{"id":"b45e934e-1518-4b52-8065-9a602f35e653","arxiv_id":"2412.00944","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Bilinear convolution layers can be decomposed into eigenfilters that separate channel and spatial information, enabling causal checks of concept probes in a maze-solving RL agent.","lead":"This paper replaces the nonlinear parts of a neural network with bilinear versions, making the network's internal filters easier to analyze mathematically. It tests these 'Bimpala' agents on maze and other game environments, then decomposes their weights into interpretable filters that track objects like cheese.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.4 contradicts the key causal-validation claim: it says the top-1 eigenfilter ablation fails to solve the maze, then immediately says the top-1 ablation does solve it. Until this conflict is resolved, the sufficiency of a single action eigenvector is unsupported.","rationale":"The paper's central claim is that bilinear convolutions can be decomposed into eigenfilters and that ablating all but the top action vector preserves maze-solving ability, thereby providing causal validation of concept probes. The most load-bearing piece of evidence for this claim is the ablation study in Section 4.4, and that section is internally contradictory: the top-1 ablation is first reported as failing (max steps reached) and then as succeeding (slower, with wrong moves). The reader's verdict was CONDITIONAL, and this concern supports that verdict, so UNCHANGED is appropriate. I agree only partially with the reader's identified weakest assumption: the bias-omission point is real (Section 2.1 omits biases 'for brevity,' and Section 3.1's Bsym includes only U and V), and it would invalidate the exactness of the decomposition if the trained layers contain biases, but it is speculative without code. The ablation contradiction, by contrast, is present in the manuscript text itself and is immediately checkable. The recommended concrete test is a focused reproduction of the k=1 ablation with per-seed results; this will determine whether the contradiction is a typo (k=0 vs k=1) or a genuine failure of the top-1 claim. I am not recommending REJECT because the issue is addressable and the mathematical decomposition may still be valid; I am also not recommending ACCEPT because the paper as written does not support its own central empirical claim.","tokens_in":7453,"tokens_out":7044,"duration_ms":63546,"concrete_test":"Rerun the k=1 ablation with the released code and checkpoints on the same 20 seeded environments, and report per-seed success rate and mean steps-to-solve for three conditions: (a) retaining the top 1 eigenfilter per output channel in the final convolution layer, (b) retaining the top 1 action eigenvector in the final FC layer, and (c) a true k=0 full-ablation control. If (a) or (b) gives near-100% success, the Section 4.4 'fails to solve' sentence is a typo and the claim survives; if either gives max-steps failure, Figure 8's claim is unsupported. Publishing the per-seed numbers also resolves whether the contradiction stems from a k=0 versus k=1 labeling error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central evidence for causal validation is the ablation sufficiency claim. Section 4.3.4 states that 'ablating all but the top action vector is sufficient to preserve maze-solving ability' (Figure 8). Section 4.4, however, contains two adjacent contradictory sentences about the same k=1 setting: 'With k = 1, all convolutions are ablated, and the agent fails to solve the maze (as it reaches the max steps).' and 'With k = 1 eigenfilters, we find that the agent does solve the maze, but it takes relatively more time compared to average as it makes some wrong moves.' Both sentences cannot describe the same intervention. The first also conflates k=1 with 'all convolutions are ablated' (a k=0 condition), suggesting a labeling error, while the second contradicts the claimed failure. Because the paper's headline interpretability result depends on a single retained component being sufficient, this internal inconsistency means the empirical support for the causal claim is currently self-conflicting. The bias-omission issue noted in Section 2.1 further threatens the exactness of the decomposition, but the ablation contradiction alone is sufficient to withhold full acceptance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes replacing ReLU convolutions with bilinear convolutions, BConv2D(x, U, V) = (x ∗ U) ⊙ (x ∗ V), and bilinear fully-connected layers, yielding a variant of IMPALA called Bimpala. It claims that Bimpala matches or occasionally outperforms a simplified ReLU IMPALA on ProcGen environments (Maze, Heist, Plunder, Dodgeball). The central interpretability contribution is a weight-based decomposition: for each output channel, the bilinear convolution is rewritten as a symmetric quadratic form Bsym of dimension nK², whose eigendecomposition yields \"eigenfilters\" that can be visualized and ablated. The paper then proposes a protocol that trains a linear probe for a concept (cheese location), applies SVD to the probe weights, uses the top singular vectors as output directions for the preceding BConv layer, and decomposes the layer along those directions. Experiments report probe F1 scores, eigenfilter spectra, and ablation results intended to show that a small number of eigenfilters suffices to preserve maze-solving behavior, with a single action eigenvector claimed sufficient in Section 4.3.4.","tokens_in":7658,"tokens_out":4914,"duration_ms":44068,"significance":"Conditional on the empirical claims being correct, the paper offers a concrete template for mechanism-level, interventional interpretability of an RL agent: the decomposition is a closed-form function of the trained weights, and the ablations directly modify the network rather than only analyzing activations. The linear algebra in Section 3.1 is explicit and checkable, and the symmetrization argument that Bsym agrees with B on all inputs appears sound. The paper also honestly states limitations, including the difficulty of data-independent unit interpretation and the omission of batch norm, dropout, and pooling. However, the significance is currently weakened by three load-bearing issues: the training comparison lacks error bars or seed counts, the ablation section contains a direct self-contradiction about the k=1 condition, and the decomposition omits bias terms without evidence that biases are negligible in the trained Bimpala model. These issues must be resolved before the causal claims can be evaluated.","major_comments":[{"comment":"The final paragraph of Section 4.4 contains two adjacent sentences that directly contradict each other: \"With k = 1, all convolutions are ablated, and the agent fails to solve the maze (as it reaches the max steps)\" and \"With k = 1 eigenfilters, we find that the agent does solve the maze, but it takes relatively more time compared to average as it makes some wrong moves.\" Both sentences refer to the same k=1 condition, so they cannot both be correct. Additionally, the first sentence mislabels k=1 as \"all convolutions are ablated,\" which would be the k=0 condition. This contradiction is load-bearing because the paper's headline sufficiency claim (Section 4.3.4 and Figure 8, \"keeping just 1 eigenvector for each output action is enough to preserve near 100% success rate\") is contradicted by the statement in Section 4.4 that 2 eigenfilters per channel are needed (Figure 10). Please clarify the exact intervention at k=1, report the success-rate curves for k=0,1,2, and reconcile the discrepancy between Figure 8 and Figure 10.","section":"Section 4.4"},{"comment":"The decomposition constructs Bsym exclusively from the U and V kernels, with no bias term, and BConv2D in Eq. (2) is defined without bias. However, Section 2.1 states that biases are omitted only \"for brevity\" in the fully-connected case, which implies that the actual Bimpala model used in Section 4 includes bias parameters. Standard convolutional layers in RL architectures indeed include biases. If biases are present and non-negligible, the layer's input-output map is (x ∗ U + b_U) ⊙ (x ∗ V + b_V), which is not equal to the quadratic form x^T Bsym x. Consequently, the eigenfilter spectra (Figure 6), the probe decomposition (Section 4.2), and the ablations (Section 4.4) would not describe the true computation of the layer. Please either confirm that the trained Bimpala BConv layers have no bias parameters, or extend the decomposition to include the affine terms and demonstrate that the conclusions are unchanged.","section":"Sections 2.1 and 3.1"},{"comment":"The claim that \"Bimpala matches and occasionally outperforms IMPALA\" is a central motivation for the architecture, but Figure 4 shows learning curves without error bars, confidence intervals, or a reported number of seeds. Without this information, the \"comparable performance\" claim could be within run-to-run noise. Please report mean and standard deviation (or interquartile range) over at least 3–5 independent seeds for each environment, and state the number of seeds used in the figure and in the main text.","section":"Section 4.1.1 and Figure 4"},{"comment":"The proposed protocol is partly self-referential: a linear probe is trained on the activations, its SVD directions are used to select the output directions for the BConv decomposition, and then the probe's F1 score (Table 1) and the resulting eigenfilter activations are used as evidence that the decomposition captures the concept. Since the probe is fit to the same activations it later selects, high probe accuracy is to be expected. The independent evidence must come from the behavioral ablations (maze-solving success), not from probe accuracy alone. Please report the network's behavior on a held-out set of mazes that were not used for probe training or eigenfilter selection, and explicitly state how the probe and the decomposition are evaluated out-of-sample.","section":"Section 4.2 protocol"}],"minor_comments":[{"comment":"The notation is confusing when K is used both for the kernel width and as a summation index in Eq. (3); please use K1 and K2 or explicit bounds such as -ℓ to ℓ for clarity.","section":"Section 2.2"},{"comment":"The proof that Bsym agrees with B on all inputs references \"the respective red and blue terms,\" but the colors are not visible in a monochrome printout; please recast this explanation without color references.","section":"Section 3.1"},{"comment":"The notation Xj is used for both the input patch and its flattened vector; please introduce separate symbols (e.g., X̌_j for the flattened patch) to avoid ambiguity.","section":"Section 3.1"},{"comment":"The figure would benefit from explicit axis labels and a legend that identifies ReLU vs. Bilinear curves and the number of seeds; currently the caption is the only guide.","section":"Figure 4"},{"comment":"The F1 score of 2.73% for the Value FC layer is surprisingly low; please discuss why the value head does not preserve cheese-position information, or check whether this entry is an error.","section":"Section 4.3.1 and Table 1"},{"comment":"The text says the action spectrum has \"one very large positive eigenvalue,\" but the figure does not show the negative eigenvalues or the full eigenvalue distribution; please clarify the range of the y-axis and whether negative eigenvalues are present.","section":"Section 4.3.4 and Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is under-specified for reproducibility: no architecture hyperparameters, no optimizer details, no seed counts, no code or model-release link. Given that the central interpretability claim depends on a single trained Bimpala checkpoint, the authors should be asked to provide these details and ideally release the model. The internal contradiction in Section 4.4 is the most serious issue and must be resolved before the paper can be considered for publication; the bias-omission concern is also important and needs a concrete response. The methodological idea is promising and within the scope of the journal, but the current empirical support is too thin and self-conflicting to accept as-is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about arXiv:2412.00944. First, the mathematical core is real: the symmetrized block-matrix decomposition of a bilinear convolution into eigenfilters is a correct and useful extension of Pearce et al.'s bilinear MLP work, and the SVD trick for separating channel and spatial axes is a clever addition. Second, the experimental support for the causal claims is not yet there — Section 4.4 literally contradicts itself about the k=1 ablation, saying in one sentence that the agent fails and in the next that it solves the maze with extra steps. Until that is sorted, treat the sufficiency results with caution.\n\nWhat is genuinely new: the Bsym construction, the SVD-based channel/spatial separation, and the protocol that connects probe directions to filter decompositions and then tests them by ablation. The equivalence of Bsym and B on all inputs is proved, and the spectral decomposition is parameter-free, so the algebra is solid. The paper also deserves credit for being honest about its preliminary nature and for including a clear limitations section.\n\nNow the soft spots, in proportion. The bias omission is a real gap, not a nitpick: BConv2D is defined without bias, but the paper says biases are omitted 'for brevity' elsewhere, so it's unclear whether the actual trained layers have them. If they do, the quadratic form misses a linear term and the eigenfilter analysis does not describe the layers' true input-output map. The training comparison in Figure 4 has no error bars, which for RL is a significant omission. And the ablation section is internally inconsistent: after saying 'After knocking out the entire layer, maze solving ability is preserved,' it later says with k=1 'all convolutions are ablated' (which would be k=0) and then gives two contradictory outcomes for the same intervention. The stress-test note is right about this; it is a real problem for the causal-validation narrative.\n\nOne clarification: the 'top action vector' ablation in Figure 8 targets a different object (the final FC layer) than the eigenfilter ablations in Figures 9-10 (BConv layers). The paper should keep those claims separate. The circularity concern about probes is mild: the ablations are behavioral, so they do test causal relevance, though the selection of components is probe-driven.\n\nWho should read this: people working on mechanistic interpretability of RL vision models, and anyone interested in bilinear layers as a transparency tool. The algebra alone is worth taking seriously. As a referee, I would engage: the method is new and the math checks out, but the experimental reporting needs revision. I would not accept the causal claims in the current draft.","headline":"The algebra is correct and the protocol is promising, but the ablation section contradicts itself on the key k=1 result, so the causal claims are not yet backed by the experiments.","tokens_in":8225,"tokens_out":4800,"would_cite":true,"duration_ms":42164,"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":"Replacing ReLU with bilinear convolutions yields an RL agent whose maze-solving mechanisms can be decomposed into eigenfilters and causally validated.","keywords":["bilinear convolutions","reinforcement learning interpretability","eigenfilter decomposition","concept probes","causal validation","Bimpala","ProcGen","maze solving"],"falsifier":"Open the trained Bimpala model and check whether its BConv2D layers have nonzero bias parameters. If they do, compare each layer's actual output with the bias-free quadratic form $x^\\top B_{\\mathrm{sym}} x$ on held-out maze patches; if the difference changes which eigenfilters rank highest or alters the ablation outcomes, the decomposition is incomplete. Retraining with explicit zero biases and rerunning the ablations would settle it.","tokens_in":7229,"feed_emoji":"🧀","tokens_out":8339,"duration_ms":64825,"temperature":0.7,"pith_summary":"This paper argues that swapping the ReLU nonlinearity in convolutional networks for a bilinear gating operation --- replacing Conv2D+ReLU with $(x \\ast U) \\odot (x \\ast V)$ --- gives reinforcement learning agents that perform on par with a standard IMPALA baseline on ProcGen maze and related tasks, while making the network's internals analytically tractable. Because the bilinear layer's local input-output map is a quadratic form, it can be decomposed into an orthonormal basis of eigenfilters with signed eigenvalues. The paper then shows how to combine this decomposition with a linear concept probe: SVD of the probe weights separates channel from spatial dimensions, and the resulting singular channels pick out eigenfilters whose activations implement the probed concept. Applied to a maze-solving agent, the method finds a positive/negative eigenfilter pair that acts as a cheese detector, and ablations show that keeping only the top action eigenvector preserves maze-solving success. The upshot is a path from correlational probing to causal, mechanism-level validation in an RL agent.","feed_headline":"Bilinear convs expose an RL agent's cheese detector","feed_subtitle":"Maze-solving Bimpala matches IMPALA's scores; its filters decompose into eigenfilters for probing and ablation.","key_machinery":"The load-bearing object is the symmetric bilinear matrix $B_{\\mathrm{sym}}$ of shape $[nK^2, nK^2]$ for each output channel, obtained by summing outer products $U_j^\\top V_k$ over channel pairs and symmetrizing; it turns the gated convolution into one quadratic form. From it, the protocol builds eigenfilters: for an output direction $u$ (a probe-derived singular vector or an action direction), $Q_u = u B_{\\mathrm{sym}}$ is a symmetric matrix whose eigendecomposition $Q_u = F^\\top \\Lambda F$ gives an orthonormal basis of filters $f^i$, each with eigenvalue $\\lambda^i$, and the layer's response decomposes as $\\sum_i \\lambda^i (f^i \\ast X)^2$. SVD of a probe's weight matrix reshaped to $[C, wh]$ separates channel-space singular vectors $s_j$ from spatial components, so the probe's total score is a signed, weighted sum of eigenfilter activations. These eigenfilters are the units that ablation studies remove or retain, which is what makes the interpretability claims causal rather than correlational.","core_discovery":"The central claim is that a bilinear convolution layer, defined by $\\mathrm{BConv2D}(x,U,V) = (x \\ast U) \\odot (x \\ast V)$, has a closed-form decomposition: for each output channel, the layer's action on any local patch is exactly $x^\\top B_{\\mathrm{sym}} x$, where $B_{\\mathrm{sym}}$ is the symmetric matrix built from the U and V kernels. Eigendecomposing the quadratic form $Q_u = u B_{\\mathrm{sym}}$ for a chosen output direction $u$ yields $nK^2$ eigenfilters $f^i$ with real eigenvalues $\\lambda^i$, and the layer's contribution to $u$ is $\\sum_i \\lambda^i (f^i \\ast X)^2$. Combined with an SVD of a linear probe's weights into channel singular vectors $s_j$, the probe's score becomes $\\sum_{j,i} (s_j \\lambda^i_{u_j})(f^i_{u_j} \\ast X)^2$, so each eigenfilter's relevance is quantified by a signed product of a singular value and an eigenvalue. The paper demonstrates this machinery on a Bimpala agent trained in ProcGen Maze: probes on residual-block activations reach roughly 99--100\\% F1 for cheese position, the top singular channel explains 30\\% of probe variance, the top positive and negative eigenfilters form a cheese detector, and ablating all but one action-direction eigenvector preserves maze-solving.","pith_inferences":["If the trained convolution layers turn out to include non-negligible bias parameters, the decomposition should be re-derived with a bias term; otherwise the eigenfilter that gets ablated is not the layer's true computation.","The same SVD-plus-eigendecomposition recipe should transfer to any bilinear layer stack, suggesting a general workflow: train with bilinear gates, probe a concept, decompose the preceding layer toward the probe's top channel, and ablate the ranked eigenfilters to test causality.","The finding that the top negative eigenfilter downweights non-cheese patterns without erasing cheese activation suggests a subtraction mechanism; a testable extension is to check whether this paired positive/negative structure appears for other concepts and environments."],"forward_implications":["A bilinear RL policy can match a standard ReLU IMPALA policy on ProcGen environments, so analytic interpretability does not come at an obvious performance cost.","Any linear concept probe on a bilinear layer can be rewritten as a signed weighted sum of eigenfilter activations, turning a correlational probe into a list of concrete, ablatable computational units.","Keeping only the top action-direction eigenvector in the final layer is enough to preserve near-perfect maze-solving, implying the policy's decision-relevant computation is concentrated in a low-rank subspace.","In the standard channel basis the eigenfilter spectrum has only two nonzero eigenvalues, while the probe-aligned singular basis has a nondegenerate spectrum, so the probe direction is what makes the mechanism visible."],"supporting_citations":[{"why":"Supplies the eigendecomposition method for bilinear layers that the paper adapts from MLPs to convolutions.","marker":"(Pearce et al., 2024)"},{"why":"Motivates bilinear layers as more interpretable than standard nonlinear MLPs.","marker":"(Sharkey, 2023)"},{"why":"Defines the IMPALA architecture that the simplified baseline and Bimpala are adapted from.","marker":"(Espeholt et al., 2018)"},{"why":"Provides the ProcGen environments used for the performance comparison and maze experiments.","marker":"(Cobbe et al., 2020)"},{"why":"Supplies the gated linear unit notation and formulation that the bilinear convolution and fully connected layers build on.","marker":"(Shazeer, 2020)"}],"fun_headline_variants":["Bilinear convs decompose RL agents into eigenfilters","RL interpretability: bilinear convs yield closed-form eigenfilters","Cheese-tracking RL agent decoded via bilinear conv decomposition","Closed-form bilinear decomposition exposes RL agent's cheese detector","Maze-solving RL agent's cheese sensor revealed by bilinear eigenfilters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each trained bilinear convolution layer is fully described by the quadratic form $x^\\top B_{\\mathrm{sym}} x$ built from the U and V kernels alone, with no bias term and no other nonlinearity; the paper omits biases 'for brevity' in its equations, so if the actual trained layers contain non-negligible biases, the eigenfilter spectra and ablations describe only part of the layer's true input-output map.","fun_headline_variants_meta":{"raw":{"variants":["Bilinear convs decompose RL agents into eigenfilters","RL interpretability: bilinear convs yield closed-form eigenfilters","Cheese-tracking RL agent decoded via bilinear conv decomposition","Closed-form bilinear decomposition exposes RL agent's cheese detector","Maze-solving RL agent's cheese sensor revealed by bilinear eigenfilters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1636,"prompt_tokens":1012,"completion_tokens":624,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":535}},"tokens_in":628,"tokens_out":624,"duration_ms":5388,"temperature":1.0,"reasoning_tokens":535,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:49:45.175720+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Open the trained Bimpala model and check whether its BConv2D layers have nonzero bias parameters. If they do, compare each layer's actual output with the bias-free quadratic form $x^\\top B_{\\mathrm{sym}} x$ on held-out maze patches; if the difference changes which eigenfilters rank highest or alters the ablation outcomes, the decomposition is incomplete. Retraining with explicit zero biases and rerunning the ablations would settle it.","supporting_citations":[{"cited_title":"A technical note on bilinear layers for interpretability, 2023","cited_arxiv_id":null,"evidence_quote":"Motivates bilinear layers as more interpretable than standard nonlinear MLPs."}],"review_version":1}