Coming home late

The idea

Joe Haldeman’s The Forever War (St. Martin’s Press, New York, 1974; library record) follows Private William Mandella, a reluctant conscript sent to fight an interstellar war against an enemy far away. The publisher’s description states the premise in one line: because of the time dilation caused by space travel, “Mandella is aging months while the Earth he left behind is aging centuries” (Macmillan). The novel won the Nebula Award for best novel (SFWA) and the Hugo Award in 1976 (The Hugo Awards). Haldeman is a Vietnam veteran (Macmillan).

My reading: the war is the frame, and the subject is the return. Each time Mandella comes home, the gap is larger. He has forgotten nothing. What he knows is exactly what he knew when he left, and that is the problem. The publisher’s copy puts it as a warning: home “may be even more terrifying than battle.”

The bridge

A language model is trained on a snapshot of text and then deployed. From that day on, it stays where its data stopped, and the world keeps going. The analogy holds at three specific points.

First, knowledge stops at a point and the world does not. Lazaridou and colleagues trained models on text up to a fixed date and tested them on text from later periods. Performance degraded, the degradation grew the further out they tested, and larger models did not fix it (Lazaridou et al., 2021).

Second, the stale knowledge is still there, and now it is wrong. Jang and colleagues framed the repair job as three tasks at once: keep what is still true, update what has changed, and add what is new. Their benchmark and metric measure all three (Jang et al., 2022).

Third, catching up is hard. Mandella cannot absorb centuries of change at once, and a model cannot either. Gekhman and colleagues found that during fine-tuning, models learn facts that are new to them much more slowly than facts consistent with what they already know. Once those new facts are learned, hallucination rises linearly with them. The authors conclude that models acquire factual knowledge mostly in pre-training, and that fine-tuning mainly teaches them to use what they already have (Gekhman et al., 2024).

What research says now

Staleness is gradual and uneven

The picture of a clean cutoff date is too simple. Luu and colleagues built eight tasks from social media, science papers, news and reviews, spanning five or more years. They found stronger effects of temporal misalignment than earlier work had reported. Continued pre-training on text from the new period helped only a little; fine-tuning on labelled data from that period helped more (Luu et al., 2022).

A May 2026 benchmark, TempoMed-Bench, tested models against clinical guidelines that changed over the years. Accuracy on current recommendations fell roughly linearly with time, without a sharp edge at the training cutoff. Accuracy on superseded guidance was only 25.37% to 53.89% of accuracy on the current guidance, which the authors suggest may reflect forgetting during training. Adding agentic search tools changed accuracy by between -3.15 and +14.14 percentage points (Guan et al., 2026).

Models do not know when they are

Mandella can calculate how long he was gone. A model usually cannot. Cheng and colleagues showed that a model’s effective knowledge cutoff often differs from the date its developers report, and can differ by topic and by source. They trace part of the gap to old Common Crawl data that persists in newer crawls, and to deduplication that misses near-duplicates (Cheng et al., 2024).

Zhao and colleagues found that Llama 2, despite a cutoff around 2022, mostly answered time-sensitive questions with older knowledge, often from 2019. Aligning it to 2022 improved its performance on that year’s questions by up to 62%, without mentioning time in the prompt (Zhao et al., 2024).

A June 2026 paper makes the point sharper. Hossain and colleagues define parametric temporal conflict: the model holds both the outdated fact and its replacement, and the default answer favours the outdated one. On a benchmark of 8,746 changes of office holders from Wikidata, adding a date to the prompt recovered the newer fact in 61% to 81% of these cases. Patching activations flipped the answer in 72% to 85% (Hossain et al., 2026). The new knowledge was present. The model reached for the old one first.

Retrieval as a briefing on arrival

The standard fix is to hand the model current documents at question time. Retrieval-augmented generation pairs the model’s internal memory with a searchable index and was proposed partly because updating knowledge inside the weights was an open problem (Lewis et al., 2020).

It helps, with limits. On FreshQA, a question set that tests current world knowledge, all tested models struggled with fast-changing facts and false premises, and adding search results to the prompt improved accuracy (Vu et al., 2023). RealTime QA found that GPT-3 often revised its answers when given new documents, but gave outdated answers when the retrieved documents lacked enough information (Kasai et al., 2022).

An August 2026 benchmark tests the hardest version of the briefing. TIDE asks which version of a regulation was in force on a given date, using 644 customs instruments issued in Bangladesh between 1969 and 2025. The best model reached 68.5% macro-averaged accuracy. Recognising that a supplied version did not govern the query succeeded only 26.7% of the time, and models tended to trust a confident answer from memory over the authoritative text in front of them (Sobhani et al., 2026).

Retraining without forgetting

The other path is to update the weights. Model editing changes individual facts: ROME locates factual associations in mid-layer feed-forward modules and rewrites one fact with a rank-one weight update (Meng et al., 2022). Editing one fact at a time does not scale to centuries of change, so the field also works on continual learning, where the main obstacle is catastrophic forgetting. One answer, elastic weight consolidation, slows learning on the weights that matter most for old tasks (Kirkpatrick et al., 2017). For language models specifically, Jang and colleagues found that retaining old knowledge and learning new knowledge at the same time seemed to require adding parameters (Jang et al., 2022).

Recent work makes the trade-off measurable. Lin and colleagues taught models new facts and checked how much old knowledge survived. On NaturalQuestions, F1 dropped 89% after full fine-tuning and 71% with LoRA. Updating only the memory slots that the new facts strongly activate learned a comparable amount and lost 11% (Lin et al., 2025). SEAL goes further and has the model write its own fine-tuning data and update instructions, trained with reinforcement learning on whether the update helped (Zweiger et al., 2025). Titans takes a third route: a long-term memory module that learns to store context at test time, while the attention layers handle the current window (Behrouz et al., 2024).

Where the analogy breaks

Mandella knows the size of the gap. Physics tells him how many years passed on Earth. A model has no clock unless someone gives it one, and the research above shows that even its effective cutoff can differ from the reported one. In this respect the model is worse off than the soldier. It also points to the cheapest fix: tell the model the date. In the conflict study above, a date in the prompt alone recovered the newer fact in most cases.

Mandella’s knowledge ages in one jump per trip. A model’s knowledge ages unevenly, topic by topic, and the edge is blurred, as TempoMed-Bench and the effective-cutoff work both show.

Mandella is one person who carries every return forward. A model usually comes home as a different model: the next version is trained on a new snapshot. My read: the closer image for model versions is a new recruit trained on newer manuals.

The novel’s delay is set by physics. My read: a model’s delay is set by cost, by how often a team can afford to pre-train, fine-tune or re-index. That makes staleness an engineering budget line, which the story never had to account for.

Speculation: if continual learning works at scale, the return stops being an event and becomes a steady process. The forgetting numbers above say that is not solved yet.

Follow the thread

• Mind the Gap (Lazaridou et al., 2021): the clearest early measurement that staleness grows with distance from the training period; the authors released streaming benchmarks built on WMT and arXiv.
• Dated Data (Cheng et al., 2024): a method to estimate a model’s effective cutoff per source, useful before you trust a model card’s date.
• Set the Clock (Zhao et al., 2024): code and data are public, so you can try aligning an open model to a given year.
• Right Knowledge, Wrong Answer (Hossain et al., 2026): the authors report that automatic detection of these conflicts remains unreliable, an open problem with a released benchmark to test against.
• FreshQA (Vu et al., 2023): the authors planned regular updates, so check for the latest version before you use it.
• Sparse memory finetuning (Lin et al., 2025): concrete forgetting numbers for full fine-tuning, LoRA and sparse updates, a fair baseline for any continual-learning claim.
• TIDE (Sobhani et al., 2026): built on customs regulations, but the version-in-force problem it measures applies to retrieval over any versioned documents.
• An open question with no paper attached: how should a deployed model estimate today’s date and the age of its own knowledge when nobody tells it?