{"id":"10203182-8f19-4a10-9abc-20067f90004d","arxiv_id":"2509.00622","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"BALM-TSF combines a statistical-prompt text branch with a patch-based time series branch, using scaling plus contrastive alignment to balance the two modalities, improving long-term and few-shot forecasting on five of six benchmarks.","lead":"This paper presents BALM-TSF, a time series forecasting model that feeds a frozen LLM only a short text summary of the series, then aligns the resulting text embedding with a patch-based time series encoder using scaling and contrastive learning. The method reports lower forecasting error than previous LLM-based baselines on most standard benchmarks while training only about one million parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation removes or replaces the frozen LLM text branch, so the paper never isolates whether gains come from LLM-based multimodal alignment or just from the dual-branch statistical-prompt design.","rationale":"I read the paper as claiming that a frozen LLM plus balanced scaling/contrastive alignment yields state-of-the-art results by leveraging textual semantics. For this to be true, the LLM text branch must be doing non-trivial work beyond re-encoding summary statistics. The ablations in Figure 4 remove alignment components and learnable prompt but never the LLM itself, nor a non-LLM encoder. This is the least secure link in the causal story: if a from-scratch transformer on the same statistics matches BALM-TSF, the proposed mechanism is not 'LLM-based multimodal alignment' but a dual-branch architecture with a useful statistical prompt. The paper's own Weather discussion concedes the text branch provides limited context, which makes the absence of this control more salient. The reader's truncation concern is real but secondary; even if Eq. 7 is suboptimal on other datasets, the Exchange/ETT results suggest the model has some value. The missing LLM control, by contrast, directly tests the paper's central attribution. Note that the code is available, so this control is easily run. If the control shows a large gap, the current CONDITIONAL verdict could be upgraded; if not, the abstract and contribution claims would need revision. I therefore recommend no change to the reader's CONDITIONAL verdict.","tokens_in":18549,"tokens_out":9808,"duration_ms":119600,"concrete_test":"Run one control experiment: replace the frozen GPT-2 text branch in BALM-TSF with a randomly initialized 6-layer Transformer encoder of the same hidden size trained end-to-end on the same statistical prompt embeddings, keeping scaling, contrastive loss, concatenation, and forecasting head identical. Evaluate long-term MSE/MAE on all six benchmarks and few-shot on the four ETT datasets. If the non-LLM variant is within 2% of BALM-TSF's average MSE, the central claim that LLM-based balanced multimodal alignment drives the gains is not established; the paper should then report this control and soften the 'LLM-based' attribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.4 ablates 'w/o scale + alignment', 'w/o alignment', 'w/o scale', and 'w/o learnable prompt', but every variant keeps the frozen GPT-2 text branch. The central claim is that balanced multimodal alignment lets an LLM contribute complementary semantic information; if the text branch is simply a learned encoder over the same statistics (min/max/median/trend/top-5 lags), the same scaling+contrastive recipe might reproduce the results. Since the statistical prompt is derived from the input series, the 'textual modality' is not truly exogenous, and the only source of external semantic knowledge is GPT-2's pretrained weights. Without a non-LLM or no-text control, the 8.9% MSE reduction over Time-LLM could be due to architecture/statistics/alignment alone, not to LLM-based multimodal alignment. This is a load-bearing gap because it targets the paper's stated contribution, not a peripheral hyperparameter.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces BALM-TSF, a dual-branch LLM-based time series forecasting framework. One branch encodes normalized, patched time series via a PatchTST-style encoder; the other feeds descriptive statistics (min, max, median, trend, top-5 lags) plus a learnable prompt through a frozen GPT-2. To balance the modalities, the textual embeddings are truncated according to the prediction horizon, rescaled to the time series embedding scale, and aligned with a contrastive loss. The aligned representations are concatenated and projected to the forecast. Experiments on six standard benchmarks compare against four LLM-based and four non-LLM baselines in long-term and few-shot settings, with ablations and parameter-sensitivity analyses. The claimed contributions are a diagnosis of modality imbalance, a lightweight alignment mechanism, and state-of-the-art accuracy with only 0.97M trainable parameters.","tokens_in":18753,"tokens_out":3440,"duration_ms":43434,"significance":"If the claimed results hold under a properly qualified comparison, the paper makes a useful practical contribution: a very lightweight (0.97M trainable parameters) LLM-based forecaster that can match or beat much larger models on several ETT/Exchange settings. The method is simple and the Section 3 equations are internally consistent. The paper also has positive reproducibility features: code is released, standard benchmarks are used, and the ablations show that each component individually matters. The main significance is therefore conditional on two issues: (i) the 'state-of-the-art' claim is contradicted by the paper's own tables on Weather and on some few-shot comparisons; (ii) the ablations never remove or replace the frozen LLM text branch, so it is not established that LLM semantic knowledge, rather than the dual-branch statistical-prompt architecture plus alignment, drives the gains.","major_comments":[{"comment":"The abstract claims BALM-TSF 'achieves state-of-the-art performance in both long-term and few-shot forecasting,' but Table 1 shows this is not true on Weather: BALM-TSF's average MSE/MAE are 0.243/0.281, worse than GPT4TS (0.228/0.266), Time-LLM (0.230/0.269), and PatchTST (0.232/0.273). Likewise, in the few-shot Table 2, ETTh1 average MSE 0.683 is much worse than DLinear's 0.530, and on ETTh2 DLinear also wins (0.412 vs 0.460). The claim should be qualified to 'state-of-the-art among LLM-based methods on ETT/Exchange' or 'competitive overall'; as written, the central contribution is overstated by the paper's own numbers.","section":"Abstract and Section 4.2, Table 1"},{"comment":"Every ablation variant — 'w/o scale + alignment', 'w/o alignment', 'w/o scale', 'w/o learnable prompt' — retains the frozen GPT-2 text branch with the statistical prompt. Therefore the experiments never isolate whether the gains come from LLM-based multimodal semantic knowledge or simply from the dual-branch design with a statistical-prompt encoder and the scaling/contrastive recipe. Since the statistical prompt is derived from the same input series and GPT-2 is frozen, a non-LLM encoder (or a no-text control) trained with the same scaling and contrastive objective could plausibly reproduce the results. This is load-bearing because the paper's stated novelty is 'balanced multimodal alignment' for the textual modality, not just a dual-branch architecture. Please add a control that removes the text branch or replaces the frozen LLM with a lightweight non-LLM text encoder, and report the re","section":"Section 4.4, Ablation Study"},{"comment":"The adaptive truncation rule N_E = min(N_P, floor(N_P*H/L)) is derived from the premise that, due to masked self-attention, final tokens capture richer context, and that longer horizons benefit from more text. However, the only empirical support is Figure 6, which is measured on ETTh1/ETTh2 and tests exactly the values 12, 24, 42, 64 generated by the formula for H=96,192,336,720 plus one extra value (72). This does not validate the formula on other datasets or against alternative truncation strategies (e.g., fixed length, first tokens, learned attention pooling). Since the scaling and contrastive alignment operate on the truncated text representation, an out-of-sample validation or an ablation on the truncation rule is needed to establish that the method's gains are not an artifact of fitting Eq. (7) to the ETTh datasets.","section":"Section 3.4.1, Eq. (7)"}],"minor_comments":[{"comment":"The sentence 'BALM-TSF achieves the lowest or second-lowest error on all ETT and Exchange dataset–horizon combinations' is accurate only for ETT/Exchange, not for Weather; please make this explicit earlier to avoid the impression that the aggregate 'state-of-the-art' phrase covers all six benchmarks.","section":"Section 4.2 / Table 1"},{"comment":"The text mentions 'the horizon of 512 benefits from increasing value', but the experiments use horizons 96, 192, 336, 720. This appears to be a typo; please correct.","section":"Section 4.5, Figure 7 caption/text"},{"comment":"The y-axis label reads 'T extual Embedding Length'; there is an extra space. Also, the figure would benefit from marking the N_E values selected by Eq. (7) for each horizon, so the reader can directly see the relation to the truncation rule.","section":"Figure 6"},{"comment":"TimeCMA is labeled 'LLM in model? no', yet Section 2.1 and Table 1 describe TimeCMA as an LLM-based dual-branch model using a frozen LLM for embeddings. Please clarify the distinction between 'using an LLM in the pipeline' and 'having trainable LLM components' to avoid confusion.","section":"Section 4.6, Table 3"},{"comment":"The efficiency table omits GPT4TS, UniTime, and non-LLM baselines. Since the claim is about lightweight efficiency relative to LLM-based methods, adding GPT4TS and reporting training time (not just inference speed) would strengthen the comparison.","section":"Section 4.6, Table 3"},{"comment":"The alignment weight lambda is set to 1, but Figure 7 shows horizon-dependent sensitivity. A sentence explaining how the single global lambda=1 was chosen (e.g., validation-set selection across all horizons) would preempt concerns about cherry-picking.","section":"Section 4.1.3 / Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be a camera-ready CIKM '25 version, so the authors may be constrained in how much they can add. I would still ask for a text-branch/no-text control and a qualification of the SOTA claim; both are feasible as post-hoc experiments and text changes. If the authors cannot add experiments, the paper would need to be very explicit about the limitations in scope. I did not find citation or novelty-disclosure issues beyond the overstatement described in the report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nQuick take: BALM-TSF is a solid, well-specified empirical paper on LLM-based time series forecasting, but the 'state-of-the-art' claim is broader than the paper's own tables support, and the central mechanism—the frozen LLM's contribution—is never actually isolated.\n\nWhat's new: the specific combination of a statistical prompt, a frozen GPT-2, variance-ratio scaling, horizon-dependent tail truncation, and InfoNCE alignment. Each ingredient exists in earlier work, but the package is new, and the efficiency story is credible: only ~0.97M trainable parameters (1.17% of total), with competitive accuracy and lower inference memory than Time-LLM. The math in Section 3 is internally consistent, and the ablations (w/o scale, w/o alignment, w/o learnable prompt) all show degradation, so each component earns its place. Code is available, too.\n\nThe soft spots are real but mostly addressable. First, the abstract's 'state-of-the-art' overreaches: on Weather, GPT4TS has lower average MSE (0.228 vs 0.243), and in few-shot ETTh1, DLinear beats them. That is not fatal, but the claim needs qualification. Second, the main tables report only averaged results over three runs, no error bars; the ablation figure has them, so the authors know how to do it. Third, and most importantly, the stress-test concern lands: every ablation keeps the frozen GPT-2 text branch, so we never learn whether the gains come from LLM-based semantic alignment or just from a learned transformation of the same statistics. A control with a randomly initialized GPT-2 or a simple MLP on the statistical prompt would answer this. The truncation rule is also supported by a sensitivity analysis on the same ETTh datasets that motivated it—mild circularity, not a dealbreaker. Also, the recent CALF and TEST baselines are cited but not compared.\n\nWho should read this: anyone working on LLM+time series forecasting, especially those interested in parameter-efficient designs. It is a legitimate benchmark, not a breakthrough. If I were handling it, I would send it to peer review—it is a serious, clear, reproducible-in-principle paper—but I would require the missing control and error bars before accepting. The core architecture may survive without the LLM, which would change the narrative but not the utility.","headline":"BALM-TSF is a solid, reproducible LLM-forecasting paper, but the SOTA claim overreaches and the missing no-LLM control leaves its central mechanism unproven.","tokens_in":19291,"tokens_out":3893,"would_cite":true,"duration_ms":45642,"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":"BALM-TSF claims that balancing text and time series modalities with a two-step scaling-plus-contrastive alignment lets a 0.97M-parameter forecaster beat much larger LLM-based forecasters on six benchmarks.","keywords":["Time Series Forecasting","Multimodal Learning","Prompt Learning","Large Language Models","Contrastive Alignment","Few-shot Forecasting","Modality Imbalance"],"falsifier":"Run BALM-TSF with the adaptive truncation reversed (keep the first N_E tokens) or with N_E drawn randomly across horizons; if MSE does not degrade on ETTh or degrades less on other datasets, the horizon-aware tail-retention premise is not load-bearing. Also, hold out a dataset whose statistical prompt order is permuted so statistics appear at the beginning; if gains vanish, the tail-retention assumption is confirmed as the cause.","tokens_in":18433,"feed_emoji":"📉","tokens_out":6312,"duration_ms":66676,"temperature":0.7,"pith_summary":"This paper argues that LLM-based time series forecasters suffer from modality imbalance: pre-trained language models dominate the semantic content and embedding distribution, drowning out the numerical signal. To fix this, BALM-TSF processes time series in a dedicated patch encoder and feeds only descriptive statistics to a frozen GPT-2, then aligns the two branches with a horizon-aware scaling step and a contrastive loss. On six standard benchmarks, the model claims an average 8.9% MSE and 5.8% MAE reduction over the strongest LLM baseline, Time-LLM, while training only 0.97M parameters. The takeaway: with the right balance, a small model can leverage LLM representations without letting text overwhelm the temporal data.","feed_headline":"Balanced alignment trims LLM forecast MSE by 8.9%","feed_subtitle":"A 0.97M-parameter model outperforms large LLM forecasters across six benchmarks.","key_machinery":"The central object is the balanced multimodal alignment module: a two-step procedure that first adaptively truncates textual embeddings to N_E = min(N_P, floor(N_P·H/L)) tokens based on forecast horizon, then rescales them by the standard-deviation ratio α = STD_time/STD_text, and finally applies an InfoNCE contrastive loss between the averaged, ℓ2-normalized text and time series embeddings. This machinery enforces both distributional consistency (matching variance) and semantic alignment (pulling corresponding text–series pairs together), so the concatenated representation feeding the forecast head is balanced rather than text-dominated.","core_discovery":"The central claim is that modality imbalance—not model capacity—is the key bottleneck in LLM-based time series forecasting. BALM-TSF shows that by (1) avoiding raw series tokens in the LLM and using compact statistical prompts enriched with a learnable prompt, (2) adaptively truncating textual embeddings based on prediction horizon, (3) rescaling them to match time series embedding variance, and (4) pulling matching text–series pairs together with InfoNCE while pushing others apart, the two modalities can be integrated without one dominating. The result is state-of-the-art or near-state-of-the-art MSE/MAE across ETT, Exchange, and Weather benchmarks in both long-term and few-shot (10% traini","pith_inferences":["The truncation rule assumes tail tokens are most informative; a natural extension would test whether the same retained-token count rule transfers to prompts where key statistics are front-loaded, or to datasets with different tokenizations.","Because the scaling factor is computed per-sample from standard deviations, the method could be extended to align higher moments (skew, kurtosis) or to use a learned affine map instead of a scalar, which may further reduce distributional gap.","The contrastive objective uses only paired time series/statistics; extending to supervised contrastive with forecast target similarity or to harder negatives from same-series different-horizon pairs might improve few-shot stability.","The claim that textual modality is auxiliary suggests an adaptive gating mechanism could weight text contributions per dataset or per channel, potentially recovering gains on Weather where univariate statistics were insufficient."],"forward_implications":["If correct, LLM-based forecasters can be made much lighter: training 0.97M parameters instead of 53M yields better accuracy, so parameter-efficient forecaster design is a viable alternative to scaling.","The horizon-aware truncation implies that the optimal textual context is a function of forecast length: short horizons want minimal text, long horizons benefit from richer statistical summaries.","Contrastive alignment plus scaling is sufficient to bridge text and time series distributions, suggesting simpler alternatives to cross-attention or complex fusion modules.","Benchmark results imply that descriptive statistics prompts carry enough information for the LLM to add value, without feeding raw values.","Few-shot gains (10.7% MSE reduction vs GPT4TS) suggest balance helps most when data is scarce."],"supporting_citations":[{"why":"Supplies the statistical prompt template, the benchmark protocol, and the strongest LLM baseline that BALM-TSF claims to improve by 8.9% MSE.","marker":"[15]"},{"why":"Provides the patch-based encoding used to turn raw series into tokens for the time series branch.","marker":"[30]"},{"why":"Defines the InfoNCE objective used for semantic alignment between text and time series embeddings.","marker":"[31]"},{"why":"Introduces reversible instance normalization, the source of the near-zero-variance time series embedding distribution that motivates the scaling step.","marker":"[17]"},{"why":"The dual-branch cross-attention baseline it must beat, representing the heavier alignment approach.","marker":"[24]"},{"why":"Supports the premise that final tokens carry richer context under masked self-attention, justifying the truncation rule.","marker":"[9]"},{"why":"LLM baseline that directly consumes raw series; the strongest few-shot comparator it claims to beat by 10.7% MSE.","marker":"[49]"}],"fun_headline_variants":["Modality balance, not size, drives LLM forecasting gains","Tiny model beats big LLMs by balancing text and time","BALM-TSF: 0.97M params, SOTA forecast via modal balance","Why LLM forecasters fail: modality imbalance—and a fix","Balanced text–time alignment lifts forecast accuracy"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method's gains rest on the assumption that the tail end of a prompt holds the most useful statistical information, an assumption the paper tests only on ETTh data.","fun_headline_variants_meta":{"raw":{"variants":["Modality balance, not size, drives LLM forecasting gains","Tiny model beats big LLMs by balancing text and time","BALM-TSF: 0.97M params, SOTA forecast via modal balance","Why LLM forecasters fail: modality imbalance—and a fix","Balanced text–time alignment lifts forecast accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1183,"prompt_tokens":794,"completion_tokens":389,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":298}},"tokens_in":538,"tokens_out":389,"duration_ms":4364,"temperature":1.0,"reasoning_tokens":298,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:23:22.615355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BALM-TSF with the adaptive truncation reversed (keep the first N_E tokens) or with N_E drawn randomly across horizons; if MSE does not degrade on ETTh or degrades less on other datasets, the horizon-aware tail-retention premise is not load-bearing. Also, hold out a dataset whose statistical prompt order is permuted so statistics appear at the beginning; if gains vanish, the tail-retention assumption is confirmed as the cause.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces reversible instance normalization, the source of the near-zero-variance time series embedding distribution that motivates the scaling step."}],"review_version":1}