PaperlayerView exact revision on arXiv

Source-linked AI summary

Self-Generated Feedback Destabilizes Test-Time Training: A Causal Decomposition of Long-Horizon Adaptation

Cheng Luo, Bing Li, Bernard Ghanem

arXiv:2610.05076v1cs.CLcs.AI

TL;DR

Test-time training can store information in model weights during inference, but persistent updates on self-generated text worsen prediction on independent human-written text. Matched comparisons trace this feedback pathway and evaluate candidate updates on independent text before commitment. Across the evaluated settings, the comparisons distinguish damaging self-writing from useful real-text adaptation.

  • Problem

    The paper asks which part of persistent self-writing feedback makes updates from a model’s own output fail to transfer to independent text.

    Source: §1 · p. 1

  • Method

    Matched comparisons use Fixed Generation, Recorded Replay, and paired one-update branches to separate feedback into future training text, effects of reading, and persistent update costs.

    Source: §5 · p. 5, §8 · p. 10

  • Results

    Fixed Generation removes 98.3% and 98.8% of the matched Closed Loop gap at 125M and 760M, while Recorded Replay and paired updates locate additional harm in reading degraded tokens and poor transfer.

    Source: §8 · p. 10

  • Takeaways & Limitations

    Settlement compares current and proposed states on independent real text before commitment, nearly matches Writes Off on generated streams, and retains useful updates on real-text streams.

    Source: §8 · p. 10

Abstract

from arXiv · show

Test-time training (TTT) lets a model store information in its weights during inference. When the model learns from its own output, however, each update also changes the model that generates the next training example. Across 128K-token streams, retaining generated-text updates worsens prediction on independent human-written text with three TTT-E2E model configurations (labeled 125M, 760M, and 3B). The same failure occurs when Adam updates Qwen3-4B's existing weights. The same update mechanisms can improve on real text, so writing itself is not the failure. Three matched comparisons trace the causal pathway. Fixed Generation removes over 98% of the damage at 125M and 760M by using a frozen model to generate training chunks. Recorded Replay separates the loss caused by reading degraded text from the additional loss stored by updating on it. A paired one-update comparison then shows the local conflict: an update predicts its source better but new real text worse. This cost grows after Closed Loop adaptation, with a few trajectories accounting for most large failures. Finally, Settlement evaluates the candidate state on independent real text before commitment. It leaves mean endpoint gaps of 0.07 and -0.02 nats at 125M and 760M while retaining real-text adaptation. These results motivate checking prediction on independent evidence before retaining an update.

1 INTRODUCTION

TTT can preserve information in fast weights, but persistent learning from self-generated text can damage prediction on independent human-written text. The introduction frames matched comparisons that trace the feedback pathway and evaluate updates using independent text.

1 INTRODUCTION

  • TTT fast weights can retain information after source text leaves attention, which makes test-time training attractive for long documents, streams, and long-running agents.The paper calls processing text without changing weights reading, and retaining its update writing.Source: §1 · p. 1
  • Generated-text writes worsen independent real-text prediction across TTT-E2E configurations labeled 125M, 760M, and 3B, and the failure also appears when Adam updates Qwen3-4B’s existing weights.The same update rules can improve prediction when trained on real text, so writing itself is not the failure.Source: §1 · p. 1
  • Fixed Generation removes over 98% of the damage at 125M and 760M by having a frozen model generate each training chunk while a separate model learns from it.Recorded Replay holds the text fixed to distinguish harm from reading degraded tokens from the additional persistent weight cost of updating on them.Source: §1 · p. 2
  • Settlement evaluates a candidate state on independent real text before commitment, retaining real-text adaptation while leaving mean endpoint gaps of 0.07 and −0.02 nats at 125M and 760M.It also remains stable when source labels are noisy.Source: §1 · p. 2
  • Fixed Generation, Recorded Replay, and paired one-update branches separate effects of changes to future training text, reading degraded tokens, and retaining updates.Together, these matched comparisons locate a state-dependent failure concentrated in a few trajectories.Source: §1 · p. 2

2 PRELIMINARIES: PERSISTENT TEST-TIME TRAINING

Test-time training updates part of a model while it processes input, allowing information to persist in weights beyond the attention window. TTT-E2E applies matching next-token updates during training and inference-time updates to final-layer MLP blocks.

2 PRELIMINARIES: PERSISTENT TEST-TIME TRAINING

  • TTT updates part of the model on the input it is processing, allowing information to remain in weights after the corresponding tokens leave the attention window.An ordinary language model keeps the same parameters throughout inference.Source: §2 · p. 2
  • TTT-E2E addresses the mismatch from inference-only updates by applying the same next-token updates during training and meta-learning an initialization suited to those updates.At inference, it updates MLP blocks in the final quarter of a sliding-window Transformer.Source: §2 · p. 2
  • Reading means a chunk remains available through attention, while writing means retaining its gradient update so it changes fast weights Wt.In the Closed Loop, Wt generates chunk xt, an update produces Wt+1, and that state helps generate the next chunk; Writes Off discards generated-text updates while retaining updates from the initial real-text prefix.Source: §2 · p. 2

3 EXPERIMENTAL SETUP

The experiments compare persistent generated-text updates with matched controls over 128K-token streams and evaluate prediction on independent real text. Their outcomes track within-condition deterioration and the additional change relative to Writes Off.

3 EXPERIMENTAL SETUP

  • The cross-scale comparison uses the same six books, five seeds, 128K horizon, and batch width eight at each scale.Models include TTT-E2E configurations labeled 125M and 760M and a released 3B model; most causal experiments use the 125M model.Source: §3 · pp. 2–3
  • Each 128K stream begins with 8K real-text tokens, then generates 105 chunks and receives 15 later real-text evaluations plus an initial reference.Evaluation passages are from later, disjoint regions of the same PG-19 book, and evaluation restores the preceding weights and attention state.Source: §3 · p. 3
  • For condition c, Dc is the first-to-last change in independent-real-text NLL; positive values indicate worse prediction, and Hc subtracts the Writes Off change Doff.Hc > 0 means worse than Writes Off, while Hc < 0 means better; in online generation, the contrast includes updates and the later text they cause the model to generate.Source: §3 · p. 3

4 LONG-HORIZON PERSISTENT SELF-WRITING DEGRADES PREDICTION

Across long streams, retaining generated-text updates worsens prediction on independent real text, with the effect accumulating beyond the earlier evaluation horizon. Decoder changes, real-text exposure, and ordinary Adam updates reveal boundaries and qualifications to this result.

4 LONG-HORIZON PERSISTENT SELF-WRITING DEGRADES PREDICTION

  • 3.01, 6.00, and .40 nats: retaining generated-text updates adds this much loss over Writes Off at 125M, 760M, and 3B, respectively.Every interval excludes zero; the direction replicates across scales, but the magnitude does not increase monotonically with parameter count.Source: §4 · p. 3, Table 1 · p. 3
  • The policies remain close within the earlier 8K+8K window; only after many additional self-writes does Closed Loop rise while Writes Off stays nearly flat.The failure is cumulative rather than a large penalty from the first update.Source: §4 · p. 3, Figure 1 · p. 4
  • A full-support sampler reduces the 128K endpoint gap from 3.010 to 1.452 nats, while an eight-chunk reset leaves a .175 gap but preserves only 44.1% of real-text adaptation benefit.Lower temperature, typical sampling, and a repetition penalty alter degradation without providing a reliable boundary between safe and harmful writes.Source: §4 · p. 4
  • With 33 real chunks, evenly spaced exposure leaves about 6% of the original harm, compared with about 60% when the same chunks are grouped into bursts.Even spacing limits consecutive generated chunks to three; bursts permit stretches of 12.Source: §4 · p. 4
  • At learning rate 10^-4, persistent self-writing raises Qwen3-4B real-text NLL by 1.231 nats relative to Writes Off, while the same update lowers NLL by .168 nats on real text.At 10^-5, the interval includes zero.Source: §4 · p. 4, Table 2 · p. 4
  • On eleven identical teacher-forced real-text streams, retaining updates lowers mean NLL by .0340 nats relative to No Updates.Because the input text is fixed, the gain comes from learning rather than changing future inputs.Source: §4 · p. 5

5 A CAUSAL DECOMPOSITION OF THE FEEDBACK PATH

Matched comparisons trace how self-generated training text, reading degraded tokens, and retaining updates contribute to harm on new real text. Fixed Generation, Recorded Replay, and paired one-update tests isolate distinct parts of this pathway.

5.1 KEEP LEARNING, BUT FIX THE GENERATOR

  • Fixed Generation moves farther from the initial weights than Closed Loop (.01387 versus .00857) yet causes far less damage.Real-Text Learning moves farther still (.01759) and improves NLL by .3544 nats [.1236, .6048].Source: Table 3 · p. 5, Figure 2 · p. 5, §5.1 · p. 6

5.2 RECORDED REPLAY: PROCESS THE SAME TEXT WITH AND WITHOUT WRITING

  • Replay separates harm from keeping degraded tokens in attention from the additional persistent cost of writing on those same tokens.The paired read-only versus read-and-write comparison holds the recorded content fixed.Source: Table 4 · p. 6, §5.2 · p. 6

5 A CAUSAL DECOMPOSITION OF THE FEEDBACK PATH

  • Feedback changes later training text, while reading degraded tokens and retaining updates can each contribute additional harm.The comparisons locate pathways to damage rather than assigning additive shares of one total loss.Source: §5 · p. 5, Figure 3 · p. 6, §5.2 · p. 6

5.3 ONE UPDATE CAN FIT ITS SOURCE AND HURT ANOTHER INPUT

  • Retaining an update lowers mean source-chunk NLL but raises mean NLL on the next real passage in every tested history.Fixed Generation has source-fit improvement 0.1449 and real-text cost 0.0018; Closed Loop has source improvement 0.0354 and cost 0.0381.Source: Table 5 · p. 7, §5.3 · p. 7
  • Gradient opposition and predicted loss change both rank measured real-text harm, with stronger generated–real conflict accompanying greater harm.The reported associations are rank correlations, not NLL increases, and use the real passage on which harm is measured.Source: Table 6 · p. 7, §5.3 · p. 7

6 WHICH UPDATES SHOULD BE RETAINED?

Late write costs are heterogeneous, with a small subset of trajectories accounting for high-cost passages. Settlement evaluates candidate updates on independent real text, preserving measurable adaptation while keeping endpoint gaps near zero in the reported language-model settings.

6.1 DAMAGE AMPLIFIES IN DEGRADED STATES BUT REMAINS HETEROGENEOUS

  • At the final position, mean write cost rises to 0.7512 while the median is 0.0064; only 12 of 72 late costs exceed 0.5 nats.All 12 high costs come from four of 24 source sequences, so position alone cannot identify a safe write.Source: §6.1 · p. 8, Figure 4 · p. 8
  • Milder controls leave 54% and 48% of the Closed Loop gap, while an eight-chunk reset removes 94.2% but retains only 44.1% of useful adaptation.These controls do not evaluate whether a particular update helps prediction beyond its source.Source: §6.1 · p. 8

6.2 EXTERNAL VALIDATION MEASURES WHETHER AN UPDATE TRANSFERS

  • Settlement leaves endpoint gaps of .07 nats at 125M and −.02 at 760M, with both intervals including zero.On an all-real stream, it also retains measurable learning relative to Reject All and Write All.Source: Figure 5 · p. 9, §6.2 · p. 9
  • With corrupted source labels, Source Masking accepts generated writes and rejects real ones, whereas Settlement has lower mean NLL from 21.2% observed error.Settlement’s paired intervals exclude zero at evaluated error rates of 38.9% and above.Source: §6.2 · p. 9, Figure 5 · p. 9
  • On WebShop, Settlement reaches .1853 exact success versus .1053 for Closed Loop and .1600 for shared Writes Off; on ALFWorld it achieves 88.1%–94.0% unseen-task success.Settlement accepts 10 of 180 WebShop candidate blocks and retains 19%–33% of ALFWorld update blocks across the two scales.Source: §6.2 · p. 9, Figure 5 · p. 9

7 RELATED WORK

Related work spans writable inference-time state, test-time adaptation and stability, generated-data feedback, and update selection. This paper distinguishes its focus as feedback within one adapting model’s long text stream and uses controlled comparisons plus Settlement to study it.

Writable inference-time state

  • Fast weights, linear attention, dynamic evaluation, and TTT variants provide writable inference-time state; this work studies when that state generates its next training example.The cited TTT-related systems include TTT layers, TTT-E2E, Titans, and In-Place TTT.Source: §7 · pp. 9–10

Test-time adaptation and stability

  • Prior test-time adaptation uses deployment inputs, self-supervised losses, entropy minimization, or retrieved neighbors, while stability methods limit erroneous pseudo-label accumulation.The paper distinguishes its setting because the adapted model produces later inputs within the stream.Source: §7 · p. 10

Generated-data feedback and update selection

  • Recursive synthetic-data studies examine successive model generations, while decoding work addresses repetitive continuations; this paper studies one adapting model in one stream.Its decoder and replay controls test why diversity and synthetic origin alone do not determine persistent cost.Source: §7 · p. 10
  • Continual-learning, validation-based reweighting, and self-improvement methods constrain or select updates; aTTT and VANE study feedback or validation in other settings.Here, Settlement evaluates the accumulated fast-weight state on independent real text before retaining an update.Source: §7 · p. 10

8 DISCUSSION AND CONCLUSION

The feedback loop makes some trajectories persistently harmful, while fixed-text controls separate harm from reading degraded tokens from the additional cost of retaining their updates. Settlement screens proposed states on independent real text, reducing endpoint gaps while preserving real-text adaptation.

Concentrated failures

  • Four of 24 source sequences account for every fixed-receiver write cost above 0.5 nats, showing that large late failures are concentrated rather than universal.Affected source–receiver pairs remain above that threshold at later sampled positions.Source: §8 · p. 10

Causal decomposition

  • Breaking feedback removes 98.3% and 98.8% of the matched Closed Loop gap at 125M and 760M, respectively.Recorded Replay finds that reading degraded tokens raises real-text loss and retaining their updates raises it further, including at 3B.Source: §8 · p. 10

Reducing damage

  • With 31% real chunks, evenly spaced exposure leaves .0694 nats of excess harm versus .7155 for bursts; Settlement instead evaluates proposed states on independent real text.Settlement leaves mean endpoint gaps of .07 and −.02 nats at 125M and 760M while retaining real-text adaptation.Source: §8 · p. 10

A EXPERIMENTAL SCOPE AND IMPLEMENTATION

The appendix reports controlled comparisons of feedback, decoder policies, real-text exposure, and update retention, with boundaries on model scope and reproducibility. Across these comparisons, feedback and update retention add harm, while resets and real-text cadence can limit it at a cost to useful adaptation.

A EXPERIMENTAL SCOPE AND IMPLEMENTATION

  • The scope uses shared controlled settings and book-based bootstrap intervals; screening removes a visible contamination anomaly but does not prove all contamination absent.The 760M and 3B training lineages differ, so their damage magnitudes do not establish a scaling law. Checks do not establish bitwise reproducibility of all stochastic generation trajectories.Source: A EXPERIMENTAL SCOPE AND IMPLEMENTATION · pp. 13, 15, §A.3 · p. 15

A.2 FEEDBACK AND DISPLACEMENT CONTROL

  • Closed Loop exceeds Writes Off by 2.8308 [1.8171, 3.8598], while Fixed Generation exceeds it by only .0707 [.0510, .0924].Fixed Generation and Real-Text Learning move farther from initialization than Closed Loop while causing far less harm, so parameter displacement alone is insufficient to explain the closed-loop loss.Source: §A.2 · p. 15, Table 8 · p. 15

B.1 FIXED GENERATION CONFIRMATION AT 125M AND 760M

  • Fixed Generation leaves only a small fraction of the Closed Loop endpoint gap at both 125M and 760M.The separate 760M confirmation uses the same six canonical books and five seeds per book; the table caption identifies the comparison scope but gives no cell values.Source: §B.1 · p. 16, Table 10 · p. 16

B.2 DECODER AND RESET CONTROLS CHANGE THE RATE OF DAMAGE

  • Resetting every eight chunks leaves an endpoint gap of .1749 [.084, .321], versus 1.9218 [.949, 3.158] for resets every 32 chunks.The eight-chunk reset preserves 44.1% of the persistent real-text adaptation benefit; repetition penalties and decoder choices change damage without identifying safe writes.Source: §B.2 · pp. 16–17, Table 11 · p. 16

B.3 EXTERNAL-TEXT EXPOSURE BOUNDS THE FAILURE REGIME

  • At 31% real-text exposure, grouping the same 33 real slots increases the longest self-generated stretch from 3 to 12 and raises excess NLL change by an order of magnitude.Evenly spaced real-text exposure reduces excess harm while preserving a positive real-text adaptation benefit.Source: §B.3 · p. 17, Table 12 · p. 17

C.1 FIXED GENERATION AND RECORDED REPLAY

  • Recorded Replay gives 1.606 nats of excess without writes and 3.874 with writes, an additional 2.268 nats from retaining updates on identical recorded tokens.Comparing live generation with replay does not isolate additional feedback because the text differs.Source: §C.1 · p. 18

C.2 3B FIXED-TEXT REPLAY CONFIRMATION

  • For the 3B fixed-text replay, the paired write-minus-read change is J = 0.1927 [0.1043, 0.3272] under source/receiver clustering.The paired trajectories have equal initial NLL and identical received-token hashes.Source: §C.2 · p. 18, Table 13 · p. 18

C.3 PAIRED ONE-UPDATE COMPARISON

  • The one-update experiment pairs keep-versus-discard branches from shared starting states, current text, and matched random numbers.Table 14 defines negative source NLL change as improved source fitting and positive real NLL change as harm to independent prediction; the supplied passages give no cell values.Source: §C.3 · p. 18, Table 14 · p. 18

C.4 GRADIENT CONFLICT AND ONE-UPDATE TRANSFER COST

Settlement evaluates pending updates on independent real text before commitment, while the preceding analyses trace local update conflict and test partial repairs. Across the reported scales, it improves over Closed Loop while retaining real-text adaptation.

C.4 GRADIENT CONFLICT AND ONE-UPDATE TRANSFER COST

  • After Closed Loop history, generated–real gradient cosine is −.093 at 125M and −.103 at 760M, while real–real cosine remains positive.After Writes Off history, generated–real cosine is −.005 and −.004, respectively; real–real cosine is .125 and .097.Source: §C.4 · p. 20

C.5 IN-PLACE ADAM: DOSE AND HISTORY ARE SEPARATE CONSIDERATIONS

  • The larger-rate in-place Adam excess is 1.231 [.580,2.350], versus .026 [-.071,.122] for the smaller rate.Both learning rates improve independent real-text prediction; the 10−4 configuration can help on real text while harming prediction on a self-generated stream.Source: Table 17 · p. 20, §C.5 · p. 20

D.1 A MINORITY OF SOURCE PASSAGES DRIVES THE LATE MEAN

  • All cases above .5 nats come from four of 24 source rows, and those source–receiver pairs remain above that threshold at later sampled positions.The early-to-late repetition-bin change is unresolved at .0587 [−.0382, .3286], so position and repetition indicate risk but do not identify every harmful passage.Source: §D.1 · pp. 20–21

D.2 REPETITION WEIGHTING MITIGATES BUT DOES NOT REMOVE DAMAGE

  • Repetition weighting improves on random and uniform dose assignment by .4921 and .9224 nats, respectively, but remains 1.5637 nats above Writes Off.The result supports repetition as a useful symptom, not a sufficient admission rule; the tested partial repairs do not determine whether a particular update transfers.Source: D FROM HETEROGENEOUS HARM TO PARTIAL… · p. 20, §D.2 · p. 21, Table 20 · p. 21

D.3 SMALLER WRITES EXPOSE A STABILITY–ADAPTATION TRADE-OFF

  • Reducing the update scale delays and reduces generated-stream damage, but also reduces real-text benefit, tracing a stability–adaptation frontier.Uniform scaling does not separate useful from harmful writes.Source: §D.3 · p. 21, Table 21 · p. 21

D.4 REAL READ-ONLY CONTEXT HELPS BECAUSE OF ITS CONTENT

  • Read-only real text lowers final NLL by 1.5693 nats versus a no-op schedule, yet ends 1.3979 nats above Writes Off.It outperforms random, shuffled-real, and Fixed Generation anchors in direct paired comparisons, making it a partial repair rather than a complete selection rule.Source: §D.4 · p. 22, Table 22 · p. 22

E.1 SEQUENTIAL VALIDATION AND COMMITMENT

  • Settlement sequentially evaluates each pending increment on eligible real text against the accumulated accepted state, then commits accepted changes.It uses external validation text, with retained real text as a fallback; evaluation passages are separate from admission text.Source: E EVALUATING TRANSFER WITH INDEPENDENT… · p. 22, §E.1 · pp. 22–23

E.2 CROSS-SCALE OUTCOMES

  • At 125M Settlement improves on Closed Loop on 11 of 12 books, and at 760M on all four; real-text acceptance is 83% and 100%, respectively.Residual intervals include zero and do not establish superiority to Writes Off.Source: §E.2 · p. 23, Table 23 · p. 23

E.3 COMPONENT ABLATION

The component comparisons distinguish external validation from weaker update and context controls. Further tests show that sequentially validating the committed state preserves adaptation, and that validation can remain effective when source labels are corrupted.

E.3 COMPONENT ABLATION

  • With Settlement, every tested component combination finishes within .04 nats of Writes Off; read-only real context lowers NLL by a further .057 nats [.024, .089].Without validation, quarter-sized updates and periodic read-only real context each leave a visible gap; the incremental effect of Quarter Dose is unresolved.Source: §E.3 · p. 23, Table 24 · p. 23

E.4 ADMISSION AND ADAPTATION IN ONE ALL-REAL STREAM

  • Settlement admits 78.8% of real-text proposals, with paired benefit over Write All of .0083 [.0020, .0161].The result shows that low admission on self-generated streams is not a fixed tendency to reject all learning.Source: §E.4 · p. 23, Table 25 · p. 24

E.5 PRESERVING ADAPTATION AND VALIDATING THE COMMITTED STATE

  • Sequential validation retains about 56% of ordinary real-text benefit and avoids the joint-commit loss, whose terminal drift is .0862 versus .0176 for ordinary sequential writing.Joint commit approves candidates individually at a shared base before deploying their unevaluated sum; Settlement accepts 98–101 of 112 real candidates with point benefit close to ordinary writing.Source: §E.5 · p. 24, Table 26 · p. 24, Figure 6 · p. 24

F EXTERNAL VALIDATION WHEN SOURCE LABELS ARE CORRUPTED

  • As observed-label error rises, Source Masking NLL increases from 3.8072 to 4.2544, while Settlement stays between 3.8196 and 3.8286 and is lower in mean from 21.2% error.At 100% label error, Settlement admits 120/120 real candidates and 0/219 self-generated candidates when regrouped by true source.Source: F EXTERNAL VALIDATION WHEN SOURCE… · p. 25, Table 27 · p. 25

G WEBSHOP FEEDBACK AND TASK-METRIC VALIDATION

The WebShop experiment tests whether feedback from an agent’s own actions affects task completion and whether update validation can reduce that harm. Fixed Generation and Settlement both beat Closed Loop in every paired seed, with Settlement achieving the highest reported exact success.

G WEBSHOP FEEDBACK AND TASK-METRIC VALIDATION

  • A common source label cannot distinguish WebShop action updates: it would either reject all of them as Writes Off or retain all as Closed Loop.Settlement instead retains a candidate only when its mean validation reward is strictly higher.Source: G WEBSHOP FEEDBACK AND TASK-METRIC… · p. 25
  • The policy comparison uses 900 episodes and the same 150 test-partition goals; validation and final evaluation select tasks with different seeds, but the runner does not enforce disjoint goal sets.Exact success means reward ≥1, not merely making a purchase.Source: G WEBSHOP FEEDBACK AND TASK-METRIC… · p. 26
  • Settlement reaches exact success .1853 ± .0202, and Fixed Generation reaches .1800 ± .0089 versus Closed Loop’s .1053 ± .0597; both beat Closed Loop in every paired seed.Settlement accepts 10 of 180 candidate blocks, and final evaluations use a learner with writing disabled.Source: G WEBSHOP FEEDBACK AND TASK-METRIC… · pp. 25–26, Table 28 · p. 25

H ALFWORLD: KEEPING THE AGENT ABLE TO COMPLETE TASKS

In ALFWorld, halving update scale still leaves two Closed Loop seeds with zero unseen-task success. Settlement retains some updates and produces higher mean success than Writes Off in these three-seed results, without establishing a reliable overall gain.

H ALFWORLD: KEEPING THE AGENT ABLE TO COMPLETE TASKS

  • The agent attempts 140 seen tasks followed by 134 unseen tasks, with learned weights persisting between tasks and a 30-action limit per task.Each task is scored before its update.Source: H ALFWORLD: KEEPING THE AGENT ABLE TO… · p. 27
  • Settlement checks the current model and temporary accumulated updates on the same 20 validation tasks every 25 tasks, retaining the copy only if it completes more tasks.It also checks at task 140; a block batches updates accepted or rejected together.Source: H ALFWORLD: KEEPING THE AGENT ABLE TO… · p. 27
  • Settlement completes 88.1%–94.0% of unseen tasks and retains 19.4%–33.3% of update blocks, while two halved-scale Closed Loop seeds have zero unseen success.Across all 274 tasks, Settlement mean success is 83.2%–83.8%, versus 82.1% for Writes Off; these results do not establish a reliable overall gain.Source: H ALFWORLD: KEEPING THE AGENT ABLE TO… · p. 27, Table 29 · p. 27, Figure 7 · p. 27

Related papers

Loading 2610.05076v1…