KV reconciliation for procedural memory switches (LoRA adapter changes). Experiments on ALFWorld, Habitat, and real-world robotic tasks show that HeMeR improves task success (e.g., +18.57% on ALFWorld, +14.47% on Habitat over MemRL) while reducing per-step inference time (e.g., from 0.90s to 0.18s on ALFWorld).","tokens_in":35136,"tokens_out":2303,"duration_ms":21372,"significance":"If the results hold, HeMeR addresses a practically important problem: efficient KV cache reuse in memory-augmented embodied agents where memory contents change across decisions while many states remain reusable. The paper provides extensive empirical evaluation across three benchmarks, including real-world hardware, with ablations, sensitivity analyses, and representation-level drift characterization. The combination of schema-guided chunk indexing and adapter-aware reconciliation is novel relative to prior work on segment-level reuse (CacheBlend) and adapter switching (LRAgent). The reported efficiency gains (3-5x latency reduction) and task success improvements are substantial. The code and experimental details are provided, supporting reproducibility.","major_comments":[{"comment":"The paper states that when the active adapter changes, HeMeR recomputes the procedural memory chunk's KV range and 'those of its direct and transitive dependents under the declared dependency annotations.' However, the paper never specifies how these dependency annotations are obtained—whether they are manually specified, automatically inferred from the foundation model's attention patterns, or derived from the memory schema's 'dependency annotation' field. The sensitivity analysis in Table 20 varies dependency density but does not validate that the actual annotations used in the experiments are correct or complete. If the annotations are incomplete, stale KV states could be reused, degrading task performance; if overapproximate, the method loses efficiency. This is a load-bearing gap because the correctness of the reconciliation plan (Equation 6 and Algorithm 1) depends on accurate dependency tracking. The authors should clarify the provenance of dependency annotations and provide evidence (e.g., human evaluation, automatic validation) that they are correct for the evaluated tasks.","section":"Section 4.3, paragraph 2; Algorithm 1, line 7","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The caption says 'Analysis on cache-processing cost across reconciliation operations' but the table includes counts for events and tokens, not just costs. Consider rewording to 'Cache-processing cost decomposition by reconciliation operation'.","section":"Table 3, caption","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'The efficiency and memory behavior metrics explain this trend' is vague. 'these metrics explain this trend' is imprecise; consider specifying which metrics (e.g., reuse ratio, active action ratio) explain the success gain.","section":"Section 5.2, paragraph 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The notation `use the same backbackground` is ambiguous: in Algorithm 1, line 13, the RoPE remapping formula should be written as `tilde{k}_j = R_theta(p'_j) R_theta(p_j)^{-1} k_j` for clarity, matching Equation (7). The current text uses inline notation but could b e more explicit.","section":"Appendix B.2, pseudocode","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The phrase 'Rather than treating the composed memory context Xt ... solely as a flat token sequence' is followed by 'HEMER organizes it as a sequence of memory chunks.' This is clear, but the subsequent sentence listing schema fields is a run-on; consider breaking it into a list or shorter sentences for readability.","section":"Section 4.1, paragraph 3","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This paper does a clean job of identifying a real problem: embodied agents with multiple memory modules recompute the full KV cache at every step because they treat updated memory contexts as completely new inputs. HeMeR’s framing of this as a KV reconciliation problem is the main conceptual contribution. The engineering integration – chunk-level indexing, RoPE remapping for position shifts, adapter-aware reconciliation for procedural switches – is sensible and put together in a way that the individual pieces alone don't cover.\n\nThe experiments are the paper’s strongest point. Three benchmarks, including real-world robot trials, with ablations and error bars. The cost decomposition (Table 3) shows that the majority of tokens are reused, and the ablation (Table 4) cleanly separates the contributions of each component. The hidden-state drift analysis (Figure 9) also gives reasonable confidence that the approximation doesn’t silently accumulate error. The authors acknowledge the shared-attention LoRA design limits cross-adapter drift but don’t eliminate it, and they quantify the drift (Table 18). That is fair.\n\nThe one thing that genuinely nags me is the dependency annotation mechanism described in Section 4.3. The paper says that when an adapter switches, HeMeR recomputes \"the procedural memory chunk’s KV range and those of its direct and transitive dependents under the declared dependency annotations.\" But it never says how these dependencies are determined – manually written, inferred from attention patterns, or something else. The stress-test note on this is correct: if the annotations are incomplete, stale KV states might be reused; if they are over-approximate, the efficiency gain erodes. Table 20 varies dependency density for sensitivity but does not validate that the actual annotations used in the main experiments are correct. This is an underspecied load-bearing detail.\n\nThat said, the paper is otherwise well-executed and the central argument holds up. This is written for people building efficient serving stacks for memory-augmented robotics agents. It deserves a serious refereeing and should be accepted with a request to clarify the dependency annotation mechanism.","headline":"Solid systems paper that frames heterogeneous memory updates as a KV reconciliation problem and validates it thoroughly, though the dependency annotation mechanism is underspecified.","tokens_in":35612,"tokens_out":1340,"duration_ms":13178},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"HeMeR selectively reuses KV cache across memory updates, cutting embodied-agent inference time by up to 80% while improving task success.","keywords":["KV cache reuse","heterogeneous memory","embodied agents","LoRA adapters","foundation model inference","long-horizon tasks","memory reconciliation"],"falsifier":"If a set of tasks requires frequent switching among highly diverse adapters and HeMeR's task success degrades to or below the level of LRAgent (which recomputes for each adapter switch), that would show cross-adapter drift undermines the efficiency gain and invalidates the claim of robust KV reuse under procedural changes.","tokens_in":37750,"feed_emoji":"🤖","tokens_out":1543,"duration_ms":35894,"temperature":0.7,"pith_summary":"Embodied agents rely on declarative, procedural, and working memory modules that change across decisions. Current systems recompute the full key-value (KV) cache each time, wasting computation. HeMeR treats memory contents as typed chunks with associated KV ranges, maps updates to those ranges, and reuses or remaps unchanged states. On ALFWorld, Habitat, and real-world manipulation tasks, it achieves higher success rates and 2–5× lower per-step latency than strong baselines, without full cache rebuilds.","feed_headline":"HeMeR speeds embodied agents 5x via selective KV cache reuse","feed_subtitle":"Only recompute what changed: maps memory updates to KV ranges and reuses the rest.","key_machinery":"Three components working together: (1) a memory schema that records each chunk's persistent ID, type, content hash, adapter identifier, and KV range; (2) chunk-level KV reconciliation that matches chunks across steps and reuses or recomputes based on schema compatibility, with RoPE remapping for position-only shifts; (3) adapter-stable KV reconciliation that includes the adapter identifier in the compatibility check and recomputes the procedural chunk's KV range when the adapter switches, while reusing other unaffected ranges.","core_discovery":"HeMeR establishes that heterogeneous memory updates—insertions, evictions, reorderings, and procedural-adapter switches—can be handled by reconciling only the affected KV ranges, preserving compatible states for unchanged content. The framework uses a memory schema to index chunks, RoPE-based position remapping for pure shifts, and adapter-stable reconciliation via shared-attention LoRA to limit cross-adapter drift. This enables efficient, long-horizon inference without full KV recomputation.","pith_inferences":["The schema-guided reconciliation approach could generalize beyond embodiment to any transformer-powered agent that maintains a mutable context (e.g., personal assistants that accumulate user preferences).","The paper's reliance on shared-attention LoRA to limit cross-adapter drift is a practical concession; if more diverse adapters are introduced, the assumption of small drift may need revisiting, perhaps through learned compatibility metrics.","The structured chunk representation hints at a principled cache-coherence protocol for LLM serving, analogous to invalidation-based coherence in distributed systems.","HeMeR's ablations suggest that even without adapter-stable reconciliation, reuse remains high but task success drops, indicating that representation quality matters more than raw cache hitrate."],"forward_implications":["Enables embodied agents to operate over longer episodes with significantly less computational overhead, making them more practical for real-world deployment.","Provides a template for KV reuse in any interactive system where the input context evolves incrementally (e.g., conversational AI, code editing).","Reduces the gap between simulation and physical robot trials by lowering per-step latency to near-real-time levels.","Opens the possibility of concurrent KV reuse in multi-agent systems once concurrency control is added."],"supporting_citations":[{"why":"Provides the ReAct baseline for reasoning-and-acting prompting without explicit memory reconciliation.","marker":"[2]"},{"why":"Provides the Reflexion baseline for self-reflective feedback from prior failures.","marker":"[7]"},{"why":"Provides the MemRL baseline for episodic memory retrieval and use without KV cache reuse.","marker":"[35]"},{"why":"Provides the CacheBlend baseline for segment-level KV reuse with selective recomputation at boundaries.","marker":"[21]"},{"why":"Provides the KEEP baseline for invalidation-based KV cache management in embodied reasoning.","marker":"[47]"},{"why":"Provides the LRAgent baseline for LoRA-based procedural skill switching without declarative/working memory updates.","marker":"[25]"},{"why":"Supplies the LoRA methodology used to implement procedural memory adapters and the shared-attention design.","marker":"[16]"},{"why":"Supplies the RoPE positional embedding used for key remapping under chunk reordering.","marker":"[24]"},{"why":"Supplies the ALFWorld text-based household simulation for environment and task definition.","marker":"[32]"},{"why":"Supplies the EmbodiedBench Habitat benchmark for vision-based multimodal embodied tasks.","marker":"[33]"}],"fun_headline_variants":["HeMeR reconciles only updated memory chunks for inference","Selective KV reuse: HeMeR avoids full recomputation on memory change","HeMeR maps memory updates to KV ranges, reuses the rest","HeMeR: partial KV recompute for heterogeneous memory modules"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The shared-attention LoRA design limits but does not eliminate cross-adapter drift; the assumption that this drift is small enough to allow KV reuse across adapter switches is load-bearing.","fun_headline_variants_meta":{"raw":{"variants":["HeMeR reconciles only updated memory chunks for inference","Selective KV reuse: HeMeR avoids full recomputation on memory change","HeMeR maps memory updates to KV ranges, reuses the rest","HeMeR: partial KV recompute for heterogeneous memory modules"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"cost":0.000258188,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000258188,"upstream_inference_prompt_cost":9.5788e-05,"upstream_inference_completions_cost":0.0001624},"total_tokens":1469,"prompt_tokens":889,"completion_tokens":580,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":513}},"tokens_in":889,"tokens_out":580,"cost_source":"openrouter_usage","duration_ms":5634,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":513,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T14:53:59.998562+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"If a set of tasks requires frequent switching among highly diverse adapters and HeMeR's task success degrades to or below the level of LRAgent (which recomputes for each adapter switch), that would show cross-adapter drift undermines the efficiency gain and invalidates the claim of robust KV reuse under procedural changes.","supporting_citations":[{"cited_title":"ReAct: Synergizing Reasoning and Acting in Language Models","cited_arxiv_id":null,"evidence_quote":"Provides the ReAct baseline for reasoning-and-acting prompting without explicit memory reconciliation."},{"cited_title":"Reflexion: language agents with verbal reinforcement learning","cited_arxiv_id":null,"evidence_quote":"Provides the Reflexion baseline for self-reflective feedback from prior failures."},{"cited_title":"MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory","cited_arxiv_id":"2601.03192","evidence_quote":"Provides the MemRL baseline for episodic memory retrieval and use without KV cache reuse."},{"cited_title":"Cacheblend: Fast large language model serving for rag with cached knowledge fusion","cited_arxiv_id":null,"evidence_quote":"Provides the CacheBlend baseline for segment-level KV reuse with selective recomputation at boundaries."},{"cited_title":"KEEP: A KV-Cache-Centric Memory Management System for Efficient Embodied Planning","cited_arxiv_id":null,"evidence_quote":"Provides the KEEP baseline for invalidation-based KV cache management in embodied reasoning."},{"cited_title":"LRAgent: Efficient KV Cache Sharing for Multi-LoRA LLM Agents","cited_arxiv_id":"2602.01053","evidence_quote":"Provides the LRAgent baseline for LoRA-based procedural skill switching without declarative/working memory updates."},{"cited_title":"LoRA: Low-Rank Adaptation of Large Language Models","cited_arxiv_id":null,"evidence_quote":"Supplies the LoRA methodology used to implement procedural memory adapters and the shared-attention design."},{"cited_title":"RoFormer: Enhanced transformer with Rotary Position Embedding","cited_arxiv_id":null,"evidence_quote":"Supplies the RoPE positional embedding used for key remapping under chunk reordering."},{"cited_title":"Rapid object detection using a boosted cascade of simple features","cited_arxiv_id":"2010.03768","evidence_quote":"Supplies the ALFWorld text-based household simulation for environment and task definition."},{"cited_title":"EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents","cited_arxiv_id":null,"evidence_quote":"Supplies the EmbodiedBench Habitat benchmark for vision-based multimodal embodied tasks."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}