acceptodds
Under review as a conference paper at ICLR 2027

Mitigating Catastrophic Forgetting in LLM Continual Learning through Attractor-Informed Latent Transition Alignment

10 views

Abstract

Continual learning enables large language models to acquire new knowledge while preserving the computational structures that support their existing capabilities. In replay-free continual learning, however, original pre-training data are unavailable, and model updates may disrupt these structures and cause catastrophic forgetting. Existing methods mainly preserve model parameters, output distributions, or isolated hidden states, while paying less attention to cross-layer transformations. We propose Latent Transition Alignment (LTA), which uses cross-layer latent transitions as a preservation target. In selected intermediate Transformer windows, our latent-dynamics analysis identifies coherent transition fields and locally attracting fixed points. We then introduce an offline dual-anchor surrogate that connects observed transitions to the residual field of a latent self-map. Under exact reconstruction, the observed latent transition equals the residual-field vector at its start state. Under approximate reconstruction, their discrepancy is bounded. Online, LTA aligns transition directions and applies one-sided magnitude control without online field fitting or fixed-point iteration. In single-domain learning, LTA outperforms all evaluated baselines in retention while matching SFT in new-domain learning; relative to Output KL, it reduces accumulated first-domain forgetting by 8.20 percentage points and improves final QASC accuracy by 7.30 percentage points in multi-domain continual learning; under fixed-protocol transfer to Qwen2.5-3B, it requires no scale-specific retuning and reduces forgetting by 63.1% while improving QASC accuracy by 3.33 percentage points.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Large Language Models, Replay-Free Continual Learning, Catastrophic Forgetting, Latent Dynamics, Semantic Flow, Attractor Dynamics, Latent Transition Alignment.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qM5fOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes latent transition alignment (LTA) for replay-free continual learning of language models. It uses frozen encoders to compare cross-layer transitions between a reference model and the model being trained. The loss aligns transition directions and penalizes excessive magnitude growth. An offline dual-anchor construction and latent fixed-point analysis motivate this objective. Experiments report lower reference-NLL drift and better first-domain retention across several task orders, with an additional transfer experiment on Qwen2.5-3B.

Critical strengths
  • LTA offers a clear alternative to matching individual hidden states. It allows common endpoint shifts and some magnitude changes, without adding inference-time components.

  • The structural baselines and component ablations offer useful controls.

Critical weaknesses
  • The novelty over existing cross-layer distillation is not that clear.

FDD (Gong et al., 2025) already aligns cross-layer changes using cosine distance in its layer-delta loss (Eq.~7). LTA uses the same form of angular penalty, but applies it in a learned latent space instead of the LM-head log-probability space. The dual-anchor construction, reference-dependent weighting, and one-sided magnitude term distinguish LTA from FDD. However, this work does not clearly establish how much these additions improve upon prior directional alignment.

Transition-MSE matches complete vectors and hence doesn't represent FDD's cosine-based objective. The Direction-only ablation is useful, but it retains LTA's dual-anchor encoder and doesn't isolate the benefit of that construction. Moreover, Table 5 reports nearly identical mean reference drift for Direction-only and full LTA, at 0.0221 and 0.0220, respectively. This provides limited evidence for the added magnitude term. A matched FDD-style baseline and an encoder ablation without the second reconstruction anchor would help show whether LTA offers a substantial methodological advance over existing transition alignment.

  • The attractor analysis doesn't show what the online LTA loss preserves.

Theorem~B.4 requires uniform bounds on both map differences and Jacobian differences over a neighborhood. LTA instead constrains sampled transition directions and excessive magnitude growth. This paper doesn't show that minimizing this loss controls either bound. For example, away from the numerical stabilizer, a positive contraction Δzθ=cΔzref\Delta z^\theta=c\Delta z^{\mathrm{ref}}, with 0<c<10<c<1, can incur no structural penalty despite changing the transition magnitude. Moreover, the fixed-point diagnostics use a different reconstruction chart from the online method. These results motivate transition alignment, but do not show attractor preservation by LTA. Measurements in the actual Qwen dual-anchor space would make this connection more convincing.

  • Baseline calibration doesn't give a fair retention-learning comparison.

LTA undergoes multi-dataset calibration, while Output KL uses a fixed weight of one, and the other structural baselines use gradient-based scale matching. Freezing these choices before formal evaluation is useful, but it doesn't establish comparable tuning budgets or optimal operating points. The learning trade-off also matters. On ReClor, LTA reaches 68.16% accuracy, compared with 70.92% for SFT. So, the abstract's statement that LTA matches SFT is too broad. Retention-accuracy curves under a shared development protocol would provide a stronger comparison than one selected setting per method.

Other minor issues, e.g.:

  • Line 052: "an transition–field" => "a transition–field"

  • Line 053: Missing a space in "reconstruction.Online"

  • Figure 9 caption: The right panel shows a fixed point and perturbation returns. Yet, the displayed right panel instead shows a cosine-similarity distribution.

Questions for authors
  1. Can you provide the full task-by-stage accuracy matrices, including later-task acquisition, and explain how the retention conclusions change after accounting for differences in first-task learning?

  2. What evidence connects the online loss to attractor preservation in the actual Qwen dual-anchor space? Does the second anchor improve results over start-anchor-only reconstruction?

4 comments

Response to Reviewer qM5f

We sincerely thank you for your careful reading and constructive feedback. We have carefully considered your concerns about the novelty of the transition objective, the attractor interpretation, and baseline calibration. We have added direct comparisons, component ablations, and mechanism measurements to address these points.

Reply 1: Complete sequential matrices and retention after later-task learning

Question: Can you provide the full task-by-stage accuracy matrices, including later-task acquisition, and explain how the retention conclusions change after accounting for differences in first-task learning?

Complete task-by-stage accuracy matrices.

We have now summarized and provided the complete task-by-stage matrices for all three task orders, five paired seeds, and five formal methods. The supplementary matrices contain 900 unique task-by-stage cells and 75 complete trajectories, with all three tasks evaluated after every 300-update stage. The complete matrices are available in supp/rebuttal_supplement/results/1P5B_CL.md.

First-task retention and later-task acquisition.

The paired comparison with Output-KL shows that Full LTA reduces first-task forgetting in all three orders:

Task orderLTA F1Output-KL F1ΔF1 [95% paired CI]ΔFinal mean [95% paired CI]
QASC→Chemistry→ReClor25.0733.27−8.20 [−13.96, −2.44]+1.46 [0.23, 2.68]
ReClor→Chemistry→QASC3.646.64−3.00 [−4.88, −1.12]−1.74 [−2.57, −0.92]
ReClor→QASC→Chemistry1.967.32−5.36 [−8.82, −1.90]+1.04 [−2.10, 4.18]

The paired confidence intervals for first-task forgetting exclude zero in all three orders. The complete matrices also report second-task acquisition and retention, third-task acquisition, and final three-task mean accuracy, allowing the learning–retention trade-off to be evaluated across the entire sequence.

The QASC-first order gives the strongest terminal gain: final QASC accuracy improves from 43.83% to 51.13% (+7.30 pp, 95% CI [+2.46, +12.14]), and final three-task mean accuracy improves from 62.77% to 64.23% (+1.46 pp, 95% CI [+0.23, +2.68]). In ReClor→QASC→Chemistry, final ReClor accuracy improves by 2.64 pp and final QASC accuracy improves by 2.53 pp. The final three-task mean increases by 1.04 pp.

We report acquisition and retention together. In the QASC-first order, the initial QASC difference is −0.90 pp and the final QASC difference is +7.30 pp. In ReClor→QASC→Chemistry, the initial and final ReClor differences are −2.72 pp and +2.64 pp. The complete matrices also include second-task forgetting F2F_2 and all final task accuracies.

Reply 2: Actual Qwen dual-anchor space, attractor-informed dynamics, and the second-anchor contribution

Question: What evidence connects the online loss to attractor preservation in the actual Qwen dual-anchor space?

Sample-conditioned dual-anchor maps and the attractor-informed motivation.

The dual-anchor construction provides a sample-conditioned dynamical interpretation. For each input and token context, the shared encoder maps both endpoints of a Transformer window into a common latent coordinate system, and the decoder reconstructs both endpoint codes as the same window-input state.

Under exact reconstruction, this induces a context-conditioned latent self-map. The end code is a fixed point of this sample-conditioned map, while the observed start-to-end transition equals the residual-field vector at the start code.

From sample-conditioned fixed points to the population-level transition field.

The surrogate is fitted jointly across many training examples and contexts. It therefore aggregates these sample-conditioned transition–field relationships into a population-level transition field that captures the original model's collective latent processing behavior. Each example and context can induce its own fixed point. The learned population field represents the aggregated dynamics across these sample-conditioned relations and does not require a single global fixed point shared by all examples.

Online LTA preserves this attractor-informed field through directional alignment and one-sided magnitude control. Direction alignment preserves the orientation of observed transitions. The magnitude term penalizes excessive radial expansion while leaving contractions unpenalized. Its speed-dependent weighting assigns greater confidence to low-speed reference transitions, protecting the low-speed region of the transition geometry while retaining flexibility for adaptation.

Evidence in the actual Qwen2.5-3B dual-anchor charts.

We measure transition–field correspondence in the actual Qwen2.5-3B dual-anchor charts. Across three domains and two frozen windows, correctly paired cosine similarities range from 0.2158 to 0.4494, compared with 0.0018–0.0673 for shuffled-context, shuffled-transition, and cross-domain controls. Correct pairing is consistently stronger across all six domain-window combinations.

The adapted QASC population-geometry analysis provides complementary evidence. Full LTA reduces covariance-shape distortion from 0.451650 for Direction-only to 0.347517, a 23.1% reduction. Effective-rank displacement decreases from 3.17927 to 1.08884, a 65.8% reduction. These measurements connect the online objective to preservation of the population-level transition-field geometry.

We retain Theorem B.4 as a conditional persistence result under its stated neighborhood-wise map and Jacobian conditions. The actual-chart measurements establish the empirical transition–field correspondence and population-geometry preservation underlying the attractor-informed formulation.

Question: Does the second anchor improve results over start-anchor-only reconstruction?

Second-anchor ablation: DA, SA, and EA.

We directly isolate the second-anchor contribution through matched representation ablations. Dual-Anchor Reconstruction (DA) reconstructs the window input jointly from both endpoint encodings. Start-Anchor Reconstruction (SA) reconstructs the window input using only the start-endpoint encoding. Endpoint Autoencoding (EA) reconstructs each endpoint from its own encoding. All three learned representations use the same 128-dimensional surrogate architecture, 3,000-step fitting budget, and online directional objective.

RepresentationReClor accuracyReClor ΔNLLSequential final meanFirst-task forgettingCumulative ΔNLL
DA68.13±0.70%0.01496±0.0011264.96±0.37%24.44±1.08 pp0.15822±0.00689
SA69.53±0.83%0.01631±0.0009263.64±0.37%27.44±1.84 pp0.17430±0.00603
EA67.67±0.31%0.02306±0.0011264.61±0.24%25.33±1.86 pp0.20509±0.00744

After selecting r=10r=10 on ReClor and freezing it for QASC→Chemistry→ReClor, DA achieves the highest final mean accuracy, the least first-task forgetting, and the lowest cumulative drift. Relative to SA, DA improves final mean accuracy by 1.32 pp and reduces cumulative drift by 9.2%. This directly demonstrates the incremental value of the second reconstruction anchor over simpler representations under the same online directional objective.

Reply 3: FDD, magnitude ablation, and matched-accuracy calibration

Question: How much do these additions improve upon prior directional alignment?

FDD as a direct cross-layer directional-alignment baseline.

We agree that FDD is a strong and closely related method, and we have now included it as a baseline in our comparisons.

We added FDD as a direct baseline for cross-layer directional alignment. FDD uses four uniformly spaced Transformer states, projects them through the model's own LM head into full-vocabulary log-probabilities, and aligns adjacent-state derivatives using cosine distance. FDD-full additionally includes full-vocabulary output KL and intermediate trajectory KL. This comparison directly tests LTA against a closely related prior cross-layer distillation objective.

Five-seed 1.5B FDD comparison.

Across QASC/Chemistry/ReClor, LTA Direction-only achieves lower five-seed mean General-reference ΔNLL\Delta\mathrm{NLL} than FDD-direction in all three domains: 0.03424/0.02498/0.00711 vs. 0.10402/0.06093/0.05058, reductions of 67.1%/59.0%/85.9%, respectively. Full LTA also achieves lower drift than FDD-full on QASC (0.03194 vs. 0.04135; 22.7% lower) and ReClor (0.00701 vs. 0.01331; 47.3% lower). On Chemistry, FDD-full has lower drift than Full LTA (0.01282 vs. 0.02693). The complete five-seed per-domain results and curves are provided in the supplementary material.

Five-seed Qwen2.5-3B Direction-only ablation: the magnitude term's independent benefit.

We added a complete five-seed Direction-only ablation on Qwen2.5-3B QASC, using the same 300-update evaluation protocol. The results directly isolate the one-sided magnitude term:

MethodQASC accuracyAggregate ΔNLL\Delta\mathrm{NLL} ↓
Full LTA77.17±3.21%0.01080±0.00235
Direction-only78.67±1.07%0.01796±0.00620
FDD-full79.67±0.47%0.02666±0.00423

Full LTA reduces aggregate drift by 39.8% relative to Direction-only and 59.5% relative to FDD-full. These results quantify the retention contribution of the complete LTA objective relative to both a matched directional ablation and a strong prior baseline.

We also completed the full five-seed 1.5B sequential Direction-only matrices. Across orders A, B, and D, Full LTA obtains a macro-average first-task forgetting of 10.22 pp, compared with 11.84 pp for Direction-only, and a final mean accuracy of 68.54%, compared with 68.12%. In the QASC-first order, Full LTA improves final QASC accuracy by 4.70 pp and reduces first-task forgetting by 3.67 pp. The complete matrix is provided in supp/rebuttal_supplement/results/1P5B_DIR.md.

Independent magnitude stress test under severe update pressure.

To further isolate the magnitude term, we conducted a separate ReClor learning-rate-pulse stress test. The multiplier q∈1,2,4q\in\\{1,2,4\\} is applied only during updates 101–120, after which the base learning rate 4×10−54\times10^{-5} is restored.

Under the strongest q=4q=4 intervention:

MethodStep-120 accuracyRecovery (120→300)Step-300 accuracyFinal aggregate ΔNLL\Delta\mathrm{NLL}
Direction-only27.73±2.10%39.67±0.90 pp67.40±1.20%0.07448±0.00472
Full LTA26.60±1.91%42.60±1.11 pp69.20±0.80%0.06103±0.00349

The magnitude term improves post-shock recovery by approximately 2.93 points, final accuracy by 1.80 points, and final aggregate reference retention by reducing drift by 18.1%. From q=1q=1 to q=4q=4, the additional drift increase is also 13.9% smaller with Full LTA. These results establish the practical role of one-sided magnitude control in recovery and final retention under strong update overshoot.

Question: Do the main retention gains hold at matched target accuracy, or under comparable coefficient-search budgets?

Matched-accuracy coefficient calibration on ReClor.

At matched target accuracy, do the main retention gains remain? Do they persist when LTA, Output-KL, and Transition-MSE receive comparable coefficient-search budgets?

We added a shared development-set coefficient sweep on 1.5B ReClor single-domain adaptation, using development seed 2026, a common relative grid r∈0.1,0.3,1,3,10r\in\\{0.1,0.3,1,3,10\\}, and a pre-specified SFT−1 pp accuracy threshold. Among candidates meeting the threshold, we selected the operating point with the lowest aggregate reference ΔNLL\Delta\mathrm{NLL}.

Under this threshold, Full LTA has a feasible point at 68.9655% accuracy / 0.0145558 drift. The best Transition-MSE point is 68.3190% / 0.0264799, while Output-KL has no candidate meeting the threshold. These results directly characterize the learning–retention trade-off under a shared development-set search protocol.

Question: Other minor issues: wording, spacing, and the Figure 9 caption.

Manuscript revisions and reproducibility.

We have corrected the manuscript's wording from “an transition–field identity” to “a transition–field identity,” inserted the missing space in “reconstruction. Online,” and corrected the Figure 9 caption to identify the right panel as the transition–field cosine distribution.

We have also revised the manuscript: Fig. 6 now includes the FDD baseline, Fig. 7 includes the Direction-only ablation, and Fig. 8 presents both the FDD and Direction-only ablation curves. We have additionally added the FDD citation at line 373.

The supplementary package at supp/rebuttal_supplement/ includes the complete per-seed result matrices, fixed experimental protocols and configurations, reproduction code, detailed experiment documentation, and machine-readable tables, facilitating independent verification and reproduction.

Reviewer_n4AYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Latent Transition Alignment (LTA) for replay-free continual learning of language models. Rather than matching outputs or individual hidden states, LTA preserves cross-layer latent transitions between selected Transformer depths. An offline dual-anchor surrogate connects these transitions to the residual field of a latent self-map. During adaptation, LTA aligns transition directions and penalizes only excessive magnitude growth, allowing the model to learn new domains without rigidly fixing representations. Experiments on NanoChat and Qwen2.5-1.5B/3B show reduced reference drift and first-domain forgetting relative to SFT, Output KL, state matching, and full-transition matching, with generally competitive target-domain performance and no inference-time overhead.

Critical strengths
  1. LTA focuses on cross-layer transformations rather than parameters, outputs, or isolated hidden states. Separating direction from magnitude provides a flexible and intuitively motivated retention objective.

  2. The paper includes matched baselines, component ablations, five paired seeds, multiple task orders, and transfer from Qwen2.5-1.5B to 3B. LTA consistently improves retention while maintaining competitive adaptation performance.

Critical weaknesses
  1. The formal evaluation is restricted to one model family, models up to 3B parameters, three adaptation datasets, and relatively short training horizons. Retention is measured mainly through reference-set NLL, leaving broader capabilities and longer continual-learning sequences insufficiently evaluated.

  2. The role of the attractor interpretation is not fully established. Online LTA directly optimizes transition direction and magnitude rather than fixed points, and the direction-only ablation performs almost identically to full LTA in mean reference drift. The experiments therefore support transition regularization more strongly than they support the necessity of the attractor-based formulation.

Questions for authors

see above

3 comments

Response to Reviewer n4AY

We sincerely thank you for your careful reading and constructive feedback. We have carefully considered your concerns and expanded the experimental evidence on model scale, longer training horizons, and the mechanism behind the attractor-informed construction. We respond below.

Reply 1: Model scale, longer training, and broader capabilities

Question: The formal evaluation is restricted to one model family, models up to 3B parameters, three adaptation datasets, and relatively short training horizons. Retention is measured mainly through reference-set NLL, leaving broader capabilities and longer continual-learning sequences insufficiently evaluated.

Broader capability evaluation on Qwen2.5-3B.

We have extended the evaluation beyond reference-set NLL to include downstream knowledge accuracy. On the fixed Qwen2.5-3B capability subsets, Full LTA achieves the highest MMLU accuracy among the adapted methods, while HellaSwag differences remain small:

StateMMLU (512)HellaSwag raw (1000)HellaSwag normalized (1000)
Base67.3856.073.1
SFT54.8854.872.8
Output-KL55.8654.272.1
Direction-only58.0155.572.3
Full LTA60.3555.472.7

Longer 1,000-step adaptation.

We also conducted a three-seed, 1,000-update MedMCQA adaptation study on Qwen2.5-1.5B, with 11 evaluation checkpoints and the same training budget and learning-rate protocol across methods.

MethodFinal accuracy (%)Mean reference ΔNLL ↓
SFT47.4860.11892
Output-KL47.5420.05985
Full LTA48.0760.08567
Transition-MSE45.6530.07709

Full LTA achieves the highest final task accuracy, while Output-KL retains the lowest reference drift. This provides a longer-horizon comparison of the learning–retention trade-off.

Scale extension to Qwen2.5-7B.

We further evaluated Qwen2.5-7B on QASC using SFT, KL-M3, and Full LTA. All three methods use seed 0 and 300 optimizer updates. The complete 18-row result matrix is presented below, covering the shared 60-update evaluation grid.

MethodUpdateAcc.Cand. NLLΔC4\Delta_{\mathrm{C4}}ΔGen.\Delta_{\mathrm{Gen.}}ΔWiki\Delta_{\mathrm{Wiki}}Mean ΔNLL\Delta\mathrm{NLL}
SFT083.500.5893130000
6087.130.4413600.0003620.0016380.0000480.000683
12088.330.3810090.0006750.000901-0.0012030.000124
18088.230.3505970.0009300.001843-0.0020840.000230
24088.700.3407360.0011880.002474-0.0006940.000989
30086.670.4665790.0025570.0080590.0052710.005295
KL-M3083.500.5893130000
6086.100.5038990.007885-0.003889-0.0023710.000542
12087.900.4151710.009447-0.002932-0.0040520.000821
18087.500.3841130.011375-0.005453-0.008244-0.000774
24087.170.3720340.014099-0.002105-0.0069510.001681
30085.670.4312530.014951-0.001114-0.0039180.003306
Full LTA083.500.5893130000
6086.330.4914470.000102-0.000140-0.000456-0.000165
12086.670.4535000.0007300.000159-0.0002970.000197
18087.830.4889550.0009430.000248-0.001435-0.000081
24088.000.4169030.0015930.001007-0.0006980.000634
30086.670.4467520.0018180.0033020.0017950.002305

At step 300, Full LTA matches SFT's QASC accuracy at 86.67% and achieves the lowest mean reference drift, 0.002305. This is 30.3% lower than KL-M3 (0.003306) and 56.5% lower than SFT (0.005295). KL-M3 also reduces mean drift relative to SFT, but its reference-specific behavior is strongly uneven: C4 drift reaches +0.014951, whereas General and WikiText drift are −0.001114 and −0.003918. Relative to SFT, Full LTA lowers drift on each reference set: C4 0.001818 vs. 0.002557, General 0.003302 vs. 0.008059, and WikiText 0.001795 vs. 0.005271. It therefore achieves the lowest mean drift while matching SFT's final QASC accuracy.

The complete 7B matrix and machine-readable values are available in supp/rebuttal_supplement/results/7B_QASC.md and supp/rebuttal_supplement/tables/qwen7b.csv.

Reply 2: Attractor-informed construction and the actual Qwen transition–field mechanism

Question: The role of the attractor interpretation is not fully established. Online LTA directly optimizes transition direction and magnitude rather than fixed points.

Sample-conditioned dual-anchor maps and the attractor-informed motivation.

The attractor-informed motivation follows from the sample-conditioned dual-anchor construction. For each training example and token context, the shared encoder maps both endpoints of a Transformer window into a common latent space, while the decoder reconstructs both endpoint codes as the same window-input state.

Under exact reconstruction, this induces a context-conditioned latent self-map. The endpoint code is a fixed point of that sample-conditioned map, while the observed transition from the start code to the endpoint code equals the residual-field vector evaluated at the start code.

From sample-conditioned fixed points to the population-level transition field.

The surrogate is trained jointly over many training examples and contexts. It therefore aggregates these sample-conditioned transition–field relationships into a population-level transition field that captures the original model's typical latent processing behavior. Each context can induce its own fixed point. The resulting population-level field represents the aggregated dynamics across these sample-conditioned relations and does not require all examples to share a single global fixed point.

Online LTA preserves this attractor-informed transition field without repeatedly solving for fixed points or fitting the full field. Direction alignment preserves the orientation of observed transitions. The one-sided magnitude term penalizes excessive radial expansion while leaving contractions unpenalized. Its speed-dependent confidence weighting gives greater weight to low-speed reference transitions, protecting the low-speed region of the learned transition geometry while allowing useful adaptation.

Evidence in the actual Qwen2.5-3B dual-anchor charts.

We now measure transition–field correspondence in the actual Qwen2.5-3B dual-anchor charts. Across three domains and two frozen windows, correctly paired transition–field cosine similarities range from 0.2158 to 0.4494. Shuffled-context, shuffled-transition, and cross-domain controls range from 0.0018 to 0.0673. Correct pairing is consistently much stronger across all six domain-window combinations.

The adapted-model geometry analysis provides a complementary result. In the QASC-adapted seven-state comparison, Full LTA reduces covariance-shape distortion from 0.451650 for Direction-only to 0.347517, a 23.1% reduction. Its effective-rank displacement decreases from 3.17927 to 1.08884, a 65.8% reduction. These measurements establish that the complete objective preserves the population-level transition-field geometry more accurately.

We retain Theorem B.4 as a conditional persistence result under its stated neighborhood-wise map and Jacobian conditions. The new measurements establish the empirical transition–field correspondence and population-geometry preservation in the actual Qwen dual-anchor charts.

Reply 3: Incremental contributions of the complete LTA construction

Question: The experiments therefore support transition regularization more strongly than they support the necessity of the attractor-based formulation.

The attractor-field preservation motivation is central to LTA: it directly yields the dual-anchor surrogate, directional alignment, and one-sided magnitude control, and the following matched ablations demonstrate the value of these components beyond generic transition regularization.

Five-seed 3B Direction-only ablation: the contribution of magnitude control.

The completed five-seed Qwen2.5-3B QASC comparison directly isolates the magnitude term. Full LTA achieves 77.17±3.21% accuracy and 0.01080±0.00235 aggregate reference ΔNLL, compared with 78.67±1.07% and 0.01796±0.00620 for Direction-only.

Full LTA reduces aggregate drift by 39.8%, with lower drift across all five seeds. This demonstrates a substantial practical retention contribution from the magnitude term at 3B scale.

Five-seed sequential Direction-only comparison.

We also completed the full five-seed Direction-only matrices across the three Qwen2.5-1.5B task orders. The results provide a direct comparison with Full LTA under the original five-seed continual-learning protocol.

OrderFull LTA F1 ↓Direction-only F1 ↓Full LTA final meanDirection-only final mean
A3.643.5270.74%70.89%
B25.0728.7364.23%62.63%
D1.963.2870.65%70.83%
Macro average10.2211.8468.54%68.12%

The macro-average F1 difference is −1.62 pp (paired 95% CI [−3.71, 0.46]), and the final-mean difference is +0.42 pp (95% CI [−0.27, 1.12]). The QASC-first order shows the clearest gain: Full LTA improves final QASC accuracy by 4.70 pp and reduces first-task forgetting by 3.67 pp. The complete matrices are available in supp/rebuttal_supplement/results/1P5B_DIR.md.

Independent magnitude stress test.

We designed a separate learning-rate-pulse experiment on ReClor to test the magnitude term under severe update pressure. The learning rate is multiplied by q∈1,2,4q\in\\{1,2,4\\} only during updates 101–120 and then restored to 4×10−54\times10^{-5}.

Under the strongest q=4q=4 intervention:

MethodStep-120 accuracyRecovery (120→300)Final accuracyFinal aggregate ΔNLL
Direction-only27.73±2.10%39.67±0.90 pp67.40±1.20%0.07448±0.00472
Full LTA26.60±1.91%42.60±1.11 pp69.20±0.80%0.06103±0.00349

Full LTA achieves 2.93 pp stronger post-pulse recovery, 1.80 pp higher final accuracy, and 18.1% lower final reference drift. From q=1q=1 to q=4q=4, its additional drift increase is also 13.9% smaller than that of Direction-only. This directly demonstrates the magnitude term's contribution to post-shock recovery and final retention.

Matched representation ablation: the contribution of the second anchor.

We further isolate the second-anchor contribution using three matched representations. DA reconstructs the window input from both endpoint encodings. SA reconstructs it using only the start-endpoint encoding. EA reconstructs each endpoint from its own encoding. They share the same 128-dimensional surrogate architecture, 3,000-step fitting budget, and online directional objective.

RepresentationReClor accuracyReClor ΔNLLSequential final meanFirst-task forgettingCumulative ΔNLL
DA68.13±0.70%0.01496±0.0011264.96±0.37%24.44±1.08 pp0.15822±0.00689
SA69.53±0.83%0.01631±0.0009263.64±0.37%27.44±1.84 pp0.17430±0.00603
EA67.67±0.31%0.02306±0.0011264.61±0.24%25.33±1.86 pp0.20509±0.00744

DA improves the long-horizon trade-off: it achieves the highest final mean, least first-task forgetting, and lowest cumulative drift among the three learned representations. Its coefficient r=10r=10, selected on ReClor, is frozen and transferred to QASC→Chemistry→ReClor without retuning. Relative to SA, DA improves final mean accuracy by 1.32 pp and reduces cumulative drift by 9.2%. This demonstrates the incremental value of the second reconstruction anchor.

Manuscript revisions and reproducibility.

We have revised the manuscript: Fig. 6 now includes the FDD baseline, Fig. 7 includes the Direction-only ablation, and Fig. 8 presents both the FDD and Direction-only ablation curves. We have additionally added the FDD citation at line 373.

The supplementary package at supp/rebuttal_supplement/ includes complete per-seed result matrices, fixed experimental protocols and configurations, reproduction code, detailed experiment documentation, and machine-readable tables.

Reviewer_4GuGOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Latent Transition Alignment (LTA) for replay-free continual learning of large language models. Rather than preserving output distributions or the locations of individual hidden states, LTA preserves cross-layer latent transitions. An offline dual-anchor encoder–decoder is trained on selected Transformer windows so that the latent codes of both window endpoints reconstruct the window’s input state. During adaptation, frozen encoders extract transitions from the trainable model and from a frozen rolling reference on the same current-domain inputs. LTA aligns the transition directions and penalizes only excessive magnitude growth, leaving contractions unpenalized. Only the updated language model is kept at inference time (Sections 4.1–4.2). The paper establishes exact and approximate transition–field relationships and gives a conditional result on the persistence of locally attracting fixed points. The experiments include a controlled NanoChat adaptation study, three single-domain tasks and three sequential task orders on Qwen2.5-1.5B, and QASC adaptation on Qwen2.5-3B. The reported results show encouraging reference-NLL retention and first-task retention gains (Sections 3 and 5; Appendices B and I).

Critical strengths
  1. A well-motivated preservation target. Matching transitions allows shared shifts of the two latent endpoints, which pointwise state matching does not. Separating direction from magnitude is also a sensible way to preserve cross-layer relationships without forcing exact displacement equality. The distinction from StateMatch and Transition-MSE is explained clearly (Section 4.2; Appendix C).
  2. Careful experimental controls. The formal Qwen comparisons use five paired seeds with matched initialization, batch order, token masks, and evaluation examples, and the structural methods share encoders and windows. Keeping development-stage calibration separate from the formal evaluation, and transferring the LTA coefficients to the 3B model without outcome-based retuning, both strengthen the comparisons (Section 5.1; Appendices D–F).
  3. Promising retention improvements. In the QASC → Chemistry → ReClor order, LTA’s mean final QASC accuracy is 51.13%, compared with 43.83% for Output KL. At 3B, aggregate reference ΔNLL drops from 0.0293 for Transition-MSE to 0.0108 for LTA. These are meaningful reported gains that justify further investigation (Tables 6–7).
  4. A practical training-only intervention. The online objective, masking, gradient flow through the frozen encoders, and stage-boundary updates are all described explicitly. Nothing extra is needed at inference, although training still requires a reference forward pass and encoder computation. The appendix also reports implementation-specific runtime and memory costs (Appendices C–D and K).
Critical weaknesses
  1. The sequential evaluation does not establish the full retention–adaptation trade-off. This is my main concern. Figure 7 and Table 6 report first-domain accuracy only, not the full task-by-stage accuracy matrix. As a result, they do not show how well each method learns the second and third domains, or how well the second domain is later retained. Lower first-task forgetting could partly reflect weaker later-task learning, and the current results do not rule this out. The single-domain experiments are useful but do not address acquisition from sequentially changing initializations. Differences in first-task acquisition also affect the reported forgetting gaps. For ReClor → Chemistry → QASC, Output KL goes from 70.84% to 64.20%, while LTA goes from 68.12% to 64.48%. The 3.00-point reduction in forgetting is therefore made up of a 2.72-point lower starting accuracy and a 0.28-point higher final accuracy. This does not negate the stronger QASC-first result, but it limits the claim of a uniform improvement across orders. The reported intervals are also wide: final QASC accuracy in that order is 51.13 ± 8.22 for LTA and 43.83 ± 5.78 for Output KL, so the headline 7.30-point gap is not shown to be reliable without a paired test on the per-seed differences. Please provide the complete accuracy matrices, later-task acquisition, and average final accuracy over all three tasks, with paired uncertainty estimates (Section 5.3; Table 6).
  2. The incremental benefit of the full construction is not well isolated. In Table 5, direction-only and full LTA have mean reference drifts of 0.0221 and 0.0220. The full objective improves some entries, but the aggregate difference is tiny and not consistent across domains. Without paired differences and their uncertainty, this is limited evidence for the added magnitude term. Direction-only is also missing from the sequential and 3B comparisons, where some of the strongest results are reported. More importantly, all structural baselines and component ablations share the same dual-anchor encoders. These comparisons isolate loss choices within that representation, but not the value of the representation itself. A matched directional-alignment control on raw hidden states, a simple projection, or an ordinary endpoint-reconstruction encoder would show what dual anchoring actually contributes. I am not claiming that the magnitude term or the dual-anchor construction is ineffective, only that their incremental benefit has not yet been demonstrated convincingly (Section 5.1; Appendices F and H; Tables 5–7).
  3. Single operating points do not establish superiority at comparable adaptation levels. LTA receives development-stage coefficient calibration, StateMatch and Transition-MSE use gradient-matched scales, and Output KL uses temperature 1 and weight 1. Using separate development data is good practice, but these procedures do not show that each method sits at an equally well-tuned retention–adaptation trade-off. For example, LTA’s mean ReClor accuracy is 68.16% versus 70.92% for SFT, and its 3B QASC accuracy is 77.17% versus 78.90% for SFT and 78.77% for Output KL. These mean differences are not significance tests in themselves, but they make matched-accuracy comparisons important. A modest coefficient sweep for LTA and the strongest baselines, under comparable development budgets, would clarify whether the retention advantage holds at comparable target accuracy. The abstract’s “matching SFT” claim should be qualified accordingly (Appendices D, E.3, and H; Tables 5 and 7).
  4. The attractor analysis motivates the method but does not establish its preservation mechanism. The transition–field identities follow under the stated reconstruction assumptions. Theorem B.4, however, requires uniform closeness of the reference and updated maps and their Jacobians over a neighborhood, while the online objective only constrains sampled transition directions and excessive magnitude growth. The paper does not show that the latter controls the former. In addition, the reported local-attraction diagnostics use the observational reconstruction chart, which the paper itself says is distinct from the dual-anchor chart used online. I appreciate that this distinction and the limitations are stated. My concern is not that the theorem is wrong or that a global guarantee is needed, but that the mechanistic interpretation is stronger than the demonstrated link. Diagnostics on the actual Qwen dual-anchor charts, or more restrained causal language, would help (Sections 3.2 and 4.1; Appendix B, especially Theorem B.4 and Figure 9).
Questions for authors
  1. Can the authors provide complete task-by-stage accuracy matrices for all three orders, including second- and third-task acquisition, retention of the second task, and average final accuracy, with paired confidence intervals for the key differences?
  2. Does full LTA improve over direction-only in the sequential and 3B experiments? What is the benefit of the dual-anchor encoder over a simpler representation under the same directional objective and training budget?
  3. Do the main retention gains hold at matched target accuracy, or under comparable coefficient-search budgets for LTA, Output KL, and Transition-MSE?
  4. What evidence connects the online objective to the map and Jacobian conditions of Theorem B.4 for the actual Qwen dual-anchor construction? Is attractor persistence meant as a demonstrated mechanism or as theoretical motivation?
  5. What are the exact sources, sample counts, and overlap controls for the fixed general-reference set, including its separation from the surrogate-fitting and calibration data? Can the Appendix J diagnostic scores be reported alongside response-format and parsing statistics?
5 comments

Response to Reviewer 4GuG

We sincerely thank you for your careful reading and constructive feedback. We have carefully considered your suggestions and conducted additional experiments to address the identified gaps and further clarify the empirical and mechanistic evidence. We respond to each point below.

Reply 1: Complete task-by-stage matrices and sequential retention

Question: Can the authors provide complete task-by-stage accuracy matrices for all three orders?

Complete task-by-stage accuracy matrices.

We have now summarized and provided the complete task-by-stage accuracy matrices across all three task orders, five paired seeds, and five formal methods (SFT, Output-KL, StateMatch, Transition-MSE, and Full LTA). The complete matrices are available in supp/rebuttal_supplement/results/1P5B_CL.md. They contain 900 unique task-by-stage cells and 75 complete training trajectories. Each stage consists of 300 optimizer updates, followed by evaluation on all three tasks. The evaluation sets contain 600 QASC examples, 263 single-label Chemistry examples (with 37 additional free-text examples evaluated separately using teacher-forced NLL), and 500 ReClor examples.

Acquisition-adjusted retention and paired confidence intervals.

The paired five-seed comparison with Output-KL directly addresses the acquisition–retention concern:

Task orderA11 (KL/LTA)A13 (KL/LTA)ΔA13 (LTA−KL) [95% CI]ΔF1 [95% CI]ΔFinal mean [95% CI]
QASC→Chemistry→ReClor77.10/76.2043.83/51.13+7.30 [2.46, 12.14]−8.20 [−13.96, −2.44]+1.46 [0.23, 2.68]
ReClor→Chemistry→QASC70.84/68.1264.20/64.48+0.28 [−2.69, 3.25]−3.00 [−4.88, −1.12]−1.74 [−2.57, −0.92]
ReClor→QASC→Chemistry70.84/68.1263.52/66.16+2.64 [0.38, 4.90]−5.36 [−8.82, −1.90]+1.04 [−2.10, 4.18]

Here, F1=A11−A13F_1=A_{11}-A_{13}, and Final mean is the arithmetic mean of the three final task accuracies. Negative ΔF1\Delta F_1 indicates less first-task forgetting.

The paired 95% confidence intervals for ΔF1\Delta F_1 are entirely below zero in all three orders. LTA's initial first-task accuracy is not higher than Output-KL in any order, so the retention improvement is not explained by stronger initial acquisition. The complete matrices also reveal the order-dependent final learning trade-off: the QASC-first order provides the clearest simultaneous retention and final-average gains, while ReClor→Chemistry→QASC has a lower final three-task average. We report these outcomes together to characterize the full sequential learning behavior.

Reply 2: Incremental contributions of Direction-only, FDD, magnitude control, and dual anchoring

Question: Does full LTA improve over direction-only in the sequential and 3B experiments?

Five-seed Qwen2.5-3B Direction-only ablation: the independent contribution of the magnitude term.

We have added a five-seed Direction-only ablation on Qwen2.5-3B QASC, directly comparing the complete objective with the same latent-transition objective after removing its one-sided magnitude term. This is one of our most direct tests of the magnitude component under the larger-model setting.

Under 300 updates, Full LTA achieves 77.17±3.21% accuracy and 0.01080±0.00235 aggregate reference ΔNLL. Direction-only achieves 78.67±1.07% accuracy and 0.01796±0.00620 drift.

MethodQASC accuracyAggregate ΔNLL ↓
Full LTA77.17±3.21%0.01080±0.00235
Direction-only78.67±1.07%0.01796±0.00620

Full LTA reduces aggregate reference drift by 39.8% relative to Direction-only, with lower drift for all five seed identifiers. This provides direct five-seed evidence that the magnitude term makes a substantial practical contribution to reference retention in the 3B model, while maintaining competitive target-task accuracy. The complete per-seed curves are provided in supp/rebuttal_supplement/results/3B_QASC.md.

Five-seed sequential Direction-only comparison.

We have also completed the full five-seed Direction-only task-by-stage matrices for Qwen2.5-1.5B across all three task orders, using the same five formal seed identifiers as the original paper. The matrix contains 180 task-by-stage cells, 15 trajectories, and 45 stagewise reference-NLL records. It is provided in supp/rebuttal_supplement/results/1P5B_DIR.md.

OrderFull LTA F1 ↓Direction-only F1 ↓Full LTA final meanDirection-only final mean
A: ReClor→Chemistry→QASC3.643.5270.74%70.89%
B: QASC→Chemistry→ReClor25.0728.7364.23%62.63%
D: ReClor→QASC→Chemistry1.963.2870.65%70.83%
Three-order macro average10.2211.8468.54%68.12%

Across the three orders, the paired five-seed macro-average difference is −1.62 pp for F1 (95% CI [−3.71, 0.46]) and +0.42 pp for final mean accuracy (95% CI [−0.27, 1.12]). The clearest gain appears in the QASC-first order, where Full LTA improves final QASC accuracy by 4.70 points and reduces first-task forgetting by 3.67 points. Together with the 3B results, this provides a complete five-seed evaluation of the magnitude component in both single-domain scale transfer and sequential learning.

FDD: comparison with prior cross-layer directional alignment.

We also added Feature Dynamics Distillation (FDD) as a strong cross-layer baseline. FDD projects four uniformly spaced Transformer states through the model's own LM head into the full-vocabulary log-probability space and aligns adjacent-state changes using a cosine objective. FDD-full additionally includes output KL and intermediate trajectory KL.

We include this baseline because FDD is one of the closest prior approaches to cross-layer directional alignment. A direct comparison tests the incremental value of LTA's learned dual-anchor latent representation and one-sided magnitude control over an established directional-alignment objective.

MethodQASC accuracyAggregate ΔNLL ↓
Full LTA77.17±3.21%0.01080±0.00235
Direction-only78.67±1.07%0.01796±0.00620
FDD-full79.67±0.47%0.02666±0.00423

Full LTA reduces aggregate drift by 39.8% relative to Direction-only and 59.5% relative to FDD-full. FDD-full achieves higher target accuracy, while LTA provides substantially stronger reference retention. This comparison quantifies the retention contribution of the complete LTA construction against both a matched component ablation and a strong prior baseline.

Question: Does full LTA improve over direction-only in the sequential and 3B experiments?

Independent learning-rate-pulse stress test.

To isolate the magnitude term's contribution to training stability, we designed a separate learning-rate-pulse stress test on ReClor. A multiplier q∈1,2,4q\in\\{1,2,4\\} is applied only during updates 101–120, after which the original learning rate 4×10−54\times10^{-5} is restored. This is an independent intervention rather than a change to the learning rate in the formal experiments.

Under the strongest q=4q=4 intervention:

MethodStep-120 accuracyRecovery (120→300)Step-300 accuracyFinal aggregate ΔNLL
Direction-only27.73±2.10%39.67±0.90 pp67.40±1.20%0.07448±0.00472
Full LTA26.60±1.91%42.60±1.11 pp69.20±0.80%0.06103±0.00349

Full LTA improves post-pulse recovery by approximately 2.93 points, raises final accuracy by 1.80 points, and reduces final reference drift by 18.1%. Comparing q=1q=1 and q=4q=4, the additional drift increase is 0.05215 for Full LTA and 0.06055 for Direction-only, a 13.9% smaller increase with Full LTA. These results show that the magnitude term contributes directly to recovery and final retention under severe update pressure.

Question: What is the benefit of the dual-anchor encoder over a simpler representation under the same directional objective and training budget?

Matched representation ablation: the contribution of the second anchor.

To isolate the second reconstruction anchor, we conducted a matched representation ablation. Dual-Anchor Reconstruction (DA) reconstructs the window input from both endpoint encodings. Start-Anchor Reconstruction (SA) reconstructs it using only the start-endpoint encoding. Endpoint Autoencoding (EA) reconstructs each endpoint from its own encoding. All three learned representations use the same 128-dimensional surrogate architecture, 3,000-step fitting budget, and online directional objective.

RepresentationReClor accuracyReClor ΔNLLSequential final meanFirst-task forgettingCumulative ΔNLL
DA68.13±0.70%0.01496±0.0011264.96±0.37%24.44±1.08 pp0.15822±0.00689
SA69.53±0.83%0.01631±0.0009263.64±0.37%27.44±1.84 pp0.17430±0.00603
EA67.67±0.31%0.02306±0.0011264.61±0.24%25.33±1.86 pp0.20509±0.00744

On ReClor, DA reduces drift by 8.3% relative to SA with a 1.40 pp accuracy difference, and reduces drift by 35.1% relative to EA while improving accuracy by 0.47 pp.

The frozen-coefficient QASC→Chemistry→ReClor experiment provides the strongest long-horizon evidence. The coefficient r=10r=10, selected on ReClor, is transferred without retuning. DA achieves the highest final mean, least first-task forgetting, and lowest cumulative drift among the three learned representations. Relative to SA, DA improves final mean accuracy by 1.32 pp and reduces cumulative drift by 9.2%. These results demonstrate the practical contribution of the second anchor to long-horizon retention.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.