{"id":"f74d2d1b-b5e3-4964-b257-d9a724809814","arxiv_id":"2502.04052","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ReMeDe Trees are hard, axis-aligned decision trees with a learned internal memory, trained by backpropagation through time, achieving perfect accuracy on synthetic delayed-sign and sign-memory tasks.","lead":"ReMeDe Trees add a small internal memory to decision trees, letting them learn from time sequences instead of relying on hand-built lag features. The authors show on five synthetic tasks that these recurrent trees can be trained with backpropagation through time and match LSTM accuracy while staying compact and axis-aligned.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Straight-through gradient bias is only validated on delays up to 7; the claim that BPTT reliably trains ReMeDe Trees for long-term memory is unsupported.","rationale":"The reader's weakest assumption is exactly the one I would select. The architecture is coherent if the routing at time t uses m_{t-1} rather than m_t, and the forward dynamics are otherwise well-specified. The decisive question is whether the straight-through gradients used through the hard splits and the hard write gate remain informative when the recurrence must bridge many timesteps. All five tasks use delays of at most 7, so they only exercise short-lag memory. BPTT through a rounded tree is a discontinuous optimization problem; ST is a heuristic that can fail silently on longer horizons. The paper offers no formal argument and no code or described validation split, so the perfect accuracies cannot be independently confirmed. A longer-delay experiment or a standard long-sequence benchmark would settle whether the training signal survives. If it does not, the contribution reduces to a short-horizon proof of concept, not the general recurrent-DT training claimed. I therefore keep the reader's CONDITIONAL verdict and see no basis to reject or fully accept.","tokens_in":10154,"tokens_out":14559,"duration_ms":162032,"concrete_test":"Train ReMeDe on the Delayed Sign Retrieval task with variable delays sampled from [50,100] (or on sequential MNIST with 784 steps), using the same tree depth 6 and memory size 5, and compare test accuracy against LSTM under the same hyperparameter selection procedure. If ReMeDe degrades to chance or substantially below LSTM, the ST gradients do not support long-term dependency learning, and the central claim should be restricted to short-lag proof of concept.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that hard, axis-aligned ReMeDe Trees can be trained end-to-end by BPTT and can store and recall information across variable delays. This rests on the straight-through (ST) estimator used through the rounded split decisions (Eq. 4) and the hard memory gate (Eq. 11). ST replaces the derivative of the rounding operation with the identity, which is biased whenever the underlying logit is not exactly at a rounding boundary. In a recurrent setting, this bias is injected at every timestep and can accumulate through the unfolded memory state, so training a model that must hold information for many steps is not guaranteed to work. The only evidence for the estimator is five synthetic tasks whose maximum delay is 7 (fixed delay 5; variable delays 3-7). These are all variants of the same sign-retrieval problem, and the paper does not report a validation split for the Optuna searches, so the reported 1.000 accuracies are not clearly unbiased estimates of generalization. Consequently, the broad claim that recurrent DTs overcome the limitations of fixed-size memory windows is not established; the experiments demonstrate only short-lag sign memory.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ReMeDe Trees, a recurrent decision-tree architecture that augments a hard, axis-aligned GradTree with a continuous internal memory vector. At each time step, the tree reads the input and the memory, produces an output from the selected leaf, and updates the memory using leaf-prescribed gating and a tanh-transformed linear input mapping. Training is end-to-end via backpropagation through time, with straight-through estimators used through the discrete split decisions and hard memory gates. The authors report perfect test accuracy on five synthetic delayed-sign-retrieval and sign-memory tasks, matching an LSTM baseline, and they visualize a learned minimal tree that stores and recalls a sign across a delay.","tokens_in":10362,"tokens_out":5123,"duration_ms":54848,"significance":"If taken at face value, the paper provides a useful proof of concept: it shows that a fully discrete-routing tree with a continuous memory can be trained by gradient descent and BPTT, without soft probabilistic routing or post-hoc binarization. The hand-traced example and the internal consistency across five trials (zero variance, matching LSTM) support the claim that the learned trees do implement the intended memory operations. The contribution is nevertheless narrow: the experiments are synthetic, short-lag, and not compared against a fixed-window baseline, so the paper does not yet establish practical advantages for long-term sequence modeling. The authors are appropriately careful in calling this a proof of concept, but the abstract and conclusion use language that goes beyond the evidence.","major_comments":[{"comment":"The paper's central claim is that ReMeDe Trees 'learn long-term dependencies' and 'overcome the limitations of fixed-size memory windows', but the experiments use a fixed delay of 5 and a variable delay range [3,7] for PoC1-4, and PoC5 does not state its block or delay lengths. A maximum lag of 7 is not 'long-term' in any meaningful sense. Because the straight-through estimator's bias is injected at every time step and accumulates through the unfolded memory, this short-lag evidence does not support the long-horizon claim. Please either add experiments with substantially longer lags (e.g., 50-100 steps) or restrict the abstract and conclusion to the tested short-lag regime.","section":"Sec. 4.2, Table 1; Abstract; Sec. 6"},{"comment":"The hyperparameter selection protocol is under-specified. Optuna is run for 60 trials to select learning rates, but the text never states whether a validation split is used; if the test set was used for selection, the reported 1.000 accuracies are not unbiased estimates of generalization. Please specify the train/validation/test split and report test performance for the model selected on validation.","section":"Sec. 4.2"},{"comment":"The motivating contrast with NARX/memory-window models is never tested. The authors state that 'it is clear' that a fixed-lookback model cannot solve the tasks, but this is an assertion; a concrete comparison with, e.g., GradTree using lag features or a standard NARX baseline would substantiate the claim that the recurrent memory is the source of the performance. This is particularly important because all five tasks are constructed so that a small memory suffices, making the successful learning less surprising.","section":"Sec. 4.2"}],"minor_comments":[{"comment":"The Sign Memory task does not specify the values of block length l and zero-block delay d used in the experiments; these values are needed to interpret Table 1 and to judge the actual lag.","section":"Sec. 4.1, PoC5"},{"comment":"The notation ⌊ψ_g(c_j)⌋ ψ(W^x_j x_t) is ambiguous about whether the rounding is applied componentwise and whether c_j is a scalar or a vector; please define all dimensions explicitly.","section":"Eq. (11)"},{"comment":"The pruning procedure is described only as 'removing all redundant paths'; please specify the pruning rule and report standard deviations for the tree sizes.","section":"Table 2"},{"comment":"The equations in the figure repeat computed values and mix variables in a way that is hard to follow; a cleaner legend or a separate table of the traced values would improve readability.","section":"Figure 2"},{"comment":"The sentence 'We evaluate two recurrent architecture on our datasets' contains a grammar error; it should be 'two recurrent architectures'.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about being a proof of concept, but the broad language in the abstract and conclusion exceeds the evidence. I would accept a revised version that either adds a longer-delay experiment and a NARX baseline or carefully limits the claims. The hand-traced example is a strength and should be retained. If code is available, including it would strengthen reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this paper introduces a genuinely new model class—hard, axis-aligned decision trees with a continuous internal memory, trained end-to-end by BPTT. That is not a minor tweak; it is a different way to make trees recurrent, and the related-work section is honest about what came before. The hand-traced example in Figure 2 is the strongest part: it shows a learned tree actually storing the sign of the first input and recalling it when the trigger appears. That is real evidence the mechanism does what the authors claim.\n\nThe writeup is clear and the architecture is well specified. The paper is also appropriately modest in places—it calls itself a proof of concept, and it does not claim state-of-the-art results on real benchmarks. The main soft spots are empirical. First, all five tasks are variants of the same sign-retrieval problem with delays capped at 7; that does not support the broader claim that ReMeDe Trees overcome fixed-memory-window limits in any meaningful long-term sense. Second, the Optuna runs tune learning rates on what appears to be the test set, with no validation split described, so the perfect 1.000 accuracies are not clearly unbiased generalization estimates. Third, there is no NARX baseline, which matters because the introduction frames the whole contribution around beating fixed-window approaches. The authors justify the omission, but a simple NARX comparison would have made the point concrete. Fourth, no code or data are provided, which makes the results hard to check.\n\nThe straight-through estimator concern is real but not fatal here. The paper shows the estimator works for delays up to 7, and for a proof of concept that is acceptable. The worry that ST bias accumulates over longer horizons is a research question, not a refutation of the presented experiments.\n\nOverall: the central claim—that recurrent DTs with hard splits and internal memory can be trained by BPTT on these tasks—is supported. The broader significance is not yet established, but the paper is worth engaging with. Bring it to a reading group if you care about interpretable sequence models; I would cite it as the first work to train a memory-augmented recurrent DT with BPTT. It deserves a real referee, not a desk reject, and with more thorough evaluation it could become a solid contribution.","headline":"ReMeDe Trees are a genuinely new recurrent DT architecture with a clear write-up; the proof of concept is real but thin, so it deserves conditional acceptance, not a desk reject.","tokens_in":10908,"tokens_out":1696,"would_cite":true,"duration_ms":20124,"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":"The paper claims ReMeDe Trees, hard axis-aligned decision trees with a gated internal memory, can be trained end-to-end by backpropagation through time and reach perfect test accuracy on five synthetic memory tasks, matching LSTM baselines.","keywords":["recurrent decision trees","internal memory","backpropagation through time","gradient-based decision trees","straight-through estimator","hard axis-aligned splits","long-term dependencies","sequence modeling"],"falsifier":"Train a ReMeDe tree on the sign-retrieval task and inspect its memory state after the first timestep; if the stored value does not carry the sign of the initial input while the tree still predicts the correct sign at the trigger, then the model is solving the task without using its memory as claimed.","tokens_in":9914,"feed_emoji":"🌳","tokens_out":7514,"duration_ms":70944,"temperature":0.7,"pith_summary":"This paper introduces ReMeDe Trees, a decision tree with a continuous internal memory that internal nodes can read and leaf nodes can update. The authors claim that, for the first time, such a hard, axis-aligned recurrent decision tree can be trained end-to-end using backpropagation through time, with the discrete split and gate decisions made differentiable by a straight-through estimator. On five synthetic tasks that require storing an initial value and recalling it when a later trigger appears, the model reaches perfect test accuracy, matching LSTM baselines and beating window-based and naive baselines. The significance is that a tree-based model, not a neural network, can handle long-term sequential dependencies while keeping the interpretable, axis-aligned structure of classical decision trees.","feed_headline":"Decision trees with memory match LSTM on five time-series tasks","feed_subtitle":"A hard, axis-aligned tree with a hidden memory state trains end-to-end by backpropagation through time.","key_machinery":"The mechanism is the recurrent tree function itself. ReMeDe extends GradTree's arithmetic formulation $y = \\sum_l \\lambda_l L(x|l,\\tau,\\iota)$ to the joint input–memory space: the same tree computes both the output and the next memory state, with internal nodes allowed to split on any coordinate of $x_t$ or $m_{t-1}$. The memory update is an additive, gated write: a leaf's gate parameter $c_j$ is passed through a sigmoid and rounded to 0 or 1, then multiplied by $\\tanh(W_j^x x_t)$; the straight-through estimator lets gradients flow through the rounding both in the splits and in this gate. This gives the tree a hard read/write memory whose contents condition future routing, which is what lets a fixed-depth tree solve tasks with variable delays.","core_discovery":"The central claim is that Recurrent Memory Decision (ReMeDe) Trees, a single hard, axis-aligned decision tree operating on the augmented input $\\tilde{x}_t = (x_t, m_{t-1})$, where $m_t$ is a continuous memory state, can be learned by gradient descent with backpropagation through time. Each leaf prescribes both an output and a gated memory update $m_t = m_{t-1} + \\lfloor \\psi_g(c_j)\\rceil \\psi(W_j^x x_t)$, with the rounding handled by a straight-through estimator in the backward pass. The paper reports perfect average test accuracy on all five proof-of-concept datasets, equal to the LSTM baseline, and shows a pruned tree with only 22 to 44 nodes that stores the sign of the first input and retrieves it at the trigger. The authors state this is the first approach to learn a memory-augmented recurrent decision tree using backpropagation through time.","pith_inferences":["Beyond the paper, the same read/write memory idea could be lifted into any differentiable tree ensemble, so the key contribution is not the specific tree but the proof that hard threshold logic plus a memory state is trainable.","The straight-through gradient is the transfer risk; a natural stress test would be to scale sequence length and input dimension and observe whether ReMeDe retains the LSTM-matching accuracy reported here.","Because the splits are axis-aligned, a learned ReMeDe tree can in principle be audited as a state machine: tracing which memory thresholds cause writes and reads gives a human-readable account of the model's timing logic, unlike an RNN's distributed state."],"forward_implications":["ReMeDe Trees can solve tasks that require storing information across a delay longer than any fixed input window, because the memory state compresses the past rather than truncating it.","The learned recurrent trees stay compact: after pruning, the proof-of-concept trees have between 20 and 44 nodes, which the authors argue supports interpretability and verifiability.","The training recipe transfers the standard neural toolkit, BPTT, gating, and straight-through gradients, to hard, axis-aligned trees, so the approach is not limited to the zero-order output used here.","Because the base tree is the GradTree model, the method can be combined with tree ensembles, which the authors name as a direct next step."],"supporting_citations":[{"why":"Supplies the GradTree arithmetic formulation and straight-through gradient training that ReMeDe extends with memory.","marker":"(Marton et al., 2024a)"},{"why":"Provides backpropagation through time, the training algorithm used for the recurrent tree.","marker":"(Werbos, 1990)"},{"why":"Defines the LSTM used as the comparison baseline and the gating concept the memory update adapts.","marker":"(Schmidhuber et al., 1997)"},{"why":"Grounds the straight-through estimator used to differentiate through rounded splits and memory gates.","marker":"(Yin et al., 2019)"},{"why":"Identifies the vanishing-gradient problem motivating the gated memory structure.","marker":"(Hochreiter, 1998)"}],"fun_headline_variants":["Memory-augmented decision trees match LSTM on benchmarks","Recurrent trees with hidden memory learn via backprop","Decision trees that remember get LSTM-level accuracy","Trees with memory match LSTM on time-series tasks","Hard trees with memory train end-to-end via backprop"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rides on the assumption that the approximated gradients used to bypass the tree's rounding steps are informative enough to train the memory, and that assumption is only tested on five small synthetic tasks.","fun_headline_variants_meta":{"raw":{"variants":["Memory-augmented decision trees match LSTM on benchmarks","Recurrent trees with hidden memory learn via backprop","Decision trees that remember get LSTM-level accuracy","Trees with memory match LSTM on time-series tasks","Hard trees with memory train end-to-end via backprop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001307,"raw_usage":{"total_tokens":5308,"prompt_tokens":902,"completion_tokens":4406,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":4329}},"tokens_in":518,"tokens_out":4406,"duration_ms":31836,"temperature":1.0,"reasoning_tokens":4329,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T23:43:13.215125+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a ReMeDe tree on the sign-retrieval task and inspect its memory state after the first timestep; if the stored value does not carry the sign of the initial input while the tree still predicts the correct sign at the trigger, then the model is solving the task without using its memory as claimed.","supporting_citations":[{"cited_title":"Backpropagation through time: what it does and how to do it","cited_arxiv_id":null,"evidence_quote":"Provides backpropagation through time, the training algorithm used for the recurrent tree."},{"cited_title":"Long short-term memory","cited_arxiv_id":null,"evidence_quote":"Defines the LSTM used as the comparison baseline and the gating concept the memory update adapts."},{"cited_title":"The vanishing gradient problem during learning recurrent neural nets and problem solutions","cited_arxiv_id":null,"evidence_quote":"Identifies the vanishing-gradient problem motivating the gated memory structure."}],"review_version":1}