Topics — The twenty-eight kinds of claim this anthology files against — the primary axis of both the practice registry and the reading list, and the same twenty-eight for each (ADR-026). The filing rule is the point: take a domain topic when the claim is ABOUT the domain; a claim merely discovered in one still takes its kind. The axis organizes the record, it does not bound it: content the list cannot place is evidence the list is short (ADR-059).
Adaptation and tuning — taking a trained model somewhere new — fine-tuning and transfer, preference training and safety alignment, parameter-efficient adaptation, context extension.
38 of 449 SOTA documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| SOTA-033 | continued pre-training for fine tuning | Gururangan et al. (2020), LIT-034 — ARXIV-2004.10964. | Active |
| SOTA-038 v3 | ICL permits few-shot task adaptability | Brown et al. (2020), LIT-035 — ARXIV-2005.14165. | Active |
| SOTA-051 v5 | Initialize a residual or adapter branch to exactly zero, not merely near zero | Bachlechner et al. (2020), LIT-047 — ARXIV-2003.04887. Start an added branch at exactly zero so it is the identity at initialisation, and let training raise it. Exact zero rather than small-random is the claim: a small random branch is PROBABLY harmless, a zero one is PROVABLY the identity — which is what matters when the thing being protected is a residual stack deep enough to attenuate, or a pretrained model expensive enough that adapter noise is not worth risking. | Active |
| SOTA-123 | For tiny specialized models, pretrain from scratch on the target SFT or reasoning data instead of pretrain-then-finetune | Falcon-LLM Team (2026), LIT-119 — the Falcon-H1-Tiny technical blogpost. At 90M, 25% SFT data in the pretraining mix beat a separate SFT stage by 10 IFEval points and yields one checkpoint that is both base and instruct. | Active |
| SOTA-126 | Run DPO on tiny models for one epoch only | Falcon-LLM Team (2026), LIT-119 — the Falcon-H1-Tiny technical blogpost. One epoch at LR 1e-6 to 3e-6 took a 90M model’s IFEval from about 50 to over 65; a second epoch degraded it while the reward kept rising. | Active |
| SOTA-129 v2 | Build a reasoning model in three stages: pretrain on general data, SFT on reasoning traces, then RL with verifiable rewards | Olmo Team (2025), LIT-130 — the recipe behind every open reasoning model at 7B and above, released in Olmo 3 with every stage and checkpoint in the open. Filed so its variations (SOTA-130) and its tiny-scale successor (SOTA-123) have something to be traced back to. | Active |
| SOTA-130 v4 | Skip the reasoning SFT stage and run RL with verifiable rewards directly on the base model | Olmo Team (2025), LIT-130 — the RL-Zero track: RLVR on the base model with no reasoning SFT in between, released as an open benchmark for studying RL rather than as the recommended recipe; a variation on SOTA-129 that the record keeps so the family of reasoning recipes can be traced. | Proposed |
| SOTA-145 v4 | Estimate the RL baseline from a group of samples for the same prompt instead of training a critic | Shao et al. (2024), LIT-127 — Group Relative Policy Optimization: PPO with the value model dropped and the baseline taken from the scores of several outputs sampled for the same prompt, which removes a model-sized chunk of the RL memory footprint and is what every later reasoning recipe in this record actually runs. | Active |
| SOTA-146 | Correct the GRPO objective before running it — the published form has three independently identified defects | Liu et al. (2025), LIT-167; Yu et al. (2025), LIT-168; Zheng et al. (2025), LIT-180 — three groups, three defects in GRPO’s objective, three published fixes, none of them the same fix. “We ran GRPO” does not say which of these you ran. | Proposed |
| SOTA-154 v4 | Fine-tune with evolution strategies instead of policy-gradient reinforcement learning | Qiu et al. (2025), LIT-211 — evolution strategies over the full parameter space of a billion-scale LLM, which the field had assumed impossible. +36.4% over base on average against GRPO’s +21.3% and PPO’s +17.9%, with ES on one fixed hyperparameter set while RL got a sweep per experiment. Active since Gan and Isola replicated the comparison on six further tasks with the RL arm tuned harder; contested, because Gu et al. argue direct ES at LLM scale does not work — on models at 1.5B and below. | Active |
| SOTA-184 v2 | Adapt a pretrained model by training a low-rank update to each weight matrix, not the matrix itself | Hu et al. (2021), LIT-046 — ARXIV-2106.09685. Freeze the pretrained weights and learn a rank-r product BA beside each one, so the trainable parameter count and the optimizer state fall by orders of magnitude and the adapter folds back into the weight at inference. | Active |
| SOTA-199 | Regularize a narrow fine-tune against the pre-fine-tuning model’s own samples | Active | |
| SOTA-201 | Give each pretrained component its own learning rate when fine-tuning a composite model | Active | |
| SOTA-210 v3 | Report pass@k as well as pass@1 after post-training: reinforcement learning raises one and lowers the other | Ba et al. and Hayes et al. (2026), LIT-230 and LIT-234. GRPO finishes below its own base model on pass@16 and pass@32 in 15 of 18 comparisons while improving pass@1; across Qwen2.5, Qwen3 and published RL checkpoints up to 32B the base model overtakes the RL checkpoint at large k. A post-training result reported at pass@1 alone cannot distinguish a model that learned something from one that stopped trying anything else. | Active |
| SOTA-212 | Post-train by scoring many random weight perturbations in one parallel pass and majority-voting the best of them | Qiu et al. introduced it (LIT-211); Gan and Isola (2026), LIT-233 — ARXIV-2603.12228 — measured this form of it. Sample N Gaussian weight perturbations, score them on a few hundred held-out examples, keep the top K and majority-vote at inference. At equal training FLOPs it matches or beats PPO, GRPO and ES across seven tasks at 0.5B–8B, in one parallel step instead of hundreds of sequential ones — 3.2 minutes for OLMo3-7B on Countdown across 200 GH200s. Filed Proposed: one group, K forward passes per answer, and discrete answers only. | Proposed |
| SOTA-213 | Control evolution-strategies drift with a larger population or an anchor penalty, not by stopping training early | Schweighofer et al. and Liang et al. (2026), LIT-238 and LIT-236, on the scaling Hoy et al. derived in LIT-235 — ES drift is a random walk whose size falls with population size, so raising the population from 30 to 128 halves the update norm and monotonically reduces prior-task degradation. Anchored Weight Decay buys the same reduction at population 30 for 1-2% runtime. Do not stop early instead: the prior-task dip is often transient and recovers by the end of training. | Proposed |
| SOTA-230 | Quantize the frozen base to 4-bit and keep the adapters in 16-bit | Dettmers et al. (2023), LIT-378 — ARXIV-2305.14314. Store the frozen base weights in 4-bit NormalFloat and dequantize to BFloat16 for every matrix multiply, training only 16-bit LoRA adapters. Fine-tuning a 65B model falls from >780GB to <48GB with no measured loss against a 16-bit fully fine-tuned baseline, because the arithmetic never happens in 4 bits. | Active |
| SOTA-231 | Put an adapter on every linear layer, not just the attention query and value projections | Dettmers et al. (2023), LIT-378 — ARXIV-2305.14314. Adapters on the query and value projections alone — the default inherited from the LoRA paper — do not reach full fine-tuning on large base models. The number of adapted matrices is the hyperparameter that decides it; the projection rank r, which is the one people tune, does not affect performance across the paper’s sweep. | Active |
| SOTA-233 | Train self-correction with multi-turn online RL on the model’s own traces, rather than prompting for it | Kumar et al. (2024), LIT-383 — ARXIV-2409.12917. Self-correction is a capability to be trained, not a behaviour to be requested. Prompting a model to revise its own reasoning does not reliably help and can hurt; supervised fine-tuning on correction traces fails by distribution mismatch or behaviour collapse; multi-turn online RL on self-generated traces reaches +15.6% on MATH and +9.1% on HumanEval. | Active |
| SOTA-235 v3 | Update the weights at inference on the test instance when the task is structurally novel | Akyürek et al. (2024), LIT-379 — ARXIV-2411.07279. Build a loss from the test instance’s own in-context examples, take gradient steps at inference, then discard the update. Up to 6x a fine-tuned baseline on ARC-AGI (53.0% at 8B) and +7.3 points on BIG-Bench Hard at 10-shot. The examples are worth several times more as gradient than as context. | Active |
| SOTA-279 v2 | Put worked reasoning steps in the few-shot exemplars when the task needs more than one step, and only once the model is large enough | Wei et al. (2022), LIT-467 — eight exemplars carrying worked steps more than doubled PaLM 540B on GSM8K, beating a fine-tuned model with a verifier. Three ablations rule out the equation, the extra tokens and knowledge activation. Below ~100B it does nothing or hurts. | Active |
| SOTA-283 | To fine-tune a model that is already quantized, bank the part of each update that is smaller than the lattice spacing, and rebuild the accumulator from seeds rather than storing it | Proposed | |
| SOTA-284 v2 | Choose the checkpoint you post-train from by how much mass it puts on good rare responses, not by validation cross-entropy | Proposed | |
| SOTA-299 | Finetune the whole generative model, encoder and decoder, for a dense perception task instead of attaching a head to a backbone | Khangaonkar and Pirsiavash (2025), LIT-488 — the freshly-initialized mask head is the part that cannot generalize, because it has only ever seen the finetuning categories. Keep every parameter generatively pretrained, encode the target as an image, and a model finetuned on furniture and cars segments people, x-rays and paintings. The same backbone under a conventional head scores 1.4–2.4 mIoU. | Proposed |
| SOTA-302 v2 | Steer a distilled generator by modulating its input noise, not by fine-tuning its weights | Eyring et al. (2025), LIT-491 — train a LoRA hypernetwork to predict an improved initial noise for a frozen step-distilled generator. GenEval on SANA-Sprint goes 0.70 → 0.75 for 0.1 s of added latency, recovering about half of what 30-second test-time optimization buys. Reward fine-tuning the same model instead takes it 0.73 → 0.62: the anchoring KL term is intractable in weight space and tractable in noise space. | Proposed |
| SOTA-318 | Choose the rank per matrix when you compress a transformer, because low-rank structure varies by component and by depth | Jaiswal et al. (2024), LIT-516 — “transformer weights are low rank” is true of some matrices and false of others. Query, Key and MLP Gate converge to low rank; MLP Up, MLP Down and Value do not; middle blocks resist while the first and last few give way. Setting the rank per matrix instead of globally is ~6.4× better in perplexity than uniform reduction at 30% on LLaMA-2 7B and ~47× at 40% on 13B. | Active |
| SOTA-323 | Test a claim about what pretraining produces on a model trained with the pretraining objective, not one trained on the task family you are testing | Shen, Mishra and Khashabi (2023), LIT-536 — “there exist weights such that” and “pretraining produces weights such that” are different claims, and training a model on the task family you are about to test collapses them. Their separation of these as Hypothesis 2 and Hypothesis 1 is what let them show that a literature everyone read as evidence about pretrained language models had tested neither the models nor the objective. | Proposed |
| SOTA-329 | For short-output tasks, adapt a frozen model with a low-rank intervention on a few positions of the residual stream, not a weight adapter; do not use it for long chain-of-thought | Wu, Arora et al. (2024), LIT-549 — LoReFT sets an r-dimensional subspace of the hidden state at a few prompt positions to a learned linear function of that state, with the model frozen. At 0.03% of parameters against LoRA’s 0.8%, it leads on commonsense QA (LLaMA-7B 80.2 against 74.7) and is level on GLUE. It loses on arithmetic chain of thought (GSM8K 26.0 against 37.5). All baseline numbers are copied from other papers. | Proposed |
| SOTA-347 | To write many facts into a model’s weights, update a range of early-to-mid MLP layers jointly in one batched solve (MEMIT), choosing the range by measured edit success rather than causal tracing | Meng et al. (2023), LIT-576 — sequential single-fact edits collapse after tens of facts (ROME scores 50.3 at 10,000). Distributing each update over a range of MLP layers in one least-squares solve holds 85.8 on GPT-J, with neighborhood success falling from 83.5 to 73.7. Choose the layers by editing performance: per-fact causal tracing does not predict it (LIT-574). Evaluate implications too, because weight editors fail them (LIT-577). | Proposed |
| SOTA-348 | Evaluate a knowledge edit on the facts it implies — logical consequences, multi-hop compositions, aliases and the subject’s other facts — not only on the edited prompt, its paraphrases and unrelated neighbours | Cohen et al. (2023), LIT-577 — the standard edit metrics (efficacy, paraphrase, neighborhood) cannot see whether an edit changed what the model knows or only what it says to one question. On six ripple criteria, ROME, MEMIT and MEND average 38–66, and logical generalization falls as low as 5.5. Prompting with the new fact scores 81–83 on LLaMA-7B. Report ripple metrics beside the standard ones, and compare against in-context editing. | Active |
| SOTA-393 | Freeze the image model’s spatial layers when adding time if checkpoint compatibility matters; do not expect a quality gain from it | Girdhar et al. (2023), LIT-635, and Guo et al. (2023), LIT-633. Freezing a text-to-image model’s spatial layers and training only temporal ones keeps the model a drop-in for checkpoints fine-tuned from the same base. Emu Video’s one controlled test finds freezing costs nothing and gains little (55.0 / 58.1 against fine-tuning). The reason to freeze is compatibility, not quality. | Proposed |
| SOTA-407 v2 | Interpolate the zero-shot and fine-tuned weights at about half way rather than shipping the fine-tuned model | Wortsman et al. (2021), LIT-674. Fine-tuning a zero-shot model spends the robustness that made it worth starting from. Ship (1 − α)·θ_zero-shot + α·θ_fine-tuned at α = 0.5 instead: +3.5 to +23.2 pp under six distribution shifts against the fine-tuned model, with reference accuracy falling by at most 0.3 pp and usually rising, at no cost in training or inference. Requires that the fine-tuned weights were obtained from the zero-shot weights. | Active |
| SOTA-409 | Average the fine-tuned models from your hyperparameter sweep instead of keeping only the best, adding each one in validation order and only if it helps | Wortsman et al. (2022), LIT-675. A hyperparameter sweep ends by discarding every model but one. Average them instead — sorted by held-out accuracy, keeping each only if held-out accuracy improves, which makes the result no worse than the best individual model on that set by construction. O(1) inference against an ensemble’s O(k). +0.7 pp on CLIP and +0.5 pp on ALIGN over the best sweep member; +0.16 on ImageNet and +0.34 under distribution shift for ViT-G. Averaging all of them uniformly instead can come out worse than the best one. | Active |
| SOTA-414 | To make an image diffusion model a dense geometric predictor, fine-tune it end to end as a one-step model at t = T with a task loss | Martin Garcia et al. (WACV 2025), LIT-687. Fix the timestep at T, feed zeros as the noise, decode the prediction through the frozen VAE, and train on the task’s own loss: affine-invariant L1 for depth, angular error for normals. From 20K iterations on 74K synthetic images, the one-step model beats Marigold’s 50-step, 10-member ensemble on depth AbsRel on all five zero-shot sets, and on normals by 2.6–3.0° of mean error. Starting from plain Stable Diffusion is nearly as good. | Proposed |
| SOTA-426 | When you merge and reinitialise a low-rank adapter mid-training, prune the optimizer state and re-warm the learning rate from zero — both, not either | Lialin et al. (2023), LIT-104 — ReLoRA’s restart discipline, read as NOTE-024. At every merge-and-reinit, prune the adapter’s Adam state (99% by magnitude) and drop the learning rate to zero and re-warm it over 50–100 steps. In the 130M ablation (Table 6) the restart alone does nothing (34.25 vs LoRA’s 34.17), the re-warm without the reset does nothing (34.29), the reset without the re-warm diverges, and only the pair helps (29.77). A rule for how to restart, not a recommendation to restart. | Proposed |
| SOTA-442 | To RL-tune a flow-matching generator, roll it out as a same-marginal SDE built from its own velocity, and anchor each step with the closed-form KL to the reference | Liu et al. (2025), LIT-779. A flow model’s ODE sampler is not a stochastic policy. Swapping it for an SDE with the same marginals, built from the trained velocity alone, makes each step an isotropic Gaussian, so GRPO’s ratio and the KL to the reference model are both closed form. On SD3.5-M, GenEval goes 0.63 → 0.95 either way. Without the KL term, DrawBench aesthetic falls to 4.93 from a base of 5.39, against 5.25 with it (Table 2). Quality is judged by reward models only, one run each. | Proposed |
| SOTA-445 | When the base text-to-image diffusion model is better than the models that generated the preference pairs, train on the pairs with Diffusion-DPO rather than fine-tuning on the preferred images | Wallace et al. (2023), LIT-774, §5.5 and Supp. S6. Pick-a-Pic’s images were generated by SDXL-beta and Dreamlike, and SDXL-1.0 is better than both. Supervised fine-tuning on the preferred images helped SD1.5 (55.5% win rate over its base) and made SDXL worse at every learning rate tried. Diffusion-DPO on the same pairs improved SDXL, preferred to the untuned base 70.0% of the time on PartiPrompts. The SDXL SFT arm has no number, there is one run per setting, and no one else has repeated it. | Proposed |
| SOTA-449 | When reward-tuning a classifier-free-guided diffusion model over several rounds of sampling, train the guided prediction at the fixed sampling guidance weight, not the conditional prediction alone | Black et al. (2023), LIT-776, App. E.1 and Fig. 10. Reward fine-tuning should not train the unconditional branch, since the reward depends on the prompt. Training the conditional prediction alone made performance deteriorate after the first round of fine-tuning, which the authors attribute to the guidance weight becoming miscalibrated. Fixing the weight (5 in their runs) and training through the guided prediction used for sampling fixed it. One curve per arm, on reward-weighted regression and JPEG compressibility. | Proposed |