Pith. sign in

REVIEW 3 major objections 5 minor 13 references

Orientation, not magnitude: the causal structure of task-vector interference in merged language models

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read In merged language models, task-vector interference is carried and amplified along a stable direction that the forward pass actively rebuilds and the prompt format gates at expression; magnitude is at best a coarse, sometimes inverted…

desk verdict A careful, preregistered empirical paper that makes a real case for direction over magnitude in task-vector interference, with one shared-counterfactual caveat that keeps me at conditional rather than full accept. read the letter →

arxiv 2608.11797 v1 pith:GW3QCNXK submitted 2026-08-12 cs.LG

Chencheng Zhu This is my paper · ORCID
classification cs.LG
keywords modelmergingtaskarithmeticcross-terminterferencecausalinterventiondirectionspecificityexpressiongatinginstructiontemplatesquantizationroughness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that when two task vectors are added to a base language model, the functional interference between them is determined not by how much cross-term the network produces but by the orientation of that cross-term and what the forward pass does with it. Tracking an exact layerwise cross-term — the four-corner difference between the merged model's hidden states and the two single-task states plus the base — the paper finds that propagation transports and amplifies the existing term (roughly 65–70% of the layerwise flux, gain above 1 per late block), actively rebuilds it after erasure, and treats the carried direction as an attractor. Erasing along that direction removes expressed interference dose-dependently and saturates near full erasure, while norm-matched wrong-direction controls fail or backfire, and the same erasure finds 13× less removable interference under an instruction wrapper that internally amplifies the cross-term — so wrappers gate expression rather than removing the interference. If the paper is right, merging diagnostics should track the direction that survives propagation, format-wrapped evaluations read the gate rather than the interference, and the apparent ±15% 'universality' of naive bf16 estimates dissolves into quantization roughness.

What carries the argument

The load-bearing object is the layerwise cross-term $I_\ell = h_\ell^{AB} - h_\ell^{A0} - h_\ell^{0B} + h_\ell^{00}$ — the four-corner 'factorial ledger' difference between a merged model's residual state at the input of block $\ell$ and the two single-task states minus the base state — which is exactly the residual of the additive reconstruction $\bar{h}_\ell = h_\ell^{A0} + h_\ell^{0B} - h_\ell^{00}$. Three instruments carry the argument. First, an exact decomposition $I_{\ell+1} = G_\ell + T_\ell + M_\ell$ separates newly generated cross-term from transport of the existing term and measures the per-block gain. Second, cross-term erasure $h_\ell^{AB} - \lambda I_\ell$ at frozen depths, with four norm-matched structural controls (wrong-pair, coefficient-mismatch, cross-prompt, random), makes the direction the only varying quantity; the causal readout is the change $D = R_{\text{original}} - R_{\text{patched}}$ in the expressed-interference ratio, with the unpatched main-effect denominator frozen so the dose–response moves for exactly one reason. Third, regrowth and basin tests track the term downstream of erasure, establishing that propagation restores the carried direction in norm and orientation — the attractor property that explains why only late erasures produce durable causal effects.

What would settle it

Build the cross-term against an independently validated no-interference baseline — for example, a single model trained from the same base on both tasks jointly, or a merge of provably non-interacting adapters — and repeat the erasure dose–response; the claim collapses if removing the cross-term against that baseline no longer removes expressed interference dose-dependently, or if any norm-matched wrong-direction control removes as much as the correct direction. A second direct check: find any layer or condition where an early-depth erasure persists downstream (regrowth ratio well below 1) or where the cross-term fails to reconverge in direction from a structured starting displacement, since the attractor mechanism predicts reconvergence everywhere except where depth runs out.

Watch

Extended reading notes

Core claim

At every block of a merged model the paper defines the cross-term $I_\ell = h_\ell^{AB} - h_\ell^{A0} - h_\ell^{0B} + h_\ell^{00}$, the exact non-additive residue of merging, and asks what each block generates versus what it carries. The discovery is that the flux is transport-dominated: an exact decomposition $I_{\ell+1} = G_\ell + T_\ell + M_\ell$ attributes about 69% of the flux to transport of the existing term, with amplification above 1 per late block, and erasing the term early is undone by propagation — it regrows to about 99% of its norm at direction cosine 0.99, and six norm-matched starting displacements all reconverge, so the carried direction behaves as an attractor of the forward pass. Causally, erasing along that direction, $h_\ell^{AB} - \lambda I_\ell$, removes expressed interference dose-dependently and saturates near $\lambda \approx 1$ (restoring the additive reconstruction), while wrong-pair, coefficient-mismatch, cross-prompt, and random controls fail or backfire in proportion to their projection on the erasure direction. The same erasure removes roughly 19% of expressed interference on raw prompts but 13× less under an instruction wrapper, even though the wrapper amplifies internal cross-term generation 1.7–1.8×, because the template pins the output distribution and drowns the interaction in the main effect. As a result, task pairs whose local generation magnitudes differ by at most 1.9× differ by 14×–337× in causally removable interference, and the apparent ±15% uniformity of naive bf16 generation turns out to be 75–90% quantization roughness; the paper reports all forty-six predictions as frozen before their data, with falsifications of its own expectations reported as falsifications.

Load-bearing premise

The load-bearing premise is that the additive reconstruction $\bar{h}_\ell = h_\ell^{A0} + h_\ell^{0B} - h_\ell^{00}$ is what a no-interference merge would produce at every layer: the cross-term is defined as the deviation from this linear combination, the erasure restores states to it, and if that counterfactual is wrong the 'direction' being studied is a biased displacement and the causal conclusions could reflect the baseline chosen rather than the network's causal structure.

Editorial extensions

If this is right

  • Erasure-style mitigation must act near the output: early erasures (depths 3–8) are rebuilt by propagation and show $D \approx 0$, whereas late erasures (depths 20–25) carry the full effect, so removable interference is exactly what the remaining depth cannot reconstruct.
  • The magnitude family of diagnostics — local generation, cumulative cross-term norm, parameter cosine, and representation bias — is at best a coarse pair-level correlate and is inverted on the format axis; none of them sees the roughly 20× expression collapse under instruction wrappers, so format-wrapped evaluations of merged models read the gate, not the interference.
  • Pair difficulty in merging is decided by the orientation of the carried cross-term, not its size: pairs within 1.9× in local generation differ by 14×–337× in what erasure can remove, so causally removable interference is the quantity that separates task pairs.
  • Activation-difference analyses run at training precision can be dominated by quantization roughness: the ±15% bf16 'universality' band is 75–90% weight-rounding artifact, so such estimates need per-estimator precision audits before interpretation.
  • The directional mechanism transfers across model families and scales — it replicated at all three model points — while magnitude behavior does not, so 'universal' claims built on magnitude should be treated as family-specific.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the attractor reading is correct, the cross-term direction is maintained by the marginal single-task paths rather than stored in the merged weights; a testable extension is that the same attractor direction should reappear regardless of coefficient order or composition-grid position, making directional signatures of interference roughly invariant to which task pair is merged first.
  • The gating result implies that format choice is not a neutral evaluation variable: two evaluations of the same merge can disagree by roughly 20× while internal states remain nearly identical, so comparative claims about merged-model quality should be reported raw and wrapped, or explicitly format-scoped.
  • Since the causal effect tracks the projection $\cos(\delta, -I_\ell)$ of the displacement on the erasure direction, a cheap held-out predictor may exist in the alignment between the carried cross-term direction and the output readout direction; the paper's own direction features failed, but they were norm-contaminated, so a cleaned alignment feature is a natural next test.
  • The quantization-roughness result is a caution for the broader literature: layerwise linearity or interference measurements computed at bf16 may be partly flooring artifacts, and re-auditing such estimates in float32 is a cheap decisive check.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper studies interference in merged language models through a factorial ledger that tracks the layerwise cross-term I_l = h_AB - h_A0 - h_0B + h_00 relative to the additive reconstruction hbar_l = h_A0 + h_0B - h_00. It separates local generation G_l from transport, intervenes on the cross-term by replacing the residual state with h_AB - lambda I_l, and measures expressed interference with an output-side interaction ratio R against the logit-space additive prediction. On Qwen2.5-1.5B, with reduced replications at Llama-3.2-1B and Qwen2.5-7B, the paper reports that transport dominates generation (~65-70%), that erasure along the cross-term direction removes expressed interference dose-dependently while norm-matched controls fail or backfire, that the carried direction is rebuilt by propagation except near the output, and that instruction wrappers amplify internal cross-terms while suppressing their expression. The paper also documents that naive bf16 estimates of local generation are largely quantization roughness, and it reports 46 preregistered predictions with explicit falsifications.

Significance. If the central causal interpretation is accepted, this is a substantial contribution: it is the first activation-level causal intervention on task-vector cross-terms that I am aware of, with unusually careful numerical audits (bitwise determinism nulls, propagated-ULP floors, float32 causal runs), a frozen preregistration chain, controls with disclosed overlaps, and replications across three model points. The bf16 quantization-roughness caution is a valuable standalone methodological warning. The paper is also unusually honest in reporting falsified predictions, including two of three headline expectations. However, the intervention and the primary output estimand share the same additive counterfactual, and the only behavioral endpoint that is independent of that counterfactual (Appendix I) returns a negative result. The paper's broader claims about task-vector interference therefore need either an independent validation of the additive baseline or an explicit narrowing of the claims to the additive-reconstruction definition of interference.

major comments (3)
  1. [§3.2–§3.4, Eqs. (1)–(3)] The intervention target and the primary output estimand share the same additive counterfactual. Eq. (1) defines I_l as the deviation of h_AB from hbar_l = h_A0 + h_0B - h_00; the erasure h_AB - lambda I_l is a displacement toward that same hbar_l; and the output estimand R in Eq. (3) measures non-additivity against the logit-space additive prediction l_A0 + l_0B - l_00. D therefore measures, at least in part, whether moving the state toward hbar_l moves the output toward the additive logit prediction. If hbar_l is not the true no-interference counterfactual, D can be construction-driven. The depth-profile rebuttal in Section 5 does not fully rule this out: the regrowth statistic I_post is itself formed against the same hbar-based marginals, so convergence of I_post to I_orig is convergence to the original deviation from that baseline, and the coefficient-mismatch control lies in the same affine span, testing the specialness of the (+1,-1,+1) combination rather than the validity of the span's origin. The one endpoint that does not share the additive counterfactual—the teacher-forced NLL of Appendix I—gives a negative result, so the independent validation is absent. Please either (i) supply an independent test of the baseline, for example a family of alternative references (h_A0+h_0B, h_A0+h_0B-2h_00, or a third-task-derived reference) and show that the dose-response, saturation at lambda=1, and pair contrast are unique to the additive reference; or (ii) explicitly frame the conclusions as being about the additive-reconstruction definition of interference and adjust the abstract and title accordingly.
  2. [§4.2 and Appendix E] The 'arithmetic exclusion' is computed from ||G_l|| evaluated at the common reference hbar_l, but in the actual forward pass the state at layer l is not hbar_l, so the effective per-layer injection at that layer is the regeneration operator G_l + M_l of Eq. (4), which the paper itself separates from G_l. The observed ratio ||I_28|| / sum_l ||G_l|| = 2.6 therefore excludes only injection accounts that repeatedly evaluate generation at the reference state; it does not exclude state-dependent generation accounts whose increments are G_l + M_l. Since the transport share is an exact identity only relative to the chosen hbar, this limitation also propagates to the 'transport carries ~69%' claim. Please either recompute the bound with the actual regeneration magnitudes and report that result, or weaken the 'no per-layer injection account survives' wording to what the bound can actually support.
  3. [§5, final paragraph; also abstract and §1] The headline contrast '14x-337x' is not stable. The 337x value comes from a seed whose D(code+math) is 0.0001, i.e. indistinguishable from zero, and the text acknowledges the near-zero denominator. The robust range from the other seeds is about 14-30x (or 18-52x in unnormalized JSD, again with one seed indistinguishable from zero). Because this contrast is the paper's central quantitative claim, please report the stable range as the headline, state the unstable seed separately, and avoid repeating the 337x figure in the abstract and Section 1 without this qualification.
minor comments (5)
  1. [Abstract and §5] The text says erasure removes 'roughly 19%' of expressed interference, but with R = 10.33% for code+safety and D = 0.0301 the implied fraction is about 29%; please reconcile the numbers or define the reference R used for the 19% figure.
  2. [Figure 1 caption] The caption refers to 'the strongest composition point' without a definition; the text should state that this is the frozen point (1,1) or define the location explicitly.
  3. [§3.1] The phrase 'norm-matched to the global core-median norm' is not defined; please specify the normalization procedure and whether the median is computed over task vectors, adapters, or something else.
  4. [§B, P1-P6 memo] The paper states that all 46 predictions were preregistered before their data, but the P1-P6 memo was written after two pilot cells; since no confirmatory prediction was added after its own outcome data, this is acceptable, but please state the pilot-cell timing in the main text rather than only in the appendix.
  5. [§4.1] The statement that paired contrasts 'survive' the bf16 rounding floor should be quantified: the claim that the floor is condition-independent and cancels in differences is asserted, but no explicit paired-contrast audit is reported for the main sweep.

Circularity Check

1 steps flagged · score 4.0 of 10

Erasure target and output estimand share the same additive counterfactual, making the headline 'removable interference' partly construction-driven; a companion self-citation also anchors the format-gate facts.

  1. self definitional [Eqs. (1)-(3), Sections 3.2-3.4 and 5 ('The depth profile rules out an estimand artifact')]
    "Iℓ = h AB ℓ − h A0 ℓ − h 0B ℓ + h 00 ℓ (1) ... at λ=1, the carried cross-term is exactly erased and the state is restored to the additive reconstruction; ... expressed interference on a prompt batch is the interaction ratio R: the mean Jensen–Shannon divergence between the merged model’s next-token distribution and its no-interaction additive prediction (logit-space ℓA0 + ℓ0B − ℓ00). Intervention target and output estimand share the additive counterfactual, so the reduction might be partly by construction."

    Both the erasure target and the measured outcome are displacements from the same additive counterfactual hbar_l = h_A0 + h_0B − h_00. Patching h_AB − I_l at λ=1 instantiates hbar_l in activation space, and R is defined as divergence from the same affine combination in logit space; D therefore measures the change in distance to that chosen origin. The depth profile and the coefficient-mismatch control reduce but do not eliminate the construction-driven component: regrowth is also measured against hbar, the control has near-zero overlap with −I_l and so does not test other full-overlap directions, and the one endpoint independent of the additive counterfactual (Appendix I) returned a null/failed result.

full rationale

The central derivation is not a full definitional collapse: the depth profile (D≈0 at depths 3–8 with regrowth ρ≈0.99, rising to +0.026 to +0.058 at depths 20–25), the projection-scaling of the controls, the second-family and scale replications, and the transport decomposition are genuine empirical content that would not be entailed by the shared additive baseline alone. The coefficient-mismatch control, however, lies in the same affine span with near-zero overlap with −I_l, so it only rules out generic span displacements rather than other full-overlap directions; and the paper's independent continuous-NLL endpoint (Appendix I) falsified behavioral recovery, as the paper honestly reports. Because the erasure target and the output estimand are both defined as displacements from hbar_l = h_A0 + h_0B − h_00, the headline 'causally removable interference' is partly construction-driven. A secondary load-bearing self-citation is the companion study [Zhu, 2026], which supplies the protocol and the 20× format-collapse ground truth for the gate framing. On balance this is partial circularity rather than a reduction to the input: score 4.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or architectural entities; its contributions are measurements, interventions, and interpretations of existing quantities. The main uncharged premises are the additive-counterfactual validity, the first-token proxy, the R estimand, the three-seed statistical basis, and the representativeness of the two causal task pairs.

assumptions (5)
  • domain assumption The additive reconstruction hbar_l = h_A0 + h_0B - h_00 is a valid no-interference counterfactual state at every layer.
    Used to define I_l in Eq. 1 and the erasure intervention in Section 3.4; if the true no-interference state is not this linear combination, the intervention may remove a partially arbitrary quantity.
  • domain assumption Last-token hidden states and first-token output distributions are sufficient to characterize merging interference.
    All hidden measurements are read at the final prompt token and all causal effects are on first-token distributions (Sections 3.1 and A); full-sequence behavior is not measured.
  • domain assumption The interaction ratio R, a ratio-of-means Jensen-Shannon divergence normalized by the main effect, is a valid operationalization of expressed interference.
    R is the central output-side estimand and comes from the same-author companion paper [Zhu, 2026]; the ratio normalization is central to the 20x format-gating claim.
  • domain assumption Three training seeds are sufficient as the outer replication unit, with 3/3 directional agreement as the confirmatory criterion.
    The paper's confirmatory statistics rest on three seeds; bootstrap CIs are within-seed (Section B and limitations).
  • domain assumption The two selected task pairs (code+safety, code+math) are representative for the causal erasure claims; the held-out 15-pair analysis is predictive rather than causal.
    Headline causal erasure is measured on these two pairs only; selection is described as geometrically extreme in Section 4, and the held-out pairs are not subjected to erasure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Orientation, not magnitude: the causal structure of task-vector interference in merged language models." pith.science (2026). https://pith.science/paper/GW3QCNXK

@misc{pith2026260811797,
  author       = {Pith},
  title        = {Pith review of: Orientation, not magnitude: the causal structure of task-vector interference in merged language models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GW3QCNXK}},
  note         = {Machine review of arXiv:2608.11797}
}
read the original abstract

Model merging by task arithmetic works until it doesn't, and the field diagnoses why with magnitudes: layerwise representation bias, deviations from cross-task linearity, parameter overlap. Tracking the exact layerwise cross-term of merged LLMs through a factorial ledger and intervening on it directly, we find magnitude insufficient - and inconsistent across model families - as a diagnostic axis. An exact decomposition of the layerwise flux shows it is dominated by amplifying transport of the existing cross-term (~65-70% in both families, gain >1 per late block), and erasing the term is undone by propagation - rebuilt to 99% of its norm at cosine 0.99 - unless applied near the output; a basin test with six starting displacements establishes the carried direction as an attractor of the forward pass. That direction is causally load-bearing: erasure along it removes expressed interference dose-dependently and saturates at exact erasure, while norm-matched wrong-direction controls fail or backfire. Instruction wrappers gate the effect: the same erasure finds 13x less relative interference to remove under a wrapper that internally amplifies the cross-term, because the wrapper drowns the interaction in a template-pinned main effect rather than shrinking it - a structure that replicates across further instruction templates but not under a length-matched control. Magnitude, by contrast, is at best a coarse correlate, and the striking +-15% "universality" of naive bfloat16 generation turns out to be quantization roughness. Task pairs whose local cross-term generation differs by at most 1.9x differ by 14x-337x in causally removable interference. All 46 predictions were preregistered and frozen before their data; falsifications, including of our own headline expectations and of behavioral recovery under a validated continuous endpoint, are reported as such.

Figures

Figures reproduced from arXiv: 2608.11797 by the authors.

Figure 1
Figure 1. Orientation, not magnitude: task pairs whose local cross-term generation differs by at most 1.9× differ by 14×– 337× in causally removable interference (b), and instruction wrappers amplify internal interference while suppressing its expression 13-fold (a, c). (a) Float32 local generation AUC(G) (from the preregistered fp32 grid, Appendix L) versus expressed interference R, all 36 cells at the strongest composition … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 3 canonical work pages

  1. [1]

    Layer-Aware Task Arithmetic: Disentangling Task-Specific and Instruction-Following Knowledge

    Yan-Lun Chen, Yi-Ru Wu, Chia-Yi Yang, et al. Layer-aware task arithmetic: Disentangling task-specific and instruction- following knowledge.arXiv preprint arXiv:2502.20186,

    work page Pith review arXiv

  2. [4]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen

    arXiv:2412.00081. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations (ICLR),

    arXiv

  3. [8]

    Qwen Team

    arXiv:2305.12827. Qwen Team. Qwen2.5 technical report.arXiv preprint arXiv:2412.15115,

    arXiv

  4. [10]

    Surgeryv2: Bridging the gap between model merging and multi-task learning with deep representation surgery.arXiv preprint arXiv:2410.14389,

    Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xingwei Wang, Xiaocun Cao, Jie Zhang, and Dacheng Tao. Surgeryv2: Bridging the gap between model merging and multi-task learning with deep representation surgery.arXiv preprint arXiv:2410.14389,

    arXiv

  5. [11]

    Zhanpeng Zhou, Zijun Chen, Yilan Chen, Bo Zhang, and Junchi Yan

    arXiv:2603.06242. Zhanpeng Zhou, Zijun Chen, Yilan Chen, Bo Zhang, and Junchi Yan. On the emergence of cross-task linearity in the pretraining-finetuning paradigm. InInternational Conference on Machine Learning (ICML),

    work page arXiv

  6. [12]

    Chencheng Zhu

    arXiv:2402.03660. Chencheng Zhu. When do task vectors interfere? mapping the validity boundaries of weight-space composition.arXiv preprint arXiv:2608.09490,

    arXiv

  7. [13]

    stable reversal

    A FULL ESTIMATOR AND AGGREGATION DEFINITIONS All definitions below are transcribed from the frozen design document (§B); none was altered after data existed. Four weight-state paths.For task vectors ∆A,∆B and coefficientsα,β , we instantiate four weight statesM00,Mα0, M0β, andMαβ and, for a promptx, record the residual state at the input of blockℓ under e...

    work page 2026

  8. [2012]

    The curse of multiple mediators: Hidden interaction effects in activation patching.arXiv preprint arXiv:2606.27510,

    Sankaran Vaidyanathan, David Arbour, Aaron Mueller, Scott Niekum, and David Jensen. The curse of multiple mediators: Hidden interaction effects in activation patching.arXiv preprint arXiv:2606.27510,

    arXiv

Show all 13 references
  1. [2021]

    Disentangling task interference within neurons: Model merging in alignment with neuronal mechanisms.arXiv preprint arXiv:2503.05320,

    Zitao Fang, Guodong Du, Shuyang Yu, Yifei Guo, Yiwei Zhang, Jing Cao, Jie Ma, Sim Kuan Goh, and Ho-Kin Tang. Disentangling task interference within neurons: Model merging in alignment with neuronal mechanisms.arXiv preprint arXiv:2503.05320,

    arXiv

  2. [2023]

    PKU-SafeRLHF: Towards multi-level safety alignment for llms with human preference

    Jiaming Ji, Donghai Chen, et al. PKU-SafeRLHF: Towards multi-level safety alignment for llms with human preference. arXiv preprint arXiv:2406.15513,

    arXiv

  3. [2024]

    Activated parameter locating via causal intervention for model merging.arXiv preprint arXiv:2408.09485,

    Fanshuang Kong, Richong Zhang, Ziqiao Wang, and Yongyi Mao. Activated parameter locating via causal intervention for model merging.arXiv preprint arXiv:2408.09485,

    arXiv

  4. [2025]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, et al

    arXiv:2503.08099. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    arXiv

  5. [2026]

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi

    arXiv:2602.22538. Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. InInternational Conference on Learning Representations (ICLR),

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.