Computation and Language · arXiv:2610.31401 v1 · submitted 22 Oct 2026 · 13 references
HERD: Harness-Evolving Recursive Distillation for Continual Learning
The paper says
“Continual learning requires a model to keep acquiring new skills and knowledge without compromising its existing capabilities. Parametric methods suffer from catastrophic forgetting and loss of plasticity.”
more of the abstract
Non-parametric methods, such as prompts, skills, and harnesses, leverage in-context learning but are limited by context length and lossy plain-text compression. To effectively combine their complementary strengths, we propose HERD (Harness-Evolving Recursive Distillation), a hybrid continual learning framework that learns efficiently in the harness and internalizes experience into cumulative capabilities via distillation. On the harness side, HERD continually refines its skills to capture reusable patterns, while retaining high-residual examples that the skills fail to capture in a bounded memory. On the parameter side, HERD performs drift-constrained distillation with dual anchors, substantially reducing forgetting over long recursive runs. On the 25-stage task stream of TRACE+17, under fixed memory and compute budgets, HERD matches the strongest parametric baseline in new-task acquisition while exhibiting the least forgetting and the strongest retention of general capabilities.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 23 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A well-reported hybrid continual learning paper whose dual-anchor drift constraint is the main contribution and also the main vulnerability: the 32-token probe misses the response-length collapse its own IFEval case documents, so the abstract overreaches, but the empirical core holds and it deserves serious review. This line is wrong
2 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 13 references say.
The paper says, at Section 2.4, Eq. (4); Appendix D, Fig. 9e
The review did not quote this passage.
Pith says: Objection 1 of 2 · major
The justification that 'only the opening is scored because it decides the mode of a response' is not supported for long-form outputs. The base-anchor drift D0 is computed on the first 32 greedy tokens of fixed probes, yet the paper's own IFEval minimum-word-count case shows HERD opening like the base and then stopping at 232-258 words against the base's 499 …
more of objection 1
… words, with strict pass rate falling from 0.71-0.76 at stage 4 to 0.18-0.24 at stage 25 while the audited stage-25 D0 is 0.041, below the 0.05 budget. Since Eq. (4) is the enforcement mechanism behind the abstract's 'strongest retention of general capabilities' and Section 3.2's least-loss claim, this acknowledged limitation is load-bearing rather than a scope caveat. I recommend either adding a length- or stopping-sensitive probe (for example, EOS probability or a minimum-length completion score) and re-auditing the IFEval category, or revising the abstract and Section 2.4 to state that the dual-anchor guarantee covers opener-decided modes and explicitly excludes long-form length modes; the comparative claim against parametric baselines may survive, but it should be stated at the level of what the probe actually bounds.
The paper says, at Section 3.1, Table 2; Appendix C
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The paired 95% bootstrap intervals resample items within tasks and G-test categories, not training runs, and only three seeds are available; the intervals therefore quantify item-sampling uncertainty conditional on the three runs rather than seed-level variability. The 'significantly' and 'statistically indistinguishable' statements in Section 3.2 are attached to these item intervals, while the paper itself reports seed-level reference scales that are non-negligible (for example, 0.024 for AA).
more of objection 2
Please either qualify the statistical statements as item-conditional or supplement them with a seed-level summary, such as the three per-seed differences for the load-bearing comparisons and the sign agreement across seeds. This would make the statistical language precise without changing the direction of the main results.
4 minor objections
The paper says, at Section 2.4 and Figure 5a
Pith says: minor objection 1 of 4
The sentence 'the student ... sees only the bare prompt' is easy to read as contradicting Eq. (3) and Figure 5a, where the student conditions on y*_<k at each answer position; please clarify that the 'bare prompt' means no skill or replayed-answer scaffolding in the prefill, while teacher forcing on the verified-answer prefix is still used during training.
The paper says, at Section 3.2, first paragraph
Pith says: minor objection 2 of 4
The phrase 'loses the least general capability among them' is ambiguous because HCL-Stable has a smaller G-test decline (-0.014) than HERD (-0.063); the sentence should say 'among weight-updating methods,' as the later paragraph does.
show 2 more
The paper says, at Figure 1 caption
Pith says: minor objection 3 of 4
'Thee.g.tiles follow' appears to be a typo for 'The e.g. tiles follow.'
The paper says, at Appendix D and Limitations
Pith says: minor objection 4 of 4
The probe-coverage finding that nine or ten of the 24 earlier tasks have no probe at stage 25 is important enough to appear in the main-text Limitations section alongside the known-boundaries caveat, rather than only in the case-study appendix.
What would settle it
Compute a length-aware version of Eq. 4 on the existing stage-25 checkpoint rather than retraining: for each of the 40 IFEval G-probe records, keep the 32-token prefix and add one continuation decision, namely the KL between anchor and student over EOS versus next-token at position 32, or a probe feature equal to the anchor's total greedy length. Re-audit D0. If the modified D0 exceeds ε0 = 0.05 while the paper's original audit reports 0.041, probe truncation is the cause of the undetected length collapse and the §2.4/abstract claim must be scoped to opener-controlled modes. If the modified D0 stays within budget, the IFEval decline is not visible to any prefix KL, and the honest claim is that the base anchor covers only capabilities represented in its probe distribution, not general capability as measured by the G-test.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 23 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 2 major objections, 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The short version: HERD is a credible, unusually honest systems paper whose central claim needs a qualifier it doesn't always carry. The abstract's "strongest retention of general capabilities" is only true among weight-updating baselines — the frozen controls and HCL retain more on the G-test — and the 32-token drift probe behind Eq. 4 demonstrably misses the category where HERD loses the most.
What's genuinely new is the assembly: an evolving harness of skills plus high-residual memory, distilled into weights under a dual-anchor drift constraint, the previous checkpoint bounding per-stage drift and the base model bounding cumulative drift, with worst-group divergence and dual-ascent multipliers. The ablations are the strongest part. Removing the base anchor sends ΔCmin to -0.490 with D0 at 0.212; removing the checkpoint anchor doubles forgetting; uniform memory selection doubles it to 0.020. Each component protects what it is supposed to protect, and the drift audits are reported at every stage, not just the endpoint.
The reporting standards deserve credit: three seeds, paired bootstrap intervals on every headline comparison, per-task and per-category trajectories, honest cost tables, a public code repository, and an explicit statement that no test score selected a candidate or hyperparameter. Appendix D documents three real failure modes — tasks losing their checkpoint probes (one seed of superni:192 drops from 0.83 to 0.52), a self-contradictory skill that passes the commit guard, and the probe-length failure on IFEval minimum-word-count items. These are substantive limitations, not decoration.
The load-bearing soft spot is that last one, and the stress-test note lands. Eq. 4 asserts the opening decides the mode of a response; Figure 9e shows otherwise for long-form outputs. HERD opens like the base model, then stops at 232–258 words where the base writes 499, and its strict pass rate falls from about 0.74 at stage 4 to 0.18–0.24 at stage 25 while the audited D0 sits at 0.041, under the 0.05 budget. The drift constraint does not bound the category where HERD loses the most. The empirical comparison survives — HERD loses 13 points on IFEval where Seq-LoRA loses 40 — and the authors flag probe length in Appendix D, so nothing here is hidden. What remains unresolved is that the gap is in the central mechanism, not just the benchmark scope.
Minor soft spots: one base model and one stream (acknowledged in Limitations); the non-parametric baselines are configured into near-failure, ACE pinned at its context ceiling and HCL committing almost nothing, which the paper reports honestly but which softens the "frontier" framing; ablations are single-seed with a stated rationale.
Who this is for: anyone working on drift-constrained distillation, harness-to-weight internalization, or memory-bounded continual learning. It deserves a serious referee; the probe-coverage question is exactly what peer review should push on. A revision should re-scope the abstract to "among weight-updating methods," validate or replace the fixed 32-token probe for long-form tasks, and add a second base model or stream. With that, conditional accept.
The strongest objection
HERD's frontier claim [b92] rests on §2.4's assertion that 'only the opening is scored because it decides the mode of a response' (Eq. 4). The paper's own Figure 9e contradicts that premise for long-form outputs: on IFEval minimum-word-count items, the stage-25 opener looks like the base model's, but the base produces 499 words while HERD produces 232–258 words. HERD's strict pass rate falls from 0.71–0.76 at stage 4 to 0.18–0.24 at stage 25, against the base's 0.94. Because both anchors compute KL on the first 32 greedy tokens, this length-mode collapse is invisible to D0; the audited stage-25 D0 of 0.041 sits under the 0.05 budget even while instruction-following capability drops by roughly 50 points between stages 4 and 25. Appendix D lists 'probe length' as a limitation, so this is not a hidden flaw; the unresolved point is that it attacks the central mechanism rather than only the benchmark scope. Eq. 4 is supposed to bound per-stage and cumulative drift, and the one G-test category where HERD loses the most is exactly where the probe is blind. The empirical comparison may survive, because HERD still loses less than the parametric baselines, but the claim that drift-constrained distillation preserves general capability, and the abstract's 'strongest retention of general capabilities,' are not supported for this category.
The paper is unusually transparent: paired bootstrap intervals, per-task breakdowns, drift audits, per-stage costs, and detailed case studies are reported, and the code is released. I read the central claim as an engineering claim with a mechanism: the dual-anchor drift constraint is what lets HERD retain earlier tasks and general capabilities while matching parametric acquisition. The main tables are internally consistent, and the paired intervals support the reported comparisons; I see no fatal flaw. The most load-bearing spot is the faithfulness of Eq. 4's probe to the quantities the anchors are supposed to bound. The paper's own case study shows a large IFEval drop that is invisible to the 32-token opener and is not flagged by the audited D0. This is disclosed in Appendix D, so it is not a hidden defect, but it is more than a scope caveat: it undercuts the §2.4 justification for the central mechanism in exactly the category where HERD's largest general-capability loss occurs. A length-aware probe would settle whether the failure is a probe artifact or a genuine coverage limitation. Because the reader's CONDITIONAL verdict already captures this concern and the empirical comparison is not overturned, I recommend no change to the verdict.
The referee's report
If the reported comparisons hold, HERD is a significant systems contribution: it combines near-parametric acquisition with much lower forgetting and bare-weight deployment, and its fixed per-stage cost plus audited drift budgets address a real stability problem in recursive self-distillation. The paper is unusually transparent, with code and dataset links, three seeds, per-task and per-category reporting, drift audits, cost tables, and case studies, and the direction of the ablations is consistent with the proposed mechanisms. The main caveat is that the central probe measures only the first 32 greedy tokens, so the abstract's 'strongest retention of general capabilities' overstates what the dual-anchor mechanism is shown to guarantee; the relative claim against weight-updating baselines is more robust than the absolute mechanism claim.
Pith's reading of the paper (a machine's summary, not the authors')
Continual learning forces a choice between updating a model's weights, which forgets old abilities, and keeping experience in an external context store, which is bounded and lossy. HERD claims to escape that choice by running both layers at different speeds: each stage it evolves a small harness — a registry of reusable skills plus a bounded memory of examples the skills cannot yet predict — then distills that harness into the weights through self-distillation constrained by two drift anchors. On the 25-stage TRACE+17 stream, at matched memory and compute budgets, the paper reports HERD matching the strongest parametric baseline in new-task acquisition ($0.713$ vs $0.715$), reaching the highest final average over all seen tasks ($0.706$), forgetting least among intact weight-updating methods ($0.010$), and losing the least general capability ($\Delta C = -0.063$ vs $-0.20$ to $-0.24$), with no harness attached at inference. If this holds, it offers a practical route to models that keep learning over long streams without wrecking what they already know, at a per-stage cost that does not grow with the stage index.
What carries the argument. The carrying object is drift-constrained distillation with dual anchors, enforced on a fresh LoRA adapter (a low-rank adapter). The student is the previous checkpoint seeing only the bare prompt; the teacher is the same frozen checkpoint under a privileged context — the routed skill as system message and the verified answer $y^*$ replayed as the model's own previous turn — and the student minimizes full-vocabulary forward KL divergence at every position of the verified answer. Two constraints, each enforced by dual ascent on a multiplier tied to a divergence budget $\varepsilon = 0.05$, bound the update: checkpoint drift $D_C$ on probes drawn from the residual memory and grouped by task bounds per-stage drift, and base drift $D_0$ on reserved general-capability probes grouped by source bounds cumulative drift. On the harness side, the residual $r(x, y^*; \theta, K)$ — the mean negative log-likelihood of the verified answer under the skill but with the answer absent from the context — drives every decision: only the model's mistakes reach the skill writer, a candidate skill is committed only if it lowers the window's mean residual, and the bounded memory keeps whole records in decreasing residual. A probe is a prompt plus the first 32 greedy tokens of an anchor, because the opening decides the mode of a response. The two anchors complement each other: the checkpoint keeps every step small, and the base keeps accumulated drift from growing.
Core claim. The paper's central claim, in its own sentence [b92], is that HERD 'lies on the frontier of the three families: it acquires as much as the strongest parametric method, retains more than any method that updates weights, loses the least general capability among them, and deploys without context.' The mechanism is a two-layer memory: the harness is the fast, selective layer that learns the current task and decides what to keep, and the weights are the slow layer that internalizes the harness through distillation under explicit drift constraints. Pure non-parametric methods, the paper argues, cannot solve continual learning because everything they know must pass through a bounded context, and plain-text compression loses information at a high rate; its own baselines confirm the point, with ACE and HCL finishing within 1.4 points of the frozen base model. Pure parametric methods pay for acquisition with old abilities: the intact weight baselines lose 20 to 24 points on the general test while HERD loses 6.3. The ablations attribute the gap to the specific components: removing skills costs 4.5 points of acquisition, uniform memory selection doubles forgetting, dropping the checkpoint anchor triples it, and dropping the base anchor lets cumulative drift grow until mathematics falls from 0.885 to 0.395.
The premise it rests on. The load-bearing assumption, stated in Section 2.4 and Eq. 4, is that drift measured on the first 32 greedy tokens of fixed probes controls the mode of a response; the paper's own case study (Figure 9e) shows this fails for long-form outputs, and a second acknowledged scope condition is that every training record has a verified answer $y^*$ and the stream has known task boundaries, one task per stage, with no revisits.
If the paper is right (Pith's inferences, not the authors')
A model trained by HERD deploys with no context attached: re-scoring the stage-25 weights with the stage-25 harness prepended to every prompt changes AA from 0.706 to 0.708, forgetting from 0.010 to 0.007, and no task by more than 0.4 points, indicating the harness has been internalized.
HERD's training time tracks answer length rather than the stage index (316 s at stage 1, 284 s at stage 25; 10,723 s total, within 3% of the plain Seq-LoRA budget), so the recursive loop does not compound compute as the stream grows.
Under an identical token budget, uniform memory selection doubles forgetting to 0.020, the same value as the uniform replay baseline, showing that retention depends on what the memory keeps and what constrains the update, not on how much is stored.
On the first eight tasks, HERD's mean stays within half a point of its stage-8 level through stage 25 (0.664 to 0.665), while every parametric baseline moves more, including Seq-LoRA's dip to 0.372 at stage 12.
The 27-point gap between HERD and the two non-parametric baselines supports the paper's argument that external context alone, however evolved, is a bounded and lossy channel that cannot accumulate experience in the model itself.
My inference: the dual-anchor scheme reads as a general stabilization recipe for recursive training — per-step anchoring to the previous checkpoint plus cumulative anchoring to an immutable base, with budgets in divergence units rather than tuned penalties — and could be lifted out of HERD and tested on other self-distillation loops.
My inference: the documented loss of probe coverage (nine or ten of 24 earlier tasks have no probe at stage 25, and the task that lost its last probe fell from 0.83 to 0.52 in one seed) suggests that a per-task floor of probes, the paper's own proposed repair, would extend retention to tasks the weights have fully absorbed.
My inference: the requirements of a verified answer per record and known one-task-per-stage boundaries restrict HERD's direct use to graded or curated streams; the paper's proposed substitutes (rubrics or majority rollouts as targets, stage boundaries opened by residual change) are the natural test of whether the guarantees survive in open-ended agent settings.
My inference: the 'opening decides the mode' premise could be tested in isolation by comparing forgetting measured on 32-token probes with forgetting measured on full-length completions of the same earlier tasks; the paper's claim predicts sign-change failures concentrate in the first few tokens, while the IFEval counterexample predicts long-horizon failures the probe cannot see.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Continual learning requires a model to keep acquiring new skills and knowledge without compromising its existing capabilities. Parametric methods suffer from catastrophic forgetting and loss of plasticity. Non-parametric methods, such as prompts, skills, and harnesses, leverage in-context learning but are limited by context length and lossy plain-text compression. To effectively combine their complementary strengths, we propose HERD (Harness-Evolving Recursive Distillation), a hybrid continual learning framework that learns efficiently in the harness and internalizes experience into cumulative capabilities via distillation. On the harness side, HERD continually refines its skills to capture reusable patterns, while retaining high-residual examples that the skills fail to capture in a bounded memory. On the parameter side, HERD performs drift-constrained distillation with dual anchors, substantially reducing forgetting over long recursive runs. On the 25-stage task stream of TRACE+17, under fixed memory and compute budgets, HERD matches the strongest parametric baseline in new-task acquisition while exhibiting the least forgetting and the strongest retention of general capabilities.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-23
The authors' main claim, as the classifier picked it out 1.00
“On the 25-stage task stream of TRACE+17, under fixed memory and compute budgets, HERD matches the strongest parametric baseline in new-task acquisition while exhibiting the least forgetting and the strongest retention of general capabilities.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.16 · human participants 0.03 · controlled comparison 0.92 · randomized assignment 0.01 · states quantitative result 0.77 · compares with prior results 0.96 · mentions code or data availability 0.02
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Continual learning requires a model to keep acquiring new skills and knowledge without compromising its existing capabilities. | background 1.00 | 0.04 | 0.03 | asserted 1.00 |
| Parametric methods suffer from catastrophic forgetting and loss of plasticity. | background 0.91 | 0.06 | 0.02 | asserted 1.00 |
| Non-parametric methods, such as prompts, skills, and harnesses, leverage in-context learning but are limited by context length and lossy plain-text compression. | background 0.90 | 0.06 | 0.02 | asserted 1.00 |
| To effectively combine their complementary strengths, we propose HERD (Harness-Evolving Recursive Distillation), a hybrid continual learning framework that learns efficiently in the harness and internalizes experience into cumulative capabilities via distillation. | contribution 0.96 | 0.82 | 0.19 | asserted 0.97 |
| On the harness side, HERD continually refines its skills to capture reusable patterns, while retaining high-residual examples that the skills fail to capture in a bounded memory. | method 0.99 | 0.81 | 0.04 | asserted 1.00 |
| On the parameter side, HERD performs drift-constrained distillation with dual anchors, substantially reducing forgetting over long recursive runs. | method 0.98 | 0.88 | 0.08 | asserted 1.00 |
| On the 25-stage task stream of TRACE+17, under fixed memory and compute budgets, HERD matches the strongest parametric baseline in new-task acquisition while exhibiting the least forgetting and the strongest retention of general capabilities. | result 1.00 | 0.95 | 0.81 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.86
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-23: 1 request, 5,614 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (9)
Showing 8 of the paper's 9 figures. Figures are the authors', reproduced from the paper.
References (13)
- [8] How much can language models memorize? InProceedings of the 43rd International Conference on Machine Learning, volume 306 ofProceedings of Machine Learning Research, pages 90392–90411. PMLR. Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy V... on Pith
- [9] Progressive prompts: Continual learning for language models. InThe Eleventh International Con- ference on Learning Representations. Annette Rios, Nicolas Spring, Tannon Kew, Marek Kostrzewa, Andreas Säuberli, Mathias Müller, and Sarah Ebling. 2021. A new dataset and efficient ... on PithPith's reviewarXiv
- [10] SLoRA: Balancing plasticity and forgetting in large language models for continual learning. In 12 Proceedings of the 64th Annual Meeting of the Associ- ation for Computational Linguistics (Volume 1: Long Papers), pages 5437–5454. Association for Compu- tational Linguistics. Zh... on PithPith's reviewarXiv
- [11] Self-consolidation for self-evolving agents. arXiv preprint arXiv:2602.01966. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a ma- chine really finish your sentence? InProceedings of the 57th Annual Meeting of the Association for C... on PitharXiv
- [13] The 200-item probe scored at the other stages shows what the anchors skip: Seq-LoRA’s probe mean falls to 0.28, 0.16, and 0.37 at stages 11 to 13 and is back at 0.59 by stage 14, a collapse and recovery that lies between the anchors of Figure 3b; HERD’s probe mean never falls ... on Pith
- [2017] Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13):3521–3526. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Effi- cient memory man... on PithPith's reviewarXiv
- [2019] Abstractive summarization of Reddit posts with multi-level memory networks. InProceedings of the 2019 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2519–2531, Mi... on Pith
- [2020] Measuring massive multitask language under- standing.arXiv preprint arXiv:2009.03300. 9 Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. LoRA: Low-rank adap- tation of large language models.arXiv preprint arXi... on PithPith's reviewarXiv
- [2021] Training verifiers to solve math word prob- lems.arXiv preprint arXiv:2110.14168. A. Feder Cooper, Mark A. Lemley, Allison Casasola, Ahmed Ahmed, Aaron Gokaslan, Amy B. Cyphert, Christopher De Sa, Daniel E. Ho, and Percy Liang on PithPith's reviewarXiv
- [2023] MeetingBank: A benchmark dataset for meet- ing summarization. InProceedings of the 61st An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16409– 16423, Toronto, Canada. Association for Computa- tional Linguistics. Jianheng Huang, ... on PithPith's reviewarXiv
- [2025] Spurious forgetting in continual learning of language models. InThe Thirteenth International Conference on Learning Representations. Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Sid- dhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. 2023. Instruction-following evalu- ... on PithPith's reviewarXiv
- [2026] Extracting memorized pieces of (copyrighted) books from open-weight language models. InThird Conference on Language Modeling. DeepSeek-AI. 2026. DeepSeek Harness: Everything is a plugin. Open-source software, developer preview. Grégoire Delétang, Anian Ruoss, Paul-Ambroise Duq... on Pith
- [2521] Curran Associates, Inc. Shuai Lu, Daya Guo, Shuo Ren, Junjie Huang, Alexey Svyatkovskiy, Ambrosio Blanco, Colin B. Clement, Dawn Drain, Daxin Jiang, Duyu Tang, Ge Li, Lidong Zhou, Linjun Shou, Long Zhou, Michele Tufano, Ming Gong, Ming Zhou, Nan Duan, Neel Sundare- san, and 3 ... on PithPith's reviewarXiv
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment