Self-Preservation in Language Models

• Task Banks separate the specification of what to replay from the examples generated for it, enabling tests of combined targets and model-guided selection.Source: §1 · p. 1
• Held-out tasks exclude examples, labels, descriptions, and scores from bank construction and target selection, testing whether protection extends beyond replay targets.Source: §1 · p. 1
• Figure 1 illustrates new-task gains alongside prior-task losses, while replay from real prior-task data reduces those losses without eliminating them.Source: §1 · p. 1, Figure 1 · p. 2
• Generated replay reduced held-out losses, combined partial banks improved mean held-out Retention at all tested sizes, and incoming-aware selection helped over random only at 4B.The selection advantage did not appear at smaller tested sizes.Source: §1 · p. 2
• The incoming task and learning rate shape which prior capabilities degrade, while tracking tasks and previously correct answers reveals losses favorable averages can hide.Source: §1 · p. 2

2.1 Preservation Beyond Replay Targets

• Repeating examples for limited targets deepens preservation evidence but does not establish broader coverage; rehearsal amount and target choice are distinct design decisions.Source: §2.1 · pp. 2–3
• Task Banks test complementary target protection and incoming-task-informed selection using held-out tasks excluded from replay construction and target selection.Source: §2.1 · p. 3

2.2 Task Banks

• A Task Bank contains compact task entries with a name, behavior and input/output description, and optional demonstrations; replay can use one, a subset, or the full bank.Entries may describe overlapping skills; Task-Bank Self-Replay obtains replay answers from a frozen pre-update model.Source: §2.2 · p. 3
• Entries are selected before replay generation, with random, bank-only, and incoming-task-informed selectors; the selector, frozen reference model, and learner have distinct roles.Source: §2.2 · p. 3
• Replay is generated before the learner update and mixed with incoming examples; evaluation separates new-task learning from preservation on replay and held-out tasks.The fixed-allocation study generates answers to supplied questions, while other studies generate both questions and answers for supplied targets.Source: §2.2 · p. 3

3 Experiments

• The studies use Qwen models, MedQA and C-STANCE updates, four bank tasks, and four held-out tasks excluded from replay construction and target selection.Protocols include five-seed endpoint studies and separate controlled forgetting and fixed-example-allocation comparisons.Source: §3 · p. 3, §3.1 · p. 3

3.2.1 Aggregate accuracy can conceal lost behavior

• Across four held-out tasks, 3.56% of examples became incorrect and 3.61% became correct, leaving a +0.05 pp net accuracy change despite substantial answer turnover.Retention tracks the fraction of initially correct answers that remain correct.Source: §3.1 · p. 3, Figure 2 · p. 4, §3.2.1 · p. 4

3.2.2 Broad replay improves preservation outside its inventory

• Mean held-out accuracy change was +1.07 pp for Task-Bank Self-Replay and +0.86 pp for Replay, versus −3.99 pp for Single-Task Replay and −4.34 pp for SFT; incoming gains remained +27.74 to +27.86 pp.Source: §3.2.2 · p. 4, Figure 3 · p. 5

3.2.3 Combining supplied targets improves mean held-out Retention

• 88.70%, 93.28%, and 96.01% mean held-out Retention with combined banks exceeded the better partial bank by 4.42, 3.32, and 2.99 pp at 0.6B, 1.7B, and 4B.Incoming gains were 0.45, 0.60, and 0.28 pp below SFT, respectively.Source: §3.2.3 · p. 4, Figure 4 · p. 6

3.2.4 Incoming-task information can improve target selection

• At 4B, incoming-aware two-entry selection reached 95.69% mean held-out Retention versus 92.94% for random and 93.42% for bank-only selection, with no corresponding advantage at smaller sizes.Its incoming gain was 0.04 pp above random selection, and Retention was within 0.32 pp of the full bank.Source: §3.2.4 · p. 4, Figure 5 · p. 6

3.2.5 Preservation under a fixed example allocation

• With 59 replay rows, Task-Bank Self-Replay improved mean seed-wise worst held-out change from −1.56 to −0.68 pp, while MedQA gain fell from +1.17 to +0.49 pp.Rows and updates were matched, but token exposure and generation cost could differ.Source: §3.2.5 · p. 5, Figure 6 · p. 7

3.2.6 The incoming task and learning rate shape preservation

• At 1e-6, Qwen3.5-4B gained +4.7 pp on MedQA and +7.8 pp on C-STANCE, while prior-task changes differed; SVAMP rose about six points after MedQA and one after C-STANCE.Source: Figure 7 · p. 7, §3.2.6 · p. 7
• At 1e-7, Qwen3-1.7B C-STANCE achieved +12.01 pp incoming gain and 97.02% mean held-out Retention, versus +9.67 pp and 82.22% at 3e-6.In the separate 256-row, one-epoch sweep, 1e-5 reduced incoming-task accuracy across all six model–task combinations, indicating failed adaptation.Source: Figure 8 · p. 8, Table 1 · p. 9, §3.2.6 · p. 9

Forgetting and continual learning

• Replay and interference-aware methods rehearse prior experience or constrain disruptive updates, while MIR selects examples by anticipated interference.TRACE evaluates continual adaptation across diverse tasks; large-scale analyses separate losses of previously correct behavior from offsetting gains.Source: §4 · p. 9

Generated preservation evidence and task exploration

• Task Banks make the organization of supplied replay targets explicit, enabling study of how target composition and selection affect protection from generated replay.The main comparisons use supplied targets; model-proposed additions remain exploratory.Source: §4 · p. 9

5 Discussion

• Maintaining an inventory across updates remains open; proposed entries need demonstrated competence, distinct coverage, and incremental preservation value.The single-arm expansion pilot produced negative accuracy changes, while the two-update, one-seed pilot had mixed changes under a prescribed addition.Source: §5 · pp. 9–10
• The inventory pilots do not establish that answering a task predicts update damage or effective replay choices.The reported expansion pilot was single-arm, and the two-update pilot used one seed.Source: §5 · pp. 9–10
• SFT provides a fixed update rule for examining replay source and target choice, but findings may differ under other training signals.The discussion notes that on-policy training can reduce forgetting while some degradation remains, and its learning signals may change useful preservation targets.Source: §5 · p. 10

6 Conclusion

• Generated replay reduces forgetting beyond the bank, and combining partial banks improves mean held-out Retention; incoming-task-informed selection helps at 4B but shows no corresponding advantage at smaller sizes.Average accuracy can conceal lost previously correct answers, and replay can reduce new-task gains when it replaces incoming examples.Source: §6 · p. 10

A.1 Capability changes and loss of previously correct answers

• The reported worst held-out change is the average of seed-wise minima in Tables 1, 6, and 8, whereas endpoint profiles use minima of task means.Paired seed-level uncertainty cannot be reconstructed from the supplied endpoint aggregates.Source: §A.1 · p. 12
• F_j measures the fraction of initially correct answers lost; Retention is its complement, which pre/post accuracy alone cannot recover when new correct answers offset losses.Paired transitions also decompose net task change into newly correct answers and previously correct answers lost; these behavioral transitions do not establish erasure of underlying knowledge.Source: §A.1 · p. 12

A.2 Capability inventory and access boundary

• Held-out examples, labels, descriptions, and scores are excluded from bank construction and target selection, so evaluation tests protection beyond replay targets.The task split provides replay data only for replay tasks; held-out tasks are evaluated after the update rule is frozen.Source: §A.2 · p. 12, Table 2 · p. 12
• Controlled evaluations cap each fixed pre/post set at 512 examples and require at least 400 after filtering; fixed-allocation forgetting summaries include tasks with at least 200 initially correct answers.Other studies use their respective evaluation panels and aggregation rules.Source: §A.2 · p. 13

A.3 Model and optimization details

• The main sweep uses full-parameter AdamW, 256 selected input rows for one epoch, and shared optimizer settings across runs.The protection comparison fixes nominal rows and optimizer steps, but replay replaces incoming rows and realized token exposure varies; Qwen3-4B protection replication used TP=4 on v6e-4 with pinned weights.Source: §A.3 · p. 13, Table 3 · p. 13

A.4 Hyperparameter sweeps and condition selection

• The 36-run sweep crosses three Qwen models, two incoming tasks, three learning rates, and two seeds, reporting all runs rather than selecting by prior-task accuracy.Each run uses 256 selected rows for one epoch; the smaller-model follow-up also tests 3e-7 and 3e-6.Source: §A.4 · p. 13
• Figure 9 selects conditions with +4 to +10 pp C-STANCE gains, preferring more completed seeds, gains closer to +7 pp, then lower learning rates.The plotting rule is descriptive, not an independent test of a tuned algorithm.Source: §A.4 · p. 13
• The fixed-allocation replay comparison uses Qwen3-4B on MedQA at 1e-6, with two seeds and 16 updates; replay arms allocate 197 incoming and 59 replay rows.The condition was promoted after the SFT screen and is not a separately held-out development scenario.Source: §A.4 · p. 13
• The bank-composition and target-selection studies use Qwen3 models from 0.6B to 4B, 1,024 MedQA incoming examples, one epoch at 1e-5, and five seeds.These studies report paired Retention separately from endpoint accuracy summaries.Source: §A.4 · p. 13

Additional 0.6B allocations and evaluation panels

• At 3e-6, the additional Qwen3-0.6B C-STANCE calibration using 256 allocated rows averages two seeds with +7.1 pp incoming gain.This calibration is separate from the larger-allocation follow-ups.Source: §A.4 · p. 13
• At 3e-6, the 1,024-example 0.6B follow-ups report MedQA development gains of +7.42 and +8.59 pp with held-out mean changes of −3.81 and −5.57 pp; C-STANCE development gains are +5.47 and +9.38 pp with held-out changes of −2.69 and −3.56 pp.The C-STANCE incoming test gains are +11.33 and +11.91 pp; these are analyzed separately from the calibration and main sweeps.Source: §A.4 · p. 14

A.5 Five-seed single-task endpoint study

• Table 4 reports five-seed Qwen3-4B endpoint scores after one epoch at 1e-5 for C-STANCE; the study reports endpoints, not within-update trajectories.The pre-update score is shown separately, and other entries are means ± SE.Source: Table 4 · p. 14, §A.5 · p. 14

A.6 Confidence and behavioral probes of competence

• Better item-level calibration and discrimination do not necessarily yield a better task-level competence ranking; larger probe budgets reduce accuracy-estimation error.Table 5 compares confidence metrics across seven multiple-choice capabilities and probe-based MAE on disjoint examples.Source: §A.6 · p. 14, Table 5 · p. 15

A.7 Supplied-question Task-Bank Self-Replay protocol

• In the fixed-allocation protocol, a frozen model answers supplied questions using two verified within-task demonstrations, while the question’s source answer is withheld.Generation checks enforce formatting and access boundaries, not correctness; archived manifests connect replay data to training runs.Source: §A.7 · p. 15

B Additional Related Work

• Task Banks organize replay targets, extending prior work on generated tasks by testing whether replay protects held-out behavior rather than merely discovering or proposing capabilities.The bank-expansion pilot is exploratory; establishing its preservation value against additional replay within the existing inventory remains open.Source: B Additional Related Work · p. 16

C Task Bank Interfaces and Auxiliary Diagnostics

• Replay interfaces vary in supplied information, and held-out evaluation excludes examples, labels, descriptions, and scores from bank construction and target selection.Some studies generate questions and answers, while the fixed-allocation study generates answers to supplied questions.Source: Table 6 · p. 16, §C.1 · p. 17

C.2 Estimating competence from pre-update predictions

• Confidence-based diagnostics compare confidence with observed correctness, while a limited labeled probe budget estimates task accuracy against a disjoint remainder.The probe split is repeated 100 times at each budget.Source: §C.2 · p. 17

C.3 Replay mixture and fixed-example allocation

• Replay replaces some incoming examples at fixed nominal allocation; replay weights specify example allocation, while tokenization and loss normalization also affect training loss.The fixed-allocation comparison uses 59 replay rows and 197 incoming rows out of 256, with optimizer-step count fixed.Source: §C.3 · p. 17

C.4 Criteria for interpreting target discovery

• Target proposals require independent evidence of distinct behavior and competence, plus controlled comparisons to establish incremental preservation value.Structural validity and exact-duplicate filtering do not establish semantic novelty; a single operating point does not establish an acquisition–retention frontier.Source: §C.4 · p. 18

C.5 Negative bank-aware expansion pilot

• The expanded-bank pilot’s mean held-out accuracy change is −17.92 pp, but this single-arm result does not measure expansion’s effect relative to a fixed bank.It tests Qwen3-1.7B on MedQA with seed 1 and 512 matched examples per endpoint; the pilot also loses incoming-task accuracy.Source: Table 7 · p. 18, §C.5 · p. 18
• SVAMP extraction failures constrain interpretation of the pilot’s accuracy loss: no valid final number was extracted for 92 base and 303 post-update responses out of 512.The unchanged extractor’s failures prevent attributing the accuracy loss solely to latent mathematical competence.Source: §C.5 · p. 18

C.6 Prompt templates for replay, target selection, and exploration

• Selection chooses existing bank entries before replay generation, whereas exploration proposes descriptions; structural validity and exact-duplicate filtering do not establish semantic novelty or competence.The selection comparison uses two of four targets; its templates were not tested for robustness to alternative wording.Source: §C.6 · pp. 19–20

C.7 Complete results at fixed training allocation

• At fixed allocation, generated replay reduces measured forgetting while using 59 of 256 input rows, thereby replacing new-task examples.Table 8 describes Qwen3-4B on MedQA at 1e-6 for one epoch, with two seeds per arm; the post-hoc selector is analysis-only.Source: §C.7 · p. 20, Table 8 · p. 21

D Full Reported Learning-Rate Sweeps

• The complete sweep separates negligible adaptation, useful new-task gain, and damage to both new and prior tasks; destructive conditions are stress tests, not the sole explanation for losses.Figure 10 retains every measured learning rate, so prior-task changes can be interpreted alongside new-task gain.Source: D Full Reported Learning-Rate Sweeps · p. 21, Figure 10 · p. 22
• Selected C-STANCE operating points show that positive average changes can conceal individual task losses, but do not isolate model scale from generation and learning rate.The 0.6B panel uses a separate 256-row calibration, not the 1,024-row follow-up.Source: Figure 9 · p. 21
• Learning-rate sensitivity depends on the model and incoming task, and the sparse grid establishes neither a scaling law nor a universal safe rate.Figure 11 shows measured capability profiles across the reported SFT conditions.Source: D Full Reported Learning-Rate Sweeps · p. 23, Figure 11 · p. 23

E Model Scale and Broad Preservation

• Task-Bank Self-Replay improves the four-task average held-out accuracy by 0.61, 0.31, and 0.76 pp at 0.6B, 1.7B, and 4B, respectively.Its worst-task accuracy losses are 0.52, 1.03, and 0.39 pp at those sizes, so average gains do not imply that every task is preserved.Source: Figure 12 · p. 24, E Model Scale and Broad Preservation · p. 24
• Replay leaves average held-out accuracy losses of 2.00, 1.27, and 1.25 pp at 0.6B, 1.7B, and 4B, respectively, with worst-task losses of 2.81, 1.91, and 1.69 pp.Replay losses decrease across these measured sizes but remain positive at each size.Source: Figure 12 · p. 24, E Model Scale and Broad Preservation · p. 24
• Under SFT, average held-out accuracy loss falls from 6.01 to 4.38 to 3.50 pp, while worst-task loss falls from 8.30 to 5.94 to 5.40 pp across 0.6B, 1.7B, and 4B.Task-Bank Self-Replay does not follow a monotonic scaling pattern in these results.Source: Figure 12 · p. 24, E Model Scale and Broad Preservation · p. 24

Sequential Preservation Pilot

• The two-update pilot used Qwen3-0.6B, learning MedQA before C-STANCE, with one seed and a learning rate of 3e-6.Each stage used one epoch of 256 rows; replay arms mixed 197 new-task rows with 59 replay rows.Source: F Sequential Preservation Pilot · p. 24
• Replay generation used 64 candidate prompts per task, without source answer labels or held-out data.The fixed and expanding generated-bank arms shared the first-stage checkpoint.Source: F Sequential Preservation Pilot · p. 24
• The pilot does not establish continual self-improvement or autonomous capability discovery because task additions were prescribed and only two stages and one seed were tested.Longer-run accumulation remains beyond this pilot's scope.Source: F Sequential Preservation Pilot · p. 24
• Adding MedQA improved worst held-out change but worsened worst replay-task change relative to the fixed bank in the one-seed pilot.The expanded arm remained above its initial MedQA score but fell slightly below its own first-stage score; stage-relative accuracy changes do not track individual previously correct answers.Source: Table 9 · p. 25

Further Discussion and Limitations

• Combining partial banks improves mean held-out Retention, while incoming-aware selection helps at the tested 4B operating point.The comparisons require considering new-task gains, individual-task changes, and Retention on tasks excluded from replay.Source: G Further Discussion and Limitations · p. 25
• The experiments cover Qwen checkpoints, two incoming tasks, eight prior capabilities, and four held-out tasks, so they do not establish preservation of the full pretrained repertoire.Most studies concern one update; the two-stage, one-seed pilot has mixed outcomes and does not establish sustained autonomous preservation.Source: G Further Discussion and Limitations · p. 25
• Selected model comparisons vary learning rate and model generation, while seed counts and endpoint aggregates limit uncertainty analysis.Under a fixed example allocation, improved preservation accompanies smaller new-task gains as replay replaces incoming examples; nominal row and optimizer-step matching does not equalize token exposure or generation overhead.Source: G Further Discussion and Limitations · p. 25
• Task-Bank Self-Replay depends on supplied targets and does not eliminate supervision; replay can reproduce model errors, and supplied-question replay relies on verified demonstrations.The pilot does not establish whether newly proposed targets improve preservation over a fixed partial bank, or robustness across task omissions, prompt variants, and task-addition rules.Source: G Further Discussion and Limitations · p. 25