{"id":"4d9cd544-2899-41d0-b0f8-59b6d8201d1e","arxiv_id":"2411.11465","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Transformer models trained from scratch on in-context linear function prediction learn boundary-limited interpolation, not general linear regression.","lead":"Small transformers trained from scratch to do in-context learning of linear functions fail to extrapolate beyond the range of values seen during training, clamping predictions at what the paper calls boundary values. The paper uses this to argue that these models are not implementing linear regression in the way prior work claimed.","discovery_kind":"replication","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Saturation at boundary values is evidence of a bounded readout, not of the absence of an internal linear-regression algorithm; the no-ICL2 conclusion is therefore unsupported.","rationale":"The paper's strongest claim is that transformer ICL of linear functions is limited by training-range saturation rather than by an underlying general algorithm. The load-bearing issue is not whether boundary values equal the exact training extrema (the reader's weakest_assumption), but whether boundary saturation at the output can discriminate between 'no linear regression' and 'linear regression with a saturating readout'. The paper's own logic assumes that a model implementing linear regression would necessarily extrapolate, but a bounded final projection can clip an internally correct computation. This alternative is particularly plausible because the final LayerNorm/unembedding head is a bounded linear operator, and the observed precision/generalization tradeoff is exactly the signature of adjusting output scale. The paper's numbers even weaken the memorized-extrema story: for N(0,1) training, the expected maximum of 1.3B samples of f(x)=ax+b is about 9.2, not the reported 13.7. Given that the central conclusion rests on an unexamined internal computation, the paper needs either mechanistic evidence or a softened claim. Since the reader already requires conditional acceptance with softened interpretation and removal of the speculative math, my concern sharpens the required revision but does not change the verdict. The empirical phenomenon of boundary values is credible and worth reporting; the unsupported part is the inference to the absence of ICL2.","tokens_in":15372,"tokens_out":9205,"duration_ms":97995,"concrete_test":"Take the 12L8AH model trained on DF=DI=U(−5,5). Freeze all parameters and fit a ridge linear probe on the hidden state at the last transformer layer (before the final LayerNorm/head) to predict (a,b) from prompts whose f(x_i) are all within [−30,30]. If held-out R^2 > 0.95, slope and intercept are linearly decodable from the residual stream, so the model internally knows the function and the boundary is a readout artifact, falsifying the no-ICL2 claim. A complementary control: retrain only the final scalar head on the same prompts with labels scaled by 10; if the frozen-body model then extrapolates beyond ±30, the original boundary was not in the computation. If both probes fail while in-range predictions are accurate, the paper's interpretation is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central step is a false dichotomy: because outputs saturate at B± and predictions degrade outside [B−, B+], it concludes that no model can be implementing linear regression. This does not follow. A transformer trained on in-range targets has no loss incentive to produce unbounded outputs; the final LayerNorm/unembedding head is a bounded linear map, and the reported accuracy/generalization tradeoff is exactly what an output-scale increase should produce. The paper never probes internal states or intervenes on the readout, so boundary saturation is compatible with 'latent least-squares regression followed by a clipping readout.' The invariance claim in Observation 4(i) is asserted rather than verified and even conflicts with the paper's own numbers: for N(0,1) training, f(x)=ax+b with a,b,x~N(0,1) has std sqrt(2), so among 1.3B training values the expected maximum is about 9.2, yet the reported boundary values are about 13.7. Thus the boundaries look like a scale of the output parameterization, not memorized extrema. The conclusion 'the reason our models cannot ICL2 is boundary values' is therefore unsupported unless the internal computation is examined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies in-context learning (ICL) of univariate linear functions in small transformers trained from scratch on synthetic data. The authors distinguish ICL1 (low error when training and test distributions match) from ICL2 (recovering the linear function's parameters, e.g., via linear regression). They report that their models achieve ICL1 but fail to generalize out-of-distribution, saturating at empirically determined 'boundary values' B- and B+ that they conjecture equal the extrema seen in training. They conclude that transformers do not implement linear regression for this task, propose a memory-based projection hypothesis H to explain the observations, and describe an accuracy-generalization dilemma. The paper includes a large set of experiments across architectures and distributions, with code provided.","tokens_in":15594,"tokens_out":2673,"duration_ms":29937,"significance":"If the central claim were established, it would challenge a substantial body of work suggesting that transformers trained on linear regression tasks implement gradient-descent-like linear regression. The empirical documentation of OOD degradation and saturation across many architectures and training distributions is a useful contribution, and the paper is careful to train from scratch, vary architectures systematically, and provide code. However, the paper's headline conclusion—that 'none of our transformer models use linear regression'—is not supported by the evidence presented, because the authors do not probe or intervene on internal representations, and the boundary-value observations are equally compatible with an internal algorithmic computation followed by a bounded readout. The proposed hypothesis H is also under-specified. The contribution is therefore best viewed as a challenging empirical phenomenon that needs stronger mechanistic evidence before the strong negative claim is accepted.","major_comments":[{"comment":"The inference from output saturation to 'models do not use linear regression' is a false dichotomy. A transformer whose final LayerNorm/unembedding head is a bounded linear map, trained with squared error, has no loss incentive to emit values far beyond the training range; its outputs would saturate even if the hidden computation were an accurate least-squares estimator that is subsequently clipped or normalized. The paper never inspects hidden activations, ablates the readout, or performs any intervention that would distinguish 'latent linear regression plus bounded readout' from 'non-algorithmic interpolation.' Without such evidence, the central claim that 'none of our transformer models use such techniques' (Section 1) is not established.","section":"Section 4, Observation 3; Section 6"},{"comment":"The claim that boundary values B-, B+ are determined by the maximal and minimal function values encountered during training is asserted but not verified. For training with DF=DI=N(0,1), f(x)=ax+b with a,b,x i.i.d. N(0,1) gives f(x) ~ N(0,2), whose standard deviation is sqrt(2)≈1.414; among roughly 1.3 billion training samples the expected maximum is about 9.2, yet the paper reports boundary values near 13.7 for the 12L8AH model (Section 5, Figure 9). This mismatch suggests the boundaries may reflect the scale of the output parameterization (e.g., the unembedding matrix) rather than memorized extrema. The authors should directly test the memorization hypothesis by varying the output head scale independently, or by probing the model to show it encodes the extrema.","section":"Section 5, Observation 4(i)"},{"comment":"Observation 4 is partly post-hoc: the parameter alpha_M is a free parameter introduced to fit the saturation curves, and the statement 'B+ < f(v) < B+ + alpha implies hat f_M(v) ≈ B+' is a restatement of the observed saturation rather than a falsifiable prediction. The subsequent hypothesis H (Section 6) is described only in prose, with no formal definition of the projection pi, the stored projections, or the closeness criterion that determines boundary behavior. Appendix G's Proposition 1, which ostensibly gives a precise advantage of H over the 3-nearest-neighbor baseline, contains undefined quantities (i, m, n) and an unproved inequality. A 'mathematically precise hypothesis' requires a formal model that generates testable predictions beyond the data from which it was inferred.","section":"Section 5, Observation 4(ii); Section 6, Hypothesis H"},{"comment":"The claimed accuracy-generalization trade-off ('less accuracy' for models trained on larger variance) is evaluated using raw squared error, which is not scale-invariant. A model trained on N(0,10) will naturally have larger squared errors on N(0,1) targets than a model trained on N(0,1) simply because its predictions have larger variance; this does not establish lower precision in a relative sense. Reporting normalized metrics (e.g., R^2, or squared error divided by target variance) would make the trade-off claim meaningful and comparable across training distributions.","section":"Section 4, Observation 2; Table 1"}],"minor_comments":[{"comment":"The optimization objective writes 'Exi∈DI, f∈DF' but f is sampled from DF as a function; the notation should be made explicit (e.g., f∼DF, x∼DI) to avoid confusion between the function f and the distribution of its values.","section":"Section 3"},{"comment":"Tables 1 and 2 contain overlapping rows and appear to be near-duplicates; the captions also refer to 'Table 2 in the Appendix' in a way that is easy to misread. Please consolidate the tables or cross-reference them precisely.","section":"Tables 1 and 2"},{"comment":"The statement 'P (H(x) < f(xn) + ϵ) = im × ϵn, where i > 0, m, n < 41' is not a well-formed derivation; the variables i, m, n are undefined and the expression does not follow from the previous line. This proposition needs either a rigorous proof or should be removed.","section":"Appendix G"},{"comment":"In the sentence 'the higher the probability that πM(xn) = a yz_n, for a < α_M', the symbol 'a' is used both as the linear function's slope and as a multiplier; this is confusing and should be replaced with a distinct symbol.","section":"Section 6"},{"comment":"The text says 'contrary to what [2,23] have suggested' but then discusses [18]'s nearest-neighbor proposal; please ensure all citation numbers match the intended references (there is also an apparent mismatch between [17] and [18] in Section 6).","section":"Section 5 / Appendix E"}],"recommendation":"major_revision","confidential_remarks":"This paper presents a clean, well-scoped empirical study with a reproducible setup, but the main negative claim ('no linear regression') is currently an overclaim: the evidence does not rule out internal algorithmic computation with a bounded readout. The authors should be encouraged to add mechanistic evidence (probing or interventions) or to soften the claim to 'outputs are bounded by training-range scales,' which is well supported. If they decline to add experiments, the manuscript would need substantial reframing to avoid overclaiming. The proposed hypothesis H is too under-specified to constitute the 'mathematically precise hypothesis' promised in the abstract; formalization would strengthen the paper considerably."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the empirical core is real, the interpretation overshoots. The paper documents that small transformers trained on in-context linear regression saturate at training-range boundary values and fail to extrapolate; that is worth knowing. But the conclusion that 'no transformer implements linear regression' does not follow from saturation alone.\n\nWhat's new: they test a wide range of architectures and distributions, and the boundary-value phenomenon plus the accuracy/generalizability tradeoff is a genuine contribution. The attention-only ablation is also useful. They ship code and give enough experimental detail to reproduce.\n\nSoft spots, in order of severity. First, the central inference is a false dichotomy. Outputs bounded by B± are exactly what you'd expect from a model with a bounded output head trained only on in-range targets; a latent least-squares computation feeding into a clipped readout would produce the same behavior. The paper never probes internal representations or intervenes on the readout, so it cannot distinguish these. Second, Observation 4(i) — boundaries equal memorized extrema — is asserted, not verified, and their own numbers conflict with it. For N(0,1) training, the expected maximum among 1.3B values of ax+b is around 9, yet they report boundaries near 13.7; that looks like output-layer scale, not a memory trace. Third, α in Observation 4 is a post-hoc fit to the saturation curves. Fourth, Proposition 1 in the appendix is not a valid mathematical argument; the probability expressions are ill-defined and it should be removed or replaced.\n\nThe paper does not engage with the possibility that the model could be doing regression internally and only saturating at the readout, which is the main gap. The projection hypothesis is under-specified and not directly tested, so it reads as speculation.\n\nWho this is for: people working on interpretability of ICL and on claims that transformers implement gradient descent. The empirical boundary phenomenon is a useful caution. A serious referee should engage: the paper deserves review, but it needs a major revision that softens the no-ICL2 conclusion and fixes the appendix.\n\nRead it with skepticism about the interpretation but take the data seriously.","headline":"Solid empirical documentation of boundary saturation in ICL, but the 'no linear regression' conclusion is an overreach that needs a major revision before publication.","tokens_in":16098,"tokens_out":2504,"would_cite":true,"duration_ms":25435,"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":"Transformers don't learn linear functions in context — they clip them to a training-range band.","keywords":["in-context learning","linear functions","transformers","boundary values","generalization","memorization","attention layers","out-of-distribution"],"falsifier":"Train a model on $U(-5,5)$ but with a target function whose values stay within a narrower band, e.g., $f(x)=2x$ with $x\\in[-5,5]$ so the maximum possible output is $\\pm 10$. If boundary values appear around $\\pm 10$, the memorized-extrema account is supported; if they remain near $\\pm 30$, the boundary is set by something other than training extrema, such as the unembedding matrix or loss scale.","tokens_in":1671,"feed_emoji":"🤖","tokens_out":3846,"duration_ms":87086,"temperature":0.7,"pith_summary":"The paper tries to establish that transformer models trained from scratch on synthetic data do not learn a general algorithm for in-context learning (ICL) of linear functions, even though they perform well when the test distribution matches training. Instead, every model exhibits boundary values—prediction limits set by the largest and smallest function values seen during training—and degrades sharply when asked to predict values outside that range. This challenges the common view that transformers implement something like linear regression for such tasks. If correct, it means current ICL is a memorization-and-interpolation mechanism with inherent limits, not an abstract function learner, which matters whenever test inputs fall outside the training distribution.","feed_headline":"Transformers don't learn linear functions in context","feed_subtitle":"Over 30 models clip predictions to a training-range band, so no true linear-function algorithm emerges.","key_machinery":"The central object is the pair of boundary values $B^-$ and $B^+$, which are the largest and smallest function outputs the model has memorized from training data. They act as a bandpass filter: for target values just outside the boundaries, the model predicts a constant near the boundary; for values farther out, it predicts a random value inside the interval. This mechanism carries the argument against algorithmic ICL because it explains the observed error patterns and the accuracy-versus-generalizability dilemma. The paper also sketches a projection hypothesis $\\pi$ in which the attention layers store pairs of input and output values and interpolate among the closest stored projections.","core_discovery":"The central discovery is that across over 30 transformer architectures, in-context learning of univariate linear functions is governed by training-range saturation: every model has boundary values $B^-$ and $B^+$ beyond which its predictions become constant (near the boundary) or random (within the boundary interval). The paper distinguishes two senses of learning—ICL1, matching expected error on the training distribution, and ICL2, actually recovering the parameters $a,b$ of $f(x)=ax+b$—and shows that while models achieve ICL1, they never achieve ICL2. The boundary values coincide with the extremes encountered in training data, indicating the model memorizes those extremes and uses them as a bandpass filter. Consequently, performance outside the boundaries degrades dramatically, and there is a trade-off: training on narrow distributions gives high accuracy but poor generalization, while training on wider distributions improves generalization but loses precision. The paper argues this rules out the linear-regression account of ICL and instead supports a memory-based projection hypothesis.","pith_inferences":["Beyond the paper: this boundary-value behavior may explain analogous out-of-distribution failures in larger language models on arithmetic and reasoning tasks; if ICL is memory-based interpolation, any prompt pulling outputs beyond the training range could trigger the same saturation or hallucination pattern.","Beyond the paper: the accuracy-versus-generalizability trade-off suggests that simply scaling training data may not overcome the dilemma unless the architecture can represent an abstract mapping separate from the training range; architectures that force explicit function representations would be a testable alternative.","Beyond the paper: the projection hypothesis implies ICL performance is bounded by the density of stored $(x,y)$ pairs near the query; controlling the density of training points in a region and observing whether boundary values and interpolation quality move accordingly would provide a direct test.","Beyond the paper: benchmarking ICL only on in-distribution test sets (the ICL1 criterion) is misleading; a model can pass such a test while lacking abstract understanding, so benchmark design should always include out-of-distribution probes."],"forward_implications":["In-context learning of linear functions is not an abstract algorithm: models trained on one distribution cannot extrapolate beyond the training range, even though linear regression would trivially do so.","Accuracy and generalizability trade off: training on a wider function and input distribution extends the boundary values but degrades precision on the original narrow distribution.","Attention layers are necessary and sufficient for in-context learning of this task; MLP-only models cannot do it.","The model uses the entire prompt context, not a fixed small subsequence, to make its prediction, contradicting local-averaging accounts like the three-nearest-neighbor proposal.","Larger models (with more layers and heads) have slightly wider boundary values and better performance within the boundary interval, but still fail beyond it."],"supporting_citations":[{"why":"Claimed transformers learn in-context via gradient descent and linear regression; this is the primary narrative the paper refutes.","marker":"[2]"},{"why":"Proposed that transformers learn in-context by gradient descent; another central target of the refutation.","marker":"[23]"},{"why":"Suggested transformers implement preconditioned gradient descent for in-context learning, which the paper's boundary-value results contradict.","marker":"[1]"},{"why":"Argued that one step of gradient descent is the optimal in-context learner for one-layer linear attention; refuted by the observed failure to extrapolate.","marker":"[13]"},{"why":"Provided the original setup and baseline: a transformer trained from scratch achieves low in-context error when train and test distributions are identical, defining ICL1.","marker":"[9]"},{"why":"Their figures also show boundary values even with linear attention, supporting the generality of the observed saturation phenomenon.","marker":"[10]"},{"why":"Proposed induction heads and a three-nearest-neighbor averaging mechanism for ICL; the paper tests and rejects this as an explanation for linear-function learning.","marker":"[18]"},{"why":"Claimed shifting the prompt distribution does not affect ICL performance; the paper shows performance degrades when both prompt and query distributions are shifted beyond boundaries.","marker":"[27]"}],"fun_headline_variants":["Transformers memorize, not compute, linear functions in ICL","ICL fails: models saturate at training-range boundaries","Linear functions in context? Models just clip to training range","Transformers can't infer linear rules beyond training extremes"],"cache_read_input_tokens":18304,"weakest_assumption_plain":"The claim rests on the inference that boundary values are set by the largest and smallest function values the model happened to see in training; if those limits instead come from the model's output layer scale or loss-induced variance, the memory-based explanation collapses.","fun_headline_variants_meta":{"raw":{"variants":["Transformers memorize, not compute, linear functions in ICL","ICL fails: models saturate at training-range boundaries","Linear functions in context? Models just clip to training range","Transformers can't infer linear rules beyond training extremes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000313,"raw_usage":{"total_tokens":1735,"prompt_tokens":858,"completion_tokens":877,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":822}},"tokens_in":474,"tokens_out":877,"duration_ms":7446,"temperature":1.0,"reasoning_tokens":822,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:29:20.042014+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a model on $U(-5,5)$ but with a target function whose values stay within a narrower band, e.g., $f(x)=2x$ with $x\\in[-5,5]$ so the maximum possible output is $\\pm 10$. If boundary values appear around $\\pm 10$, the memorized-extrema account is supported; if they remain near $\\pm 30$, the boundary is set by something other than training extrema, such as the unembedding matrix or loss scale.","supporting_citations":[{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Proposed that transformers learn in-context by gradient descent; another central target of the refutation."},{"cited_title":"Advances in Neural Information Processing Systems 36, 45614–45650 (2023)","cited_arxiv_id":null,"evidence_quote":"Suggested transformers implement preconditioned gradient descent for in-context learning, which the paper's boundary-value results contradict."},{"cited_title":"Advances in Neural Information Processing Systems 35, 30583–30598 (2022)","cited_arxiv_id":null,"evidence_quote":"Provided the original setup and baseline: a transformer trained from scratch achieves low in-context error when train and test distributions are identical, defining ICL1."},{"cited_title":"Journal of Machine Learning Research 25(49), 1–55 (2024)","cited_arxiv_id":null,"evidence_quote":"Claimed shifting the prompt distribution does not affect ICL performance; the paper shows performance degrades when both prompt and query distributions are shifted beyond boundaries."}],"review_version":1}