{"id":"f8245f33-2073-4293-b55f-ae8f589631d5","arxiv_id":"2602.08964","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An LLM navigation agent encodes a coarse spatial map and multi-step plans in its activations, and reasoning shifts these representations from broad environment information to immediate action selection.","lead":"This paper tests whether an LLM agent navigating a grid world really pursues its assigned goal by combining behavioural scores with probes of its internal states. It finds the agent encodes a coarse map of the grid and goal, and that reasoning narrows this representation toward the next move.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Probing results lack controls for coordinate-prior and input-copy shortcuts, leaving the central representational claim unsecured.","rationale":"The reader's weakest_assumption identifies precisely this gap: probes may recover artifacts of probe capacity, coordinate priors, or the fact that the full grid text is in the input, rather than the model's internal beliefs. I agree that this is the load-bearing vulnerability. The behavioural evaluation and the pre/post-reasoning shift are interesting, but the paper's headline contribution is the representational evaluation, and that contribution is unsecured without controls. The reader's CONDITIONAL verdict is appropriate; my analysis does not move it, hence UNCHANGED. The concrete test above is designed to settle whether the probe is exploiting coordinates/shortcuts; if it fails, the representational claims should be substantially revised or softened. The paper's own footnote 6 acknowledges the ambiguity of decoded positions, which further supports the need for controls before causal attribution of failures to fuzzy maps.","tokens_in":26300,"tokens_out":5548,"duration_ms":63574,"concrete_test":"Train the same MLP probe on the same activation–label pairs but with the (x, y) coordinates randomly permuted across training examples, breaking the correspondence between each activation and its queried cell while preserving the coordinate marginal and label distribution. If accuracy remains near the reported ~70% (chance is 20% for five classes), the probe is exploiting coordinate priors or input-copy shortcuts rather than activation content. If accuracy collapses to near chance, the activations genuinely encode cell identity. As a secondary check, compare against a no-activation baseline that receives only coordinates and labels: if that baseline already achieves high accuracy, the probe's representational claim is vacuous.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that the LLM 'non-linearly encodes a coarse spatial map' and that actions are 'broadly consistent with these internal representations' rests on the cognitive-map probes of §5.1. There, training examples are formed by augmenting each activation with the queried cell's (x, y) coordinates, yielding inputs ([act, x, y], c). Because the model has full observability, the full grid text is literally present in the input context, so the extracted activations (from final prompt tokens) plausibly contain a near-literal encoding of the grid. The probe could therefore achieve high accuracy by combining a coordinate prior (e.g., goal tends to be near center, walls on border) with a simple readout of the input grid, rather than by decoding any internal 'belief' or 'cognitive map.' No control probes are reported: no random-feature baseline, no coordinate-shuffle ablation, no majority-class or coordinate-only baseline, and no test on activations from a model that did not see the grid. This is not a minor omission: the paper's contribution is precisely that internal representations add information beyond behaviour, and the §5.2 'Rec.' metric—used to attribute suboptimal actions to 'fuzzy world representations'—is computed from these same probes. If the probe is exploiting shortcuts, the decoded maps are artifacts, and the claimed behavioural-consistency and reasoning-reorganisation findings lose their force. The plan decoder (§5.3) faces a related risk: it may be reading the grid from activations to infer a path, which would not demonstrate goal-directed planning.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework that combines behavioural evaluation with representation probing to assess goal-directedness of an LLM agent in a grid-world navigation task. Behaviourally, it compares the agent's actions to A*-optimal policies across grid sizes, obstacle densities, iso-difficulty transformations, and instrumental/implicit-goal variants. Representationally, it trains linear and MLP probes on last-token activations, augmented with queried-cell (x, y) coordinates, to decode cell types ('cognitive maps'); it then checks whether the agent's actions are optimal with respect to its decoded map, and trains a one-shot transformer decoder to read multi-step action plans from activations. The paper concludes that the agent non-linearly encodes a coarse spatial map preserving task-relevant cues, that its actions are broadly consistent with this map, and that reasoning reorganises representations from broader spatial/long-horizon information toward immediate action selection. The authors argue that these findings support the need for introspective, representation-based evaluation beyond behaviour.","tokens_in":26684,"tokens_out":6587,"duration_ms":76980,"significance":"If the representational results were secured, the paper would make a useful contribution to goal-directedness evaluation: the behavioural protocols are carefully designed, the iso-difficulty and instrumental-goal experiments are informative, the Wilcoxon signed-rank tests are appropriate, and the one-shot plan decoder is a thoughtful attempt to avoid probe-created autoregressive structure. The behavioural findings—especially the attraction to a reward-irrelevant key and robustness to iso-difficulty transformations—are credible and interesting. However, the paper's central novelty, the claim that internal representations add information beyond behaviour, rests on probe results that currently lack essential controls. The cognitive-map probes are trained on activations from a model that literally sees the full grid text, yet no baselines are given for coordinate-prior, input-copy, or majority-class shortcuts. The plan-decoder comparison is only against random guessing. Consequently, the significance is conditional on the representational claims surviving appropriate controls.","major_comments":[{"comment":"The cognitive-map probe is trained by concatenating activations with (x, y) coordinate features, and the model's input context contains the full grid text. No control probes are reported. For small grids padded to 15×15, the padding class covers 176/225 ≈ 78% of cells for a 7×7 grid, so a trivial all-padding predictor already achieves ~78% accuracy; the reported ~70% overall accuracy is therefore uninterpretable without a per-size majority baseline. More broadly, high accuracy could reflect coordinate-prior or input-copy shortcuts rather than an internal 'belief' or 'cognitive map.' The authors should add: majority-class baselines per grid size, coordinate-only probes, probes on random/shuffled activations, and an ablation that hides or scrambles the grid input. This is required to support the abstract's claim of a non-linear coarse spatial map.","section":"§5.1"},{"comment":"It is never stated whether probe training and test examples are split by grid instance. If activations from the same grid configuration appear in both training and test sets, the probe can memorize grid-specific features, inflating all decoding accuracies and the §5.2 policy-consistency metrics. Please state the instance-level split; if none exists, rerun with disjoint grids.","section":"§5.1–§5.3"},{"comment":"The Recovery metric and the conclusion that 'a substantial fraction of failures can be attributed to inaccurate, or fuzzy world representations' assume that the decoded grid is a valid estimate of the agent's internal beliefs. Because the probe lacks the controls above and is trained on ground-truth labels from the same observably available grid, the decoded map may simply mirror the input text. The Rec. metric then measures consistency with the input, not with internal beliefs. The authors need an independent validation of the probe (e.g., intervention on the decoded map, or consistency under input perturbations that change the ground truth) before this conclusion can be drawn.","section":"§5.2"},{"comment":"Plan decoding is only compared against random chance (0.25^N). Since the target action sequence is the executed trajectory of the same model, and the full grid is in the prompt, a probe could achieve above-random prefix accuracy by learning a direct input-to-action policy rather than decoding a multi-step plan. The one-shot decoder rules out autoregressive shortcuts, but not this input-copy shortcut. Please add a control using activations from a model that never received the grid, or from representations where the grid text is permuted, and report whether prefix accuracy above baseline persists.","section":"§5.3"}],"minor_comments":[{"comment":"Table 3 reports eight Wilcoxon signed-rank tests without correction for multiple comparisons. Adjusting p-values or reporting them as exploratory would help; the reported effect sizes are appreciated.","section":"§4.2"},{"comment":"No confidence intervals or significance tests are reported for probe accuracy differences (e.g., MLP vs. linear, 7×7 vs. 11×11). Given the small number of grids, bootstrap confidence intervals would be appropriate.","section":"Fig. 6"},{"comment":"Appendix A describes planned partially-observable experiments in the present/future tense ('we evaluate...', 'we ideally are able...') but these were not carried out. Please mark this clearly as a proposal rather than completed analysis.","section":"App. A"},{"comment":"The footnote acknowledges that decoded maps may reflect multiple nearby cells rather than a single location. This undercuts the single-cell decoding used for Acc-Dec and Rec. Consider deriving a distribution over policies to account for the probe's uncertainty.","section":"§5.2, footnote 6"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is valid and central. The behavioural evaluation is sound and could be published on its own, but the representational claims—the paper's main novelty—are not yet supported without probe controls and baseline corrections. The missing analyses are feasible within the scope of the paper, so I recommend major revision rather than rejection. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. First, the behavioral half of this paper is genuinely solid: the iso-difficulty transformations, the multi-goal variants, and the key-artifact bias results are cleanly measured with paired Wilcoxon tests and random grid generation. Second, the representational half—which is the real contribution—has a load-bearing hole: the probes lack the right controls, so the \"cognitive map\" and the claim that actions are \"broadly consistent with internal representations\" are not yet established.\n\nWhat is actually new: the integration of behavioral robustness tests with internal-state probing in one framework, the pre/post-reasoning decodability crossing in the plan decoder, and the observation that goal-like artifacts (a key) attract the agent even when functionally useless. The plan decoder design is thoughtful—one-shot decoding rather than autoregressive avoids a real confound, and the increasing gap from the random baseline is a meaningful observation. The paper also deserves credit for being honest about its limitations, including the partial-observability failures in the appendix.\n\nThe soft spots are concentrated in §5.1 and §5.2. The probes are trained on activations taken after the model has seen the full grid text, and they are conditioned on (x, y) coordinates. Without a padding-majority baseline, a coordinate-only baseline, or control probes on shuffled activations or random features, the 70% MLP probe accuracy could be largely a nonlinear readout of the input grid plus coordinate priors rather than evidence of an internal spatial map. The same issue undermines the \"Rec.\" metric in §5.2: attributing suboptimal actions to \"fuzzy world representations\" is circular if the fuzziness comes from probe error rather than the model's actual beliefs. The plan decoder is less exposed because the one-shot design prevents the probe from generating trajectory structure, but it still could plan using the decoded grid, so the claim that the model itself pre-encodes multi-step plans needs an extra control or a more cautious interpretation. These are addressable weaknesses rather than fatal ones—adding controls and re-running the probes would either settle or substantially soften the concern.\n\nWho this is for: interpretability and AI-safety researchers who want a reusable template for connecting behavior to internal state. The behavioral findings are a useful reference on their own. The paper does deserve a serious referee—the framework is worth engaging with and the central question is important. But my recommendation would be conditional acceptance at best: the authors should add the missing baselines and controls, release the code and data, and either provide evidence against the input-copy shortcut or pull back the causal reading of the decoded-map evaluation.","headline":"Solid behavioral package, real methodological novelty, but the representational claims are not yet secured: the probing section needs basic controls before the cognitive-map story holds.","tokens_in":27160,"tokens_out":3386,"would_cite":true,"duration_ms":46042,"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":"The paper argues that goal-directedness in an LLM agent requires both behavioural evaluation and probing of internal representations, and shows the agent encodes a coarse cognitive map and plans, with reasoning shifting from spatial cues to","keywords":["goal-directedness","LLM agents","interpretability","probing","cognitive maps","grid world navigation","plan decoding","reasoning"],"falsifier":"Train the same probe on activations collected from grids where the goal and agent characters are replaced by neutral symbols (or on random feature vectors with identical coordinate conditioning); if per-class accuracy stays well above chance and localisation precision persists, the internal-map claim collapses. A simpler control: remove the (x,y) coordinate input from the probe and check whether accuracy drops to chance; if it does not, the coordinates are the carrier.","tokens_in":26257,"feed_emoji":"🗺️","tokens_out":5781,"duration_ms":62494,"temperature":0.7,"pith_summary":"The paper is trying to establish that goal-directedness in an LLM-based agent can be assessed only by combining behavioural evaluation with inspection of the model's internal representations, and it demonstrates this in a grid-navigation case study. The agent's actions are mostly consistent with the optimal policy, robust to difficulty-preserving transformations and instrumental goals, but systematically biased by goal-like but irrelevant objects such as keys. Probing hidden activations shows the agent non-linearly encodes a coarse spatial map of the grid, with approximate positions for itself and the goal, and that these representations reorganise during reasoning: broader spatial and long-horizon plan information decays while the next action becomes more decodable. The sympathetic reader should care because this supplies a concrete method to distinguish failures of goal-directedness from failures of capability and from faulty internal world models.","feed_headline":"Probing hidden states shows LLM agents navigate by a fuzzy inner map","feed_subtitle":"Behavioral checks alone can't separate wrong beliefs from wrong goals; decoded maps and plans can.","key_machinery":"Two central objects carry the argument. The first is the cognitive-map probe: an MLP classifier that, given a residual-stream activation and the (x,y) coordinates of a queried cell, predicts the cell type (agent, goal, wall, open, padding), allowing reconstruction of the model's decoded belief over the grid state. The second is a one-shot plan decoder: a Transformer decoder that predicts an entire T-step action sequence simultaneously from a fixed set of token activations, without autoregressive conditioning, ensuring that any recovered multi-step structure must already be present in the model's representations rather than manufactured by the probe's own inference. The behavioural component","core_discovery":"The central claim is that the LLM agent non-linearly encodes a coarse spatial map of the environment, preserving approximate task-relevant cues about its position and the goal location; that its actions are broadly consistent with these internal representations; and that reasoning reorganises them, shifting from broader environment structural cues toward information supporting immediate action selection. Decoding cell identities with an MLP probe reaches about 70% accuracy versus about 39% for a linear probe, showing the encoding is non-linear. Goal and agent positions are localised with high recall but blurred precision, and goal-distance is decodable with a mean absolute error of roughly t","pith_inferences":["If the probes are reliable, this framework offers a way to adjudicate capability failure versus goal-directedness failure in real deployments: an agent that acts optimally with respect to its decoded beliefs but poorly relative to the environment is rational given a faulty model, which changes how one should intervene.","The key-attraction bias suggests a general phenomenon: LLM agents may treat semantically salient objects from pretraining as goal-relevant even when the current task specification says otherwise; this could be tested in more naturalistic environments with arbitrary distractor objects.","The pre- and post-reasoning representational shift implies that reasoning partly acts as a compression from world-model to policy features. A testable extension is to vary reasoning effort and measure whether the crossing point in plan decodability moves, predicting that higher reasoning effort compresses spatial cues more aggressively.","The coarse, blurred localisation (high recall, low precision) mirrors place-field-like uncertainty in biological navigation; one could test whether the decoded map's uncertainty correlates with the agent's behavioural entropy or distance-to-goal, connecting the representation to calibration of internal beliefs."],"forward_implications":["The agent's performance scales with grid size and obstacle density and is robust to difficulty-preserving transformations, indicating that navigation behaviour is driven by task-relevant structure rather than incidental grid configurations.","The agent reliably solves instrumental key-door subtasks but is systematically attracted to goal-like reward-irrelevant objects, detouring toward a useless key in 75% of non-optimal actions and favouring the key-containing path in 67.3% of two-path trials.","Probes decode a coarse non-linear cognitive map of the environment from hidden activations, with approximate but localisable agent and goal positions, supporting the claim that goal-relevant spatial beliefs are carried in the model's internal state.","Actions are broadly consistent with decoded beliefs: agreement between the ground-truth-optimal policy and the decoded-map-optimal policy averages about 84%, and a substantial share of suboptimal actions are optimal with respect to the agent's own faulty map, attributing many failures to fuzzy world representations rather than to a lack of goal-directedness.","Reasoning re-weights internal representations: pre-reasoning activations support longer-horizon plan decoding (e.g., 13% vs 9% prefix accuracy at N=4), while post-reasoning activations boost next-step decodability (53.9% vs 40.3% at N=1), indicating a shift from environment-structural cues to immediate action selection."],"fun_headline_variants":["LLM agents use fuzzy inner maps, not just behavior, to navigate","Non-linear probing reveals LLM agents' coarse spatial encoding","Behavior alone can't tell wrong beliefs from wrong goals","LLM agents reason by shifting from spatial maps to action cues"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The cognitive-map and consistency claims assume that probes trained on ground-truth labels with (x,y) coordinate conditioning recover the model's internal beliefs, rather than exploiting shortcuts such as the grid text literally present in the input, coordinate priors, or probe capacity; no control probes or ablations removing the coordinate conditioning are reported.","fun_headline_variants_meta":{"raw":{"variants":["LLM agents use fuzzy inner maps, not just behavior, to navigate","Non-linear probing reveals LLM agents' coarse spatial encoding","Behavior alone can't tell wrong beliefs from wrong goals","LLM agents reason by shifting from spatial maps to action cues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1207,"prompt_tokens":705,"completion_tokens":502,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":432}},"tokens_in":449,"tokens_out":502,"duration_ms":6112,"temperature":1.0,"reasoning_tokens":432,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T03:04:46.025473+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same probe on activations collected from grids where the goal and agent characters are replaced by neutral symbols (or on random feature vectors with identical coordinate conditioning); if per-class accuracy stays well above chance and localisation precision persists, the internal-map claim collapses. A simpler control: remove the (x,y) coordinate input from the probe and check whether accuracy drops to chance; if it does not, the coordinates are the carrier.","supporting_citations":[],"review_version":1}