{"id":"987c715a-c275-4794-8310-8a871a1b962b","arxiv_id":"2504.12523","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"KUP and MCT: a new benchmark and training method showing LLMs can memorize post-cutoff knowledge updates but fail to reason over them in indirect tests.","lead":"This paper creates a test bed and a training trick for teaching large language models new facts that appear after their training cutoff. The headline finding is that even models that clearly memorize the new fact still rely on outdated knowledge when asked indirectly, so current methods cannot yet keep AI assistants temporally consistent.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The '<2% indirect probing' headline is a measure of explicit entailment of the updated fact, not response accuracy; correct answers that omit the entity are binned as N/A, so the central 'memorize but cannot reason' claim may be a scoring artifact.","rationale":"The reader's stated weakest assumption is that the GPT-4O-generated fictitious updates faithfully represent real-world knowledge dynamics. That is a generalizability concern. I see a more direct, internal threat to the central claim: the quantitative evidence for 'LLMs memorize but cannot reason' is the indirect-probing result, and the paper's own table labels that result as entailment rates (UPD/OLD/N/A), not accuracy. The abstract and introduction then call the UPD rate 'accuracy,' conflating explicit entailment of the updated fact with correct reasoning. A model that correctly avoids mentioning the updated entity—for example, by listing other sponsors and omitting the festival—is binned as N/A rather than as correct. With 15–22% of responses in N/A, the true accuracy could be substantially higher than the reported <2%. This is not an attack on the direct-probing results, which appear carefully constructed, nor on the MCT method's direct-recall improvements, which have internal support from Table 5 and the free-form QA results. It is also not a claim that the models actually reason well; they may indeed fail badly. The point is that the headline dissociation between memorization and reasoning is measured by a rubric whose validity is untested and whose interpretation is inconsistent with the paper's own numbers (3.0% vs. '<2%'). A human re-scoring of the same 60 responses is a small, feasible check that would settle whether the central claim survives. Until that check is run, the central claim should be treated as unverified rather than conditionally accepted.","tokens_in":21039,"tokens_out":10801,"duration_ms":122219,"concrete_test":"Re-score the 60 indirect-probe responses for all models in Table 7 with three independent human annotators, blind to the UPD/OLD/N/A labels, judging whether each response correctly answers the question as asked, not whether it explicitly entails f_new. Report human accuracy per method. If human accuracy is materially above the UPD column (e.g., >20% for MCT while UPD is <2%), the 'cannot reason' claim is a scoring artifact; if human accuracy is also <5%, the claim stands. In the same pass, recompute Table 7 after reconciling the 3.0% Re+CoT/Mistral cell with the '<2%' abstract claim.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"Section 4.1.2 defines indirect probing as list-style questions and reports 'the fraction of times the trained model generated a response that entails the f_old knowledge vs. the f_new knowledge.' Table 7 operationalizes this as UPD/OLD/N/A, and the abstract converts UPD into 'accuracy': 'best CPT models achieving <2% in indirect probing setting (reasoning).' This conflation is load-bearing because any response that does not mention the target entity falls into N/A (14.7–22.0% of responses in Table 7). For a list-style question such as 'What events does Baillie Gifford still sponsor?', a correct answer that simply omits the no-longer-sponsored festival and lists other sponsors is scored N/A, not UPD. The paper reports no human or validated-judge correctness score for the 60-question indirect set, so the '<2% accuracy' claim is not yet established as an accuracy claim. The paper's own Table 7 also shows Re+CoT at 3.0% UPD for Mistral, contradicting the '<2% for all CPT approaches' wording in the abstract and §1. Because the central conceptual conclusion—LLMs memorize updates but cannot reason over them—rests on this indirect metric, the headline number is not secure. The RAG oracle row (69.6% UPD) is scored with the same entailment rubric, so it does not independently validate the metric.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Knowledge Update Playground (KUP), a pipeline that generates synthetic but realistic knowledge updates (f_old -> f_new) for 1000 entities, along with an evidence corpus of fictitious news articles and auxiliary scraped news. It also proposes memory conditioned training (MCT), which prepends self-generated memory tokens during continued pretraining and uses a memory-recall step at inference. Experiments on Llama-3.1-8B and Mistral-7B-v0.3 compare MCT with standard CPT and rephrased CPT, including a no-train baseline and RAG upper bounds. The paper claims that all CPT methods fail indirect probing (<2% 'accuracy'), that MCT significantly improves direct-probing memorization (up to 25.4 percentage points), and that KUP is a challenging benchmark for future work.","tokens_in":21365,"tokens_out":7401,"duration_ms":72551,"significance":"The paper has several strengths: the direct-probing experiments are carefully controlled, with f_old/f_new verified against the test models, 1% replay data added to all training runs, chain-of-thought applied uniformly, and both a no-train baseline and RAG upper bounds reported. The KUP dataset and code are released, and the manual analysis of CoT traces provides useful evidence about what MCT recalls. If the central claims were fully supported, the paper would make a valuable contribution to knowledge-update evaluation. However, the headline '<2% indirect probing' conflates entailment classification with answer accuracy, and the paper's own Table 7 contains a number that contradicts the '<2%' wording. The 'memorization but not reasoning' conclusion therefore needs re-analysis before the paper's main message can be accepted.","major_comments":[{"comment":"The headline claim that 'the best CPT models achieve <2% in indirect probing' is not supported as an accuracy claim. Table 7 reports only the fraction of responses entailing UPD, OLD, or N/A, and the N/A category (14.7–22.0%) includes responses that do not mention the target entity. For a list-style question such as 'What events does Baillie Gifford still sponsor?', an answer that correctly omits the no-longer-sponsored entity is scored N/A rather than UPD. No human-validated correctness score is reported for the 60-question indirect set, and the RAG oracle row is scored under the same N/A rubric, so it does not independently validate the metric. In addition, Table 7 shows 3.0% UPD for Mistral Re+CoT, contradicting the abstract's 'all CPT approaches' / '<2%' wording. Please re-analyze the indirect set with human annotation and distinguish answer correctness from entailment of the update.","section":"Abstract; §4.1.2; Table 7"},{"comment":"The indirect-probing protocol is underspecified with respect to the supervised fine-tuning step. The text says that for indirect probing the authors 'supervised fine-tune M>T' on the 4.2K direct-probing Q&A pairs, but it does not state whether the CPT and Re+CoT baselines received identical SFT or whether only the MCT-trained model was fine-tuned. If the training histories differ, the comparison in Table 7 is not controlled. Please specify the exact protocol for each row and, if necessary, re-run with identical SFT for all methods.","section":"§5.2; Table 7"},{"comment":"The free-form direct-probing evaluation provides the updated fact statement f_new to the model as context at test time. Table 6 therefore measures grounded question answering given the update statement, not closed-book memorization of the update from the evidence corpus. Since the paper uses Table 6 to argue that MCT outperforms baselines in direct probing, the authors should either evaluate without exposing f_new or explicitly state that Table 6 measures contextualized QA rather than memorization.","section":"§4.1.1; Table 6"},{"comment":"The external validity of the 'realistic knowledge updates' claim rests on GPT-4O-generated fictitious updates and news articles. The only validation against real knowledge dynamics is the 50-point manual annotation in Table 2, which measures the presence of narrative properties rather than whether the synthetic updates are representative of actual post-cutoff knowledge changes. Given that the paper's conclusion about CPT methods on realistic evidence corpora depends on this premise, I would like to see either a comparison on real evolving knowledge (e.g., Realtime QA or GROWOVER) or a human evaluation of update plausibility and difficulty. The limitation section acknowledges the issue but does not resolve it.","section":"§2.1; §9"}],"minor_comments":[{"comment":"The sentence 'Note that there may exist cases where the model generation does mention entity e, and therefore, cannot be classified as entailing either' should say 'does not mention entity e'; as written it states the opposite of the intended meaning.","section":"§4.1.2"},{"comment":"There is a typo: 'axillary documents' should be 'auxiliary documents'.","section":"Figure 2 and §2.1"},{"comment":"The model name is written inconsistently as 'LLama-3.1-8B' and 'LLama'; the correct spelling is 'LLaMA-3.1-8B'.","section":"Abstract; §1"},{"comment":"In the conclusion, 'systemically study' should be 'systematically study'.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially publishable after the indirect-probing metric is re-analyzed and the abstract is corrected. The synthetic-data concern is secondary but should be addressed in the revision; it does not by itself justify rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The empirical setup is more careful than most CPT benchmark papers: f_old/f_new are verified against the actual models, 1% replay is added to all methods, CoT is applied uniformly, and there are no-train and RAG upper bounds. But the abstract’s central negative result — that all CPT approaches get <2% on indirect probing — is not yet established. The indirect metric is entailment-of-updated-knowledge, not response correctness, and answers that simply omit the entity get binned as N/A. So a correct list that leaves out the no-longer-sponsoring company is scored as not-updated. The stress-test concern lands. Table 7 also contradicts the abstract: Re+CoT gets 3.0% UPD on Mistral.\n\nWhat is actually new: KUP goes beyond entity substitution. Contextual rewrites make up 89.8% of updates, and the dataset includes external trigger events and downstream impacts, which GROWOVER and ConflictBank largely lack. MCT — prepending self-generated memory chunks during CPT and recalling memory at inference — is simple and shows real gains on direct probing, especially with CoT. The direct-probing numbers look credible to me: the controls and the RAG ceiling make them interpretable. The analysis of CoT traces is a nice touch.\n\nSoft spots, in proportion. The indirect-probing set is 60 manually curated questions; that’s small. The free-form eval uses GPT-4O-MINI as judge with no human-validated subsample. The synthetic updates and news articles are generated by GPT-4O, so the transfer to real knowledge dynamics is assumed rather than demonstrated; the paper’s own manual annotation is only 50 points. None of this sinks the benchmark or the method, but it means the headline claim should be reworded as “failure to entail the updated fact,” not “accuracy below 2%.” The 3.0% number needs reconciling. I’d also want seeds or a memory-source ablation before trusting MCT’s exact gains, though the direction is plausible.\n\nWho it’s for: people working on continued pretraining, knowledge editing, or model staleness. They’ll get a useful testbed and a competitive baseline. I’d send it to peer review — it deserves referee time — but with a request to fix the indirect metric, reconcile the numbers, and add human evaluation on a sample. If that happens, this is a solid contribution.","headline":"Solid benchmark and training method, but the indirect-reasoning headline is not yet supported by the metric as defined.","tokens_in":21892,"tokens_out":2787,"would_cite":true,"duration_ms":29531,"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 that continued pretraining on a corpus of realistic news updates lets LLMs memorize new facts but not reason over them, and that memory-conditioned training improves direct recall by up to 25.4 percentage points.","keywords":["continued pretraining","knowledge update","knowledge conflict","memorization","reasoning","memory conditioned training","benchmark","chain-of-thought"],"falsifier":"Run the same continued pretraining and MCT pipeline on a corpus of real, verified post-cutoff updates, for example 2024-2025 news events with archived articles, using entities whose old facts the base model knows, and measure indirect probing of old versus new knowledge. If real-update indirect accuracy is substantially above 2% or if the MCT gains disappear, the synthetic KUP updates rather than the learning methods explain the reported results.","tokens_in":20814,"feed_emoji":"🧠","tokens_out":6955,"duration_ms":64329,"temperature":0.7,"pith_summary":"This paper tackles the problem of updating a large language model's parametric knowledge after its training cutoff. It introduces Knowledge Update Playground (KUP), an automatic pipeline that simulates realistic knowledge updates by generating fictitious news articles about 1000 entities, together with direct and indirect probes that separate memorization of a new fact from reasoning over its consequences. On KUP, the authors find that every continued pretraining method they benchmark, including data rephrasing, scores under 2% on indirect probes, meaning the models still answer as if the old fact were true. They then propose memory conditioned training (MCT), which prepends self-generated memory tokens to update documents during training, and report direct-probe accuracy gains of up to 25.4 percentage points over continued pretraining baselines. If the results hold, the field's standard way of updating LLMs teaches them to recall new facts without enabling them to apply those facts.","feed_headline":"Continued pretraining lets LLMs memorize, not reason, over updates","feed_subtitle":"A new benchmark finds under 2% accuracy when models must apply updated facts; MCT raises memorization by 25.4 points.","key_machinery":"KUP's curation pipeline: GPT-4O generates candidate mutable facts about 1000 entities, filters them for changeability and objectivity, produces contradictory updated facts with event sequences, and writes five fictitious news articles per update; the base models are probed to confirm they know the old fact and reject the new one. The training corpus combines 3.3M tokens of these evidence documents with 52.4M tokens of real auxiliary news. MCT is the learning mechanism: chunks of self-generated 'memory' completions sampled from the base model are prepended to each evidence document, the loss is masked on those memory tokens so the model focuses on the update, and at inference the model first recalls memory (treated as chain-of-thought) before answering. The direct/indirect probe split is the measuring instrument that lets the authors attribute failures to memorization versus reasoning.","core_discovery":"The central discovery is a separation between memorization and reasoning in model updates. After continued pretraining on KUP's evidence corpus, all tested CPT methods, including standard, rephrased, and memory-conditioned variants, fail at indirect probing, entailing the outdated fact in 78.5% to 83.2% of responses and the updated fact in under 2%. Direct probing tells a different story: MCT, which trains on entity evidence prefixed by chunks of the model's own Wikipedia-style completions about that entity, selects the updated fact over distractors at 60.7% (Llama) and 71.0% (Mistral) in one MCQ setting, up from 20.0% and 17.4% for standard CPT, with gains of up to 25.4 percentage points over the baselines. The paper also finds that memory recall at inference, functionally chain-of-thought, helps knowledge surfacing, and that training perplexity does not predict whether a model will answer correctly.","pith_inferences":["If the under-2% indirect-probe result holds, the bottleneck is not storing the updated fact but choosing it over the old fact during generation; retrieval-free update methods may need an explicit conflict-resolution step before they can match RAG on indirect questions.","Because KUP uses fictitious updates, the 25.4-point MCT gain may not transfer to real post-cutoff corpora; testing MCT on archived real news would separate method quality from synthetic-corpus ease.","The memory-token trick suggests a general recipe: any CPT method could elicit entity-specific context from the base model before training, and the same elicitation at inference could be chained with retrieved passages to push indirect accuracy.","The 60-question indirect set is small and manually curated; an automated generator of indirect probes over more entities would be needed to confirm the catastrophic failure rate."],"forward_implications":["Entity-substitution benchmarks likely overstate how well continued pretraining adapts LLMs to real-world knowledge dynamics, because KUP's updates are mostly contextual rewrites rather than simple attribute swaps.","Simply maximizing likelihood on an update corpus is not enough: perplexity on new evidence does not distinguish correct from incorrect MCQ answers, so training objectives need to reward surfacing the updated fact in conflict with old knowledge.","MCT is a practical, lightweight addition to CPT pipelines: it requires only self-generated memory chunks and a masked loss, yet yields the largest direct-probe gains when combined with inference-time memory recall.","Chain-of-thought prompting helps a knowledge-intensive task here, suggesting that CoT or memory recall can resolve parametric memory conflicts even when the task is not a traditional reasoning benchmark."],"supporting_citations":[{"why":"Defines the GROWOVER entity-substitution benchmark and baseline that KUP extends and compares against.","marker":"(Ko et al., 2024a)"},{"why":"Earlier continued-pretraining update benchmark using entity substitutions; motivates KUP's broader dynamics.","marker":"(Li et al., 2024)"},{"why":"Supplies the 1% RedPajama replay-data recipe and the rephrasing baseline approach.","marker":"(Yang et al., 2024a)"},{"why":"Chain-of-thought prompting, which the paper uses as the inference-time memory recall mechanism.","marker":"(Wei et al., 2023)"},{"why":"Provides the LLaMA-3.1-8B base model used in all experiments.","marker":"(Dubey et al., 2024)"},{"why":"Provides the Mistral-7B-v0.3 base model used in all experiments.","marker":"(Jiang et al., 2023)"},{"why":"ConflictBank benchmark used to show prior entity-substitution conflict datasets are unsuitable for update learning.","marker":"(Su et al., 2024)"},{"why":"RedPajama corpus provides the replay data used in the common training recipe.","marker":"(Weber et al., 2025)"}],"fun_headline_variants":["LLMs memorize facts but fail to reason over updates","New benchmark reveals LLMs can't apply updated knowledge","MCT boosts memory recall 25.4 points, reasoning still <2%","KUP benchmark: <2% reasoning accuracy on LLM updates","Memory vs. reasoning: CPT updates fall short, MCT helps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that GPT-4O-generated fictitious news articles about 1000 entities, filtered so the base model recognizes the old fact and rejects the new one, faithfully reproduce the structure of real post-cutoff knowledge updates; if those synthetic updates are easier or harder than real ones, the under-2% indirect-probe result and the MCT gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["LLMs memorize facts but fail to reason over updates","New benchmark reveals LLMs can't apply updated knowledge","MCT boosts memory recall 25.4 points, reasoning still <2%","KUP benchmark: <2% reasoning accuracy on LLM updates","Memory vs. reasoning: CPT updates fall short, MCT helps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000772,"raw_usage":{"total_tokens":3429,"prompt_tokens":966,"completion_tokens":2463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":2375}},"tokens_in":582,"tokens_out":2463,"duration_ms":17804,"temperature":1.0,"reasoning_tokens":2375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:29:54.249315+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same continued pretraining and MCT pipeline on a corpus of real, verified post-cutoff updates, for example 2024-2025 news events with archived articles, using entities whose old facts the base model knows, and measure indirect probing of old versus new knowledge. If real-update indirect accuracy is substantially above 2% or if the MCT gains disappear, the synthetic KUP updates rather than the learning methods explain the reported results.","supporting_citations":[],"review_version":1}