{"id":"dde675b2-e2ee-4537-b748-4fa4b54505ad","arxiv_id":"2508.08100","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A small language model formats optimal A* routes on a binarized floor plan into natural-language indoor walking directions.","lead":"Grid2Guide turns a floor plan into a grid, finds a route with the A* algorithm, and uses a small language model to phrase the route as numbered walking directions. It is proposed as a beacon-free indoor navigation option, but its accuracy and real-time claims depend on manually prepared maps and on an unmeasured text-generation step.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"A* diagonal moves can cut blocked wall corners, so the claimed guaranteed-optimal walkable routes are unvalidated","rationale":"The reader's weakest-assumption analysis points to diagonal handling, and I agree that diagonal walkability is the load-bearing soft spot. However, the reader's specific mechanism—Stage 3 diagonal collapse replacing (S,1)+(E,1) with (SE,1)—is less likely to appear on a true 8-way A* optimal path, since the direct diagonal would have lower cost whenever the destination cell is free. The stronger and more general problem is in Algorithm 1 itself: diagonal transitions are permitted without checking adjacent orthogonal cells, so wall-corner cutting is possible even before compression. This directly undermines the central 'accurate optimal routes' claim, and it is not a matter of missing evaluation data but of the algorithm's transition model. Because the reported 100% accuracy metric is path-existence rather than physical walkability, the evaluation cannot detect this failure. A corner-aware replay would settle it. I partially agree with the reader, and the verdict remains REJECT/UNCHANGED since the central guarantee is unsupported.","tokens_in":13742,"tokens_out":6258,"duration_ms":87273,"concrete_test":"Replay the 80 published start/goal pairs on the same occupancy grids using a modified A* transition rule: allow a diagonal (Δi,Δj) only if both orthogonal neighbors (i+Δi,j) and (i,j+Δj) are free, keeping all other code identical. If any original path contains a diagonal whose orthogonal neighbor is blocked, or if any replayed path differs at such a location and the original segment crosses a blocked cell, the route-guarantee claim fails. Also run a minimal unit test: grid with G[0][0]=1, G[1][1]=1, G[1][0]=0, G[0][1]=0; current A* returns the diagonal (SE,1), which crosses the blocked corner.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline guarantee of 'accurate optimal routes' depends entirely on the walkability model encoded in the 8-way A* transition. In Algorithm 1, the neighbor check at line 17 accepts a diagonal move if the destination cell v is in bounds and G[v]=1; it never checks whether the two orthogonally adjacent cells (i+Δi,j) and (i,j+Δj) are free. Consequently, a diagonal step can pass through the corner of two blocked cells, i.e., through a wall corner. This is not a matter of external convention: on the coarse grids used, e.g., Bergamo at only 30×107 cells, wall corners are routinely discretized into blocked cells, so the path returned as optimal may be physically non-walkable. Section III.4's diagonal collapse has the same blind spot, though on a true 8-way optimal path consecutive (S,1)+(E,1) pairs should be rare because the direct diagonal is cheaper; the fundamental issue is that A* itself already introduces diagonals without corner-adjacency checks. The reported 100% route accuracy in Table 6 cannot expose this because it only verifies that a path is produced, not that every step stays in walkable space. Thus the central claim that the system guarantees accurate, physically valid optimal routes is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Grid2Guide proposes a five-stage indoor navigation pipeline: (i) convert a floorplan image into a binary occupancy grid; (ii) build an implicit graph with 8-way connectivity; (iii) run A* with Chebyshev heuristic to obtain an optimal grid path; (iv) compress the path into terse run-length directional commands, including a diagonal-collapse step; (v) use TinyLlama-1.1B, instruction-tuned with 1000 examples, to turn the terse commands into natural-language numbered directions. The paper claims that offloading all spatial reasoning to A* guarantees accurate optimal routes, that portal nodes enable multi-floor navigation, and that the whole pipeline runs in near-real-time on commodity CPUs. Experiments on four indoor maps (80 route queries) report A* runtimes under 5 ms, total pipeline time averaging 16.92 s, and a 100% route-success rate, compared against an LLM-only baseline (Coffrini et al.).","tokens_in":13997,"tokens_out":4313,"duration_ms":50346,"significance":"The core architectural idea—decoupling deterministic path planning from language generation so that a small language model only performs text formatting—is pragmatic and potentially useful for resource-constrained indoor navigation. If the path were guaranteed physically walkable and the generated instructions were shown to be accurate and useful, the system would be a meaningful lightweight alternative to LLM-based navigation without infrastructure. However, the paper's central claims are not currently established: the A* movement model permits diagonal crossings through blocked wall corners, the reported 100% accuracy is a tautology of A* completeness rather than a measure of instruction quality, and no evaluation of the final natural-language output is provided. The contribution, as presented, is therefore more of a system sketch than a validated navigation solution.","major_comments":[{"comment":"The 8-way neighbor expansion checks only whether the destination cell v is in bounds and G[v]=1 (line 17); it never verifies that the two orthogonally adjacent cells (i+Δi,j) and (i,j+Δj) are free. On coarse grids such as the 30×107 Bergamo map, a diagonal step can therefore cut through the corner of a wall. This invalidates the claim in Contribution 1 of 'guaranteeing accurate optimal routes' that are physically walkable. The fix is standard: permit a diagonal move only when both orthogonal neighbors are free, and update Algorithm 1 and the cost model accordingly.","section":"Algorithm 1 and §III.3"},{"comment":"The compression routine replaces an adjacent pair (S,1)+(E,1) with (SE,1) without checking whether the SE cell is free. This is not just a cosmetic issue: if the A* path contains such adjacent orthogonal steps, the direct diagonal cell is usually blocked (otherwise A* would have preferred the cheaper √2 diagonal). Thus the collapse can emit an instruction that sends the user through a wall or obstacle. This further undermines the claim of 'accurate optimal routes' and needs correction before the route-integrity claim can be accepted.","section":"§III.4, Stage 3 (Diagonal Collapse)"},{"comment":"The headline '100% accuracy' is the route-generation rate of A*, which is complete by construction: for any origin–destination pair that has a path, A* will return one. Thus the 100% figure is tautological and does not test the SLM, the generated instructions, or the physical validity of the route. The paper provides no metric—human ratings, instruction correctness, agreement with ground-truth routes—for the natural-language output that is the system's stated contribution. Without such an evaluation, the comparison with Coffrini et al. in Table 6 is not meaningful.","section":"§IV.5 and Table 6"}],"minor_comments":[{"comment":"In Table 1, the cost formula lists the same condition '|Δi|+|Δj|=1' for both orthogonal and diagonal moves; the diagonal case should be '=2'. In §IV.3, the text says orthogonal moves have 'cost = 10' while Table 1 says cost=1; the text should say cost=1.","section":"Table 1 and §IV.3"},{"comment":"The example system-prompt output has two numbered steps both labeled '2.' (the escalator step and the final step). This is likely a typo but should be fixed since the prompt instructs the model to produce one number per line.","section":"§IV.4"},{"comment":"The phrase 'true real-time performance' is not supported by the reported 16.92 s average total time (Table 5). While this is far faster than the 4–5 minute LLM baseline, the paper should define its real-time target and, ideally, report perceived-latency acceptability from a user study.","section":"§IV.5"},{"comment":"The occupancy grid and portal nodes are manually constructed ('interactive grid layer' and 'portal nodes are manually defined'). This should be explicitly stated as a limitation: the 'infrastructure-free' claim refers to runtime infrastructure, not to the one-time human effort needed per map. Also, the reference formatting duplicates 'Coffrini et al.' in several places.","section":"§III.1 and §II"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is reasonable, but the current version is not yet a validated system: the path-validity flaw and the absence of any instruction-quality evaluation are load-bearing. These issues are fixable within the manuscript's scope (e.g., correct the diagonal-move rule, re-run experiments, and add an evaluation of the SLM output). The novelty is modest—A* plus a small language model is a natural engineering combination—so the editors may also wish to consider whether the contribution depth meets the journal's bar even after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a clean engineering integration—A* on a binarized grid, run-length compression, and TinyLlama for text generation—and the A* core is standard and runs fast. But the headline '100% route accuracy' is a tautology, and the 8-way move set can cut through blocked wall corners, so the central guarantee of walkable optimal paths is not actually established.\n\nThe genuinely useful part is the decoupling: letting A* do all spatial reasoning and keeping the SLM as a text formatter is a sensible design, and the runtime comparison to a raw LLM (Coffrini et al.) is a fair point. The A* implementation is correct as a graph search, and the millisecond runtimes on these small grids are believable.\n\nThe soft spots are load-bearing. First, Table 6's 100% accuracy measures only whether A* found a path when one exists—that is true by construction and says nothing about the SLM or the final directions. There is no human or text-based evaluation of the generated walking instructions, which is the actual user-facing product. Second, Algorithm 1 admits a diagonal move if the destination cell is free, without checking the two orthogonal cells; so a path may pass between two blocked cells, i.e., through a wall corner. Stage 3's diagonal collapse has the same issue. On the coarse grids used, this is a real threat to the claim of accurate, walkable routes. Third, the 'infrastructure-free' framing is weakened by the manual mask creation and portal placement. And calling 16-17 seconds per query 'true real-time' is a stretch.\n\nMinor issues: the cost formula has a typo (both entries condition on |Δi|+|Δj|=1), and the literature claim of 'no prior work' on combining path planning with LLMs is not adequately supported.\n\nVerdict: as submitted, the central claims are not supported, so a referee would reasonably reject. But the system is real, the A* engine is sound, and the flaws are fixable—add corner checks, evaluate the instructions, and report meaningful route quality. That makes it worth a serious referee rather than a desk reject. I would not cite it in its current form, but it could be a useful reading-group case study in evaluation fallacies.","headline":"A sensible A*+SLM integration undercut by a tautological accuracy metric and a diagonal-move bug that breaks walkability guarantees.","tokens_in":14516,"tokens_out":3896,"would_cite":false,"duration_ms":43016,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Grid2Guide claims that indoor navigation can be made fast and accurate by shifting all spatial reasoning to A* search on a grid, leaving a small language model only the task of turning terse route commands into natural language.","keywords":["indoor navigation","A* search","occupancy grid","small language model","path planning","natural language instructions","floor plan","TinyLlama"],"falsifier":"Take a floor plan with a corner where an orthogonal step south followed by a step east would cross a blocked diagonal cell; ask the pipeline for a route that turns that corner. If the compressed output contains a single diagonal move such as (SE,1) and the corresponding diagonal cell is occupied, the instruction is not walkable, directly contradicting the claim of 100% accuracy and guaranteed optimal routes.","tokens_in":13580,"feed_emoji":"🧭","tokens_out":4029,"duration_ms":43333,"temperature":0.7,"pith_summary":"The paper proposes Grid2Guide, a pipeline that converts a floor plan into a binary occupancy grid, runs A* search with diagonal moves to compute an optimal route, compresses that route into short directional commands, and uses a small language model (TinyLlama) to rewrite the commands as human-friendly walking instructions. The central claim is that offloading all spatial reasoning to A* yields a deterministic, optimal route in milliseconds, while the SLM only formats text, avoiding the slow and error-prone image reasoning of large language models. The authors report 100% route accuracy across 80 test cases on four real indoor maps, with total query time averaging under 20 seconds on CPU hardware. If correct, this offers an infrastructure-free, lightweight indoor navigation assistant that runs on commodity devices.","feed_headline":"A* finds the route, a small language model writes it","feed_subtitle":"Deterministic grid search replaces LLM spatial reasoning, yielding 100% route accuracy in under 20 seconds per query.","key_machinery":"The load-bearing object is the binary occupancy grid $G \\in \\{0,1\\}^{m \\times n}$, with every free cell as a node connected to its eight neighbors. A* uses edge costs $c=1$ for orthogonal and $c=\\sqrt{2}$ for diagonal steps, plus an admissible and consistent Chebyshev heuristic $h((i,j),(i_t,j_t)) = \\max\\{|i-i_t|, |j-j_t|\\}$, guaranteeing an optimal path. A three-stage compression (vectorization, run-length encoding, diagonal collapse) converts the cell sequence into terse commands, which the SLM then turns into numbered walking instructions.","core_discovery":"The core discovery is a modular decomposition: spatial reasoning is handled entirely by A* on a binarized occupancy grid with 8-way connectivity, edge costs of $1$ for orthogonal moves and $\\sqrt{2}$ for diagonal moves, and an admissible Chebyshev heuristic. Because A* is deterministic, the route is identical on every run, and because the SLM only transforms text, the system avoids the 4–5 minute image-processing latency and spatial hallucinations reported for prior LLM-based navigation. The paper reports 100% successful route generation on all four tested maps, compared with 62–82% for the ChatGPT-based baseline, and total execution time between 14 and 21 seconds, dominated by SLM text gene","pith_inferences":["The diagonal-collapse stage in Stage 3 merges adjacent orthogonal steps such as (S,1)+(E,1) into (SE,1) without verifying that the diagonal cell is walkable; on floor plans with diagonal walls or tight corners this can produce instructions that cut through blocked regions, so the reported 100% accuracy may not generalize to all real layouts.","The manual effort needed to construct the occupancy grid and place portal nodes is a hidden cost; automating grid creation from CAD files or sensor data would make the pipeline truly turnkey, an extension the paper mentions only as future work.","The 14–21 second total latency is dominated by SLM generation, not A*; using a smaller or quantized model, or caching common route phrasings, could bring the user-visible response much closer to the sub-5ms pathfinding time.","The comparison with the Coffrini baseline is based on route-generation success, not on user satisfaction or instruction comprehension; a field study with real users would be needed to judge whether the SLM output is genuinely helpful."],"forward_implications":["If correct, real-time indoor navigation instructions can be generated on CPU-only handheld devices without Wi-Fi, beacons, RFID, or other dedicated infrastructure.","The occupancy grid is built once per map and reused for every query, so the preprocessing cost is paid once while each A* query takes under 5 milliseconds.","Because the SLM receives only text and is model-agnostic, any instruction-tuned small language model can replace TinyLlama without changing the pathfinding or compression stages.","Route accuracy no longer depends on the language model's probabilistic understanding of images, decoupling correctness from model scale and prompt tuning.","The text-only output can be converted to speech, offering a path to navigation assistance for visually impaired users."],"supporting_citations":[{"why":"Supplies the A* algorithm whose completeness and optimality guarantee the claimed route accuracy.","marker":"Hart et al. (1968)"},{"why":"Provides the LLM-based baseline and its accuracy figures (62–82%) and 4–5 minute processing times that the paper compares against.","marker":"Coffrini et al. (2025)"},{"why":"Describes TinyLlama, the small language model used in the pipeline for converting terse commands to natural language.","marker":"Zhang et al. (2024)"}],"fun_headline_variants":["Grid2Guide: A* pathfinding, SLM narrates, 100% success","A* plans, SLM speaks: indoor navigation without GPS","Indoor navigation without infrastructure: A* and SLM team up","A* computes optimal path, SLM writes clear directions","Grid2Guide: A* finds route, SLM writes steps, no infrastructure"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The manually constructed occupancy grid, the manually placed portal nodes, and the diagonal-collapse compression assume that the reduced path stays on physically walkable cells, so the claimed optimality and 100% accuracy can break on real floor plans where a diagonal shortcut crosses a blocked region.","fun_headline_variants_meta":{"raw":{"variants":["Grid2Guide: A* pathfinding, SLM narrates, 100% success","A* plans, SLM speaks: indoor navigation without GPS","Indoor navigation without infrastructure: A* and SLM team up","A* computes optimal path, SLM writes clear directions","Grid2Guide: A* finds route, SLM writes steps, no infrastructure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2650,"prompt_tokens":665,"completion_tokens":1985,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":409,"completion_tokens_details":{"reasoning_tokens":1889}},"tokens_in":409,"tokens_out":1985,"duration_ms":15922,"temperature":1.0,"reasoning_tokens":1889,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:39:09.825563+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a floor plan with a corner where an orthogonal step south followed by a step east would cross a blocked diagonal cell; ask the pipeline for a route that turns that corner. If the compressed output contains a single diagonal move such as (SE,1) and the corresponding diagonal cell is occupied, the instruction is not walkable, directly contradicting the claim of 100% accuracy and guaranteed optimal routes.","supporting_citations":[{"cited_title":"E., Nilsson, N","cited_arxiv_id":null,"evidence_quote":"Supplies the A* algorithm whose completeness and optimality guarantee the claimed route accuracy."},{"cited_title":"A., Barsocchi, P., Furfari, F., Crivello, A., and Ferrari, A","cited_arxiv_id":null,"evidence_quote":"Provides the LLM-based baseline and its accuracy figures (62–82%) and 4–5 minute processing times that the paper compares against."}],"review_version":1}