Overview — This repository tracks optimization results from AutoSota pipelines. Papers are included only when the internal ledger marks optimization as successful and optimized_code exists for the paper.

---

🛠️ AutoSota CLI

We also provide AutoSota CLI, a command-line tool for automatically optimizing research codebases.

With AutoSota CLI, you can provide a runnable codebase and an optimization goal, and the tool will iteratively propose strategies, modify code, run experiments, and report improved results.

For installation and usage instructions, please see cli_guide.md.

---

📢 News

• 2026-06-28 🚀 Released v0.3.0 — Ported the multi-agent optimization "brain" into the local CLI: resumable runs (autosota --resume); a continuation supervisor that fixes premature aborts by relaunching until the budget is spent; an independent read-only evaluation pass (evaluation_verdict.json); removal of percentage-target early stopping in favor of a budget-only, structured multi-metric objective with guardrails; deep-research web-search verification (web_search VERIFIED / UNVERIFIED); and a fix for workspace cross-talk.

• 2026-05-15 🔧 Released v0.2.1 — fixes CLI invocation failure on clean machines after global install; adds autosota login for one-time terms-of-service acceptance.

• 2026-05-13 📝 Added detailed configuration instructions for both DeepSeek official direct connection and calling DeepSeek via OpenRouter.

• 2026-05-12 🚀 Released v0.2.0 — Supported direct access to domestic models like DeepSeek (no longer tied to OpenRouter), allowed the use of separate API keys for code optimization and literature review, and added anti-tampering protection for evaluation scripts.

• 2026-04-29 🔧 Released v0.1.1 — Added inter-round pause functionality (-i), allowing you to check progress or add instructions without killing the process; introduced new interactive commands including ask, steer, and sessions.

• 2026-04-22 🎉 Released v0.1.0 — Initial release, achieving a closed loop for automated research code optimization: automatically explores repository configurations, reviews relevant literature, iteratively modifies code, and exports the optimal code and patch upon completion.

• 2026-04-21 💻 Initialized the AutoSOTA CLI project on the laboratory server.

---

🏆 Optimized Papers

Sorted by Paper ID. 🚀 >10%

ID	Paper Title	Ours_Optimization
1	SAVVY: Spatial Awareness via Audio-Visual LLMs through Seeing and Hearing	5.90%
2	Pinet: Optimizing hard-constrained neural networks with orthogonal projection layers	16.72%
3	Discount Model Search for Quality Diversity Optimization in High-Dimensional Measure Spaces	7.32%
4	Decentralized Attention Fails Centralized Signals: Rethinking Transformers for Medical Time Series	4.45%
5	Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability	2.25%
6	PhySense: Sensor Placement Optimization for Accurate Physics Sensing	4.16%
7	Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment	2.68%
8	Mean Flows for One-step Generative Modeling	0.14%
9	Score Matching with Missing Data	0.14%
10	Suitability Filter: A Statistical Framework for Classifier Evaluation in Real-World Deployment Settings	1.56%
11	Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection	0.25%
12	EfficientQAT: Efficient Quantization-Aware Training for Large Language Models	6.08%
13	APPL: A Prompt Programming Language for Harmonious Integration of Programs and Large Language Model Prompts	14.29%
14	PIGuard: Prompt Injection Guardrail via Mitigating Overdefense for Free	3.66%
15	FR-Spec: Accelerating Large-Vocabulary Language Models via Frequency-Ranked Speculative Sampling	13.44%
16	MathFusion: Enhancing Mathematical Problem-solving of LLM through Instruction Fusion	20.97%
17	Synergizing LLMs with Global Label Propagation for Multimodal Fake News Detection	0.83%
18	CoT-based Synthesizer: Enhancing LLM Performance through Answer Synthesis	3.68%
19	Dynamic Scaling of Unit Tests for Code Reward Modeling	2.12%
20	Aristotle: Mastering Logical Reasoning with A Logic-Complete Decompose-Search-Resolve Framework	1.68%
21	Turning Trash into Treasure: Accelerating Inference of Large Language Models with Token Recycling	19.50%
22	Chain-of-Reasoning: Towards Unified Mathematical Reasoning in Large Language Models via a Multi-Paradigm Perspective	3.38%
23	Don’t Reinvent the Wheel: Efficient Instruction-Following Text Embedding based on Guided Space Transformation	17.50%
24	Enhancing Automated Interpretability with Output-Centric Feature Descriptions	2.69%
25	DEEPER Insight into Your User: Directed Persona Refinement for Dynamic Persona Modeling	21.78%
26	A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning	12.40%
27	CiteEval: Principle-Driven Citation Evaluation for Source Attribution	24.10%
28	Segment-Based Attention Masking for GPTs	0.22%
29	Conditional Dichotomy Quantification via Geometric Embedding	14%
30	An Efficient and Precise Training Data Construction Framework for Process-supervised Reward Model in Mathematical Reasoning	1.12%
31	Circuit Stability Characterizes Language Model Generalization	35.80%
32	Personal Travel Solver: A Preference-Driven LLM-Solver System for Travel Planning	4.48%
33	Enhancing Unsupervised Sentence Embeddings via Knowledge-Driven Data Augmentation and Gaussian-Decayed Contrastive Learning	2.81%
34	Ensemble Watermarks for Large Language Models	3.46%
35	Comparing Moral Values in Western English-speaking societies and LLMs with Word Associations	8.59%
36	Synergistic Weak-Strong Collaboration by Aligning Preferences	28.42%
37	Mitigating Confounding in Speech-Based Dementia Detection through Weight Masking	2.45%
38	TinySAM: Pushing the Envelope for Efficient Segment Anything Model	0.90%
39	CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning	0.19%
40	Granite Guardian: Comprehensive LLM Safeguarding	1.37%
41	Auto-Regressive Moving Diffusion Models for Time Series Forecasting	1.07%
42	xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend Decomposition	6.92%
43	VHM: Versatile and Honest Vision Language Model for Remote Sensing Image Analysis	0.67%
44	Elevating Flow-Guided Video Inpainting with Reference Generation	3.40%
45	Unlocking the Power of LSTM for Long Term Time Series Forecasting	0.86%
46	Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation	2.27%
47	TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data	8.77%
48	Proxy-SPEX: Sample-Efficient Interpretability via Sparse Feature Interactions in LLMs	30.30%
49	Hogwild! Inference: Parallel LLM Generation via Concurrent Attention	4%
50	CausalPFN: Amortized Causal Effect Estimation via In-Context Learning	15.86%
51	FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials	0.70%
52	Non-stationary Diffusion For Probabilistic Time Series Forecasting	1.28%
53	$K²$VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic Time Series Forecasting	1.52%
54	TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting	0.36%
55	CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding	6.25%
56	InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic Perspective	1.60%
57	MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification	14%
58	Mind-the-Glitch: Visual Correspondence for Detecting Inconsistencies in Subject-Driven Generation	1.80%
59	Not All Data are Good Labels: On the Self-supervised Labeling for Time Series Forecasting	0.30%
60	IA-GGAD: Zero-shot Generalist Graph Anomaly Detection via Invariant and Affinity Learning	1.83%
61	Hierarchical Shortest-Path Graph Kernel Network	2.20%
62	VisionTS: Visual Masked Autoencoders Are Free-Lunch Zero-Shot Time Series Forecasters	0.80%
63	TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster	13.90%
64	Improving Time Series Forecasting via Instance-aware Post-hoc Revision	10.60%
65	Predicting mutational effects on protein binding from folding energy	15.48%
66	Tropical Attention: Neural Algorithmic Reasoning for Combinatorial Algorithms	15.55%
67	KAN-AD: Time Series Anomaly Detection with Kolmogorov-Arnold Networks	0.89%
68	SEMPO: Lightweight Foundation Models for Time Series Forecasting	0.12%
69	Tree Ensemble Explainability through the Hoeffding Functional Decomposition and TreeHFD Algorithm	36.60%
70	Certified Unlearning for Neural Networks	63.64%
71	Neural MJD: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction	19%
72	One Arrow, Two Hawks: Sharpness-aware Minimization for Federated Learning via Global Model Trajectory	4.37%
73	Regularized Langevin Dynamics for Combinatorial Optimization	0.24%
74	Least squares variational inference	0.02%
75	Tree-Sliced Entropy Partial Transport	0.50%
76	AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation	6.79%
77	Advancing Constrained Monotonic Neural Networks: Achieving Universal Approximation Beyond Bounded Activations	0.30%
78	Conformal Anomaly Detection in Event Sequences	0.11%
79	Differentially Private Federated $k$-Means Clustering with Server-Side Data	2.01%
80	Latent Score-Based Reweighting for Robust Classification	6.58%
81	Meta-Black-Box-Optimization through Offline Q-function Learning	1.01%
82	Efficient Training-Free Online Routing for High-Volume Multi-LLM Serving	1.61%
83	STaRFormer: Semi-Supervised Task-Informed Representation Learning via Dynamic Attention-Based Regional Masking for Sequential Data	3.15%
84	Towards Accurate Time Series Forecasting via Implicit Decoding	2.30%
85	NeuralSurv: Deep Survival Analysis with Bayesian Uncertainty Quantification	23%
86	On the Integration of Spatial-Temporal Knowledge: A Lightweight Approach to Atmospheric Time Series Forecasting	2.42%
87	Multi-Task Vehicle Routing Solver via Mixture of Specialized Experts under State-Decomposable MDP	51.40%
88	Channel Normalization for Time Series Channel Identification	15.20%
89	Distinguishing Cause from Effect with Causal Velocity Models	1.66%
90	Information Bottleneck-guided MLPs for Robust Spatial-temporal Forecasting	0.38%
91	Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains	4.30%
92	Modified K-means Algorithm with Local Optimality Guarantees	0.84%
93	FedWMSAM: Fast and Flat Federated Learning via Weighted Momentum and Sharpness-Aware Minimization	1.77%
94	BounDr.E: Predicting Drug-likeness via Biomedical Knowledge Alignment and EM-like One-Class Boundary Optimization	1.27%
95	Accelerating Feature Conformal Prediction via Taylor Approximation	5.49%
96	Multi-Class Support Vector Machine with Differential Privacy	2.37%
97	Wasserstein Transfer Learning	3.74%
98	X-Mahalanobis: Transformer Feature Mixing for Reliable OOD Detection	1.07%
99	Fast Non-Log-Concave Sampling under Nonconvex Equality and Inequality Constraints with Landing	4.83%
100	Inverse Methods for Missing Data Imputation	2.50%
101	Measure-Theoretic Anti-Causal Representation Learning	0.57%
102	Stochastic Forward-Forward Learning through Representational Dimensionality Compression	1.33%
103	Distributed Conformal Prediction via Message Passing	6.33%
104	Balanced Active Inference	24.80%
105	Aligning Evaluation with Clinical Priorities: Calibration, Label Shift, and Error Costs	2.40%

---

⭐ Star History

---

📝 Citation

If you find AutoSOTA useful for your research, please cite:

@article{li2026autosota,
  title={Autosota: An end-to-end automated research system for state-of-the-art ai model discovery},
  author={Li, Yu and Shao, Chenyang and Liu, Xinyang and Zhao, Ruotong and Liu, Peijie and Su, Hongyuan and Chen, Zhibin and Yang, Qinglong and Xu, Anjie and Fang, Yi and others},
  journal={arXiv preprint arXiv:2604.05550},
  year={2026}
}

---

Per-paper optimization summaries

The paragraphs below are hand-authored for this README (not produced by sync_autosota_list.py). Each paper’s long-form write-up, tables, and logs live in OPTIMIZATION.md when that file exists in the paper folder; otherwise start from README.md. Edit optimization memos under autosota_manual_docs/optimization/ and re-sync.

---

1 — SAVVY

SAVVY: Spatial Awareness via Audio-Visual LLMs through Seeing and Hearing

SAVVY predicts distances between ego-centric and exo-centric views; gains came from calibrating those distances rather than from a bigger model. The pipeline mixed Stable-Diffusion text, piecewise ego/exo scales with clipping, and a final nudge to ego lo_scale so predicted geometry better matches supervised distances.

The closed-loop run exceeded its target comfortably; the emphasis is on stable calibration and bounded outputs rather than on novel architectures.

→ paper-1-SAVVY/OPTIMIZATION.md

---

2 — PINet

PINet: Optimizing hard-constrained neural networks with orthogonal projection layers

PINet’s cost driver is the differentiable projection solve inside the loop. The best configuration cut test-time DR iterations (n_iter_test 50→10) because the bundled QP is small and was over-solved, disabled JAX float64 in favor of float32 on A100, and swapped ReLU for SiLU to match fused kernels better.

Together these give a clear drop in small-batch inference time versus the paper baseline while still meeting the hard feasibility constraints the method is built around.

→ paper-2-PINet/OPTIMIZATION.md

---

3 — DMSQD

Discount Model Search for Quality Diversity Optimization in High-Dimensional Measure Spaces

Quality–diversity here is driven by a multi-emitter CMA-ES style search. Increasing emitters per domain (15→20) adds parallel search threads with diverse covariances, which improves coverage of the measure space and raises the averaged QD score across benchmark domains without changing the core algorithm.

The change is simple to describe but effective because diversity metrics are sensitive to how many independent search processes are exploring at once.

→ paper-3-DMSQD/OPTIMIZATION.md

---

4 — DecentralAttn

Decentralized Attention Fails Centralized Signals: Rethinking Transformers for Medical Time Series

PTB accuracy rose 0.8431→0.8806 at iter9 after scaling d_model to 384, v_layer=4, and d_core=d_model//2 with label smoothing + AdamW; iter8 already beat 0.86 once d_core moved off d_model//4. Iter10–12 (wider model, dropout mask, deeper v) regressed.

→ paper-4-DecentralAttn/OPTIMIZATION.md

---

5 — TSAE

Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability

For SAEBench autointerp, the model highlights tokens for the interpreter LLM using an activation threshold. Raising act_threshold_frac from 0.01 to 0.05 keeps only high-contrast activations, which gives the explainer a cleaner mask and improves the autointerp score at modest cost.

This is a measurement-and-supervision tweak: the sparse autoencoder is unchanged, but the interface presented to the judge model is less noisy.

→ paper-5-TSAE/OPTIMIZATION.md

---

6 — PhySense

PhySense: Sensor Placement Optimization for Accurate Physics Sensing

PhySense optimizes sensor layouts under stochastic physics; variance reduction mattered as much as the mean. Antithetic sampling (paired ± noise) with a moderate ensemble size (K=25 pairs) cut variance in the placement objective, and the pipeline stepped K upward until the lower-is-better relative_l2 stabilized.

The final solution keeps the same forward model but averages Monte Carlo estimates more efficiently.

→ paper-6-PhySense/OPTIMIZATION.md

---

7 — ReasoningIQA

Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment

PLCC improved 0.7803→0.8012 by fusing CLS logits with patch_mean / patch_max instead of a single pooling path—explicit blend weights let global reasoning and local artifacts both vote on perceived quality.

→ paper-7-ReasoningIQA/OPTIMIZATION.md

---

8 — MeanFlows

Mean Flows for One-step Generative Modeling

FID improved slightly (2.8112→2.8074, lower is better) using an EMA blend (~98.7% slow EMA + ~1.3% live weights). Seed sweeps, two-step ODE paths, and three-way EMA soups did not beat the shipped checkpoint.

→ paper-8-MeanFlows/OPTIMIZATION.md

---

9 — ScoreMissing

Score Matching with Missing Data

Iter4 hit AUC≈0.7691 with a multi-model ensemble, while final reporting averaged 0.7613 over three eval passes—variance and rubric targets diverge, so read “best iterate” vs “mean eval” separately. Fixing device placement in the scoring script was prerequisite noise reduction.

→ paper-9-ScoreMissing/OPTIMIZATION.md

---

10 — SuitabilityFilter

Suitability Filter: A Statistical Framework for Classifier Evaluation in Real-World Deployment Settings

Suitability OOD score improved to 0.9838 (from 0.9687) by combining isotonic calibration, 10-fold multi-fold training, and a 3-feature subset (conf / logit / loss) aggregated with Stouffer Z-score; this combination gave the best stability in later iterations.

→ paper-10-SuitabilityFilter/OPTIMIZATION.md

---

11 — OSD

Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection

Final BigGAN accuracy reached 99.6 (from 99.35) with an inference-side recipe: horizontal-flip TTA, L2 feature normalization, and LEAP-style feature usage (CLS + patch mean), then a final asymmetric 3-view TTA weighting orig/flip/cls-only = 0.7/0.2/0.1.

→ paper-11-OSD/OPTIMIZATION.md

---

12 — EfficientQAT

EfficientQAT: Efficient Quantization-Aware Training for Large Language Models

WikiText-2 PPL 7.1654→6.73 (≈−6.08%) by post-hoc calibration of quantization scales and RMSNorm weights (Adam on a small train slice; int2 qweight/qzeros untouched), after bf16 + SDPA stabilization in QuantLinear. C4 ticks up slightly (8.9043→8.95); the rubric headline is WikiText-2.

→ paper-12-EfficientQAT/OPTIMIZATION.md

---

13 — APPL

APPL: A Prompt Programming Language for Harmonious Integration of Programs and Large Language Model Prompts

The AST-size metric rewards compact programs. Inlining marginalize into the return path removes a redundant assignment chain, shaving nodes without changing semantics.

It is a classic compiler-style cleanup chosen because the optimizer’s score is literally the AST node count.

→ paper-13-APPL/OPTIMIZATION.md

---

14 — PIGuard

PIGuard: Prompt Injection Guardrail via Mitigating Overdefense for Free

PIGuard balances injection detection with benign utility. Lowering the decision threshold (0.5→0.10) while keeping top_k=None allows borderline benign prompts—often containing trigger-like phrases—to stay classified as benign, which raises over-defense accuracy substantially across splits.

The backbone stays DeBERTa; the improvement is almost entirely decision-rule tuning on a well-calibrated scorer.

→ paper-14-PIGuard/OPTIMIZATION.md

---

15 — FRSpec

FR-Spec: Accelerating Large-Vocabulary Language Models via Frequency-Ranked Speculative Sampling

Throughput on the large-vocab MT-Bench setting improved by tightening speculative decoding knobs (num_iter, tree_size) and fixing stability issues uncovered across optimizer iterations. FR-Spec’s tree-based drafts need both width and depth to stay ahead of verification costs.

The final configuration clears an aggressive tokens/sec target versus both the internal baseline and the paper-reported reference.

→ paper-15-FRSpec/OPTIMIZATION.md

---

16 — MathFusion

MathFusion: Enhancing Mathematical Problem-solving of LLM through Instruction Fusion

Accuracy gains came from a stronger instruction-fusion recipe: a dedicated CoT+boxed template, four in-context shots, and enough generation budget for full derivations. Weaker shot counts and alternate prefixes were tried but did not beat the best configuration.

The result is a clear win on the benchmark suite’s primary accuracy metric with reproducible prompting only.

→ paper-16-MathFusion/OPTIMIZATION.md

---

17 — MultimodalGLP

Synergizing LLMs with Global Label Propagation for Multimodal Fake News Detection

Skipping full repro, the loop still improved accuracy by deepening graph propagation: n_label_iters 1→3 on train/test so pseudo-labels stabilize before metrics fire. Multi-seed checks caught a shallow-iter regression and forced a rollback before the final config landed.

→ paper-17-MultimodalGLP/OPTIMIZATION.md

---

18 — CoTSynth

CoT-based Synthesizer: Enhancing LLM Performance through Answer Synthesis

The synthesizer is treated as one vote among six: majority voting over assistant answers, with gold labels taken from a clean data[answer] field, outperformed fancier grouping ideas tried later. Stochastic synthesis errors are damped when its output is not allowed to dominate.

Incremental tweaks after that baseline did not move the needle further.

→ paper-18-CoTSynth/OPTIMIZATION.md

---

19 — DyScaleUT

Dynamic Scaling of Unit Tests for Code Reward Modeling

Pass@1 for code reward modeling rose after tightening what counts as a passing unit test (stricter than 50% case pass rate), then re-ranking with variance-aware weights over a filtered UT set. Noisy tests that never discriminate solutions are down-weighted.

Each step is interpretable: threshold, filter, aggregate.

→ paper-19-DyScaleUT/OPTIMIZATION.md

---

20 — Aristotle

Aristotle: Mastering Logical Reasoning with A Logic-Complete Decompose-Search-Resolve Framework

Logical proofs that branch on “true” paths benefit from deeper negation search; increasing search_round on that branch (10→20) helped the engine find contradictions the shallow search missed. A heavier model swap was attempted and rolled back when it did not help.

The takeaway is search-depth tuning on the structured proof side, not prompt fluff.

→ paper-20-Aristotle/OPTIMIZATION.md

---

21 — TokenRecycling

Turning Trash into Treasure: Accelerating Inference of Large Language Models with Token Recycling

Token-recycling speculative decoding needed joint tuning of draft tree / MAT geometry and recycle policy. Mean accepted tokens rose ~+19.5% with the best optimizer score near iter 10; later width experiments that ignored verification cost were rolled back.

→ paper-21-TokenRecycling/OPTIMIZATION.md

---

22 — Chain-of-Reasoning

Chain-of-Reasoning: Towards Unified Mathematical Reasoning in Large Language Models via a Multi-Paradigm Perspective

Self-consistency (N=5, temperature 0.6) with majority vote on the main math task lifted accuracy meaningfully over the single-sample baseline. Diversity in reasoning paths reduces variance on competition-style questions.

It is a standard inference-time compute trade that paid off here without retraining.

→ paper-22-ChainOfReasoning/OPTIMIZATION.md

---

23 — GuidedEmbed

Don't Reinvent the Wheel: Efficient Instruction-Following Text Embedding based on Guided Space Transformation

With repro skipped, SOTA focused on training/eval hygiene for the guided transformer embedding. v_measure climbed 36.05→42.37 after a dozen iterations (strongest around iter 12), mostly from LR/batch and projection-schedule moves rather than architectural edits.

→ paper-23-GuidedEmbed/OPTIMIZATION.md

---

24 — OutputCentric

Enhancing Automated Interpretability with Output-Centric Feature Descriptions

The headline metric was switched to VocabProj output success instead of an ensemble-concatenated score. That single evaluation choice aligned training signals with how features are judged and immediately cleared the target.

Architecture stayed the same; the improvement is definitional honesty between training objective and reported metric.

→ paper-24-OutputCentric/OPTIMIZATION.md

---

25 — DEEPER

DEEPER Insight into Your User: Directed Persona Refinement for Dynamic Persona Modeling

Persona MAE fell after enriching prompts with explicit rating history and simple rating statistics, then applying domain-aware floor calibration for users who only give top scores. The model sees both narrative persona text and concrete behavioral evidence.

Post-processing for skewed domains prevents optimistic drift on always-five-star users.

→ paper-25-DEEPER/OPTIMIZATION.md

---

26 — GARTKG

A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning

Primary accuracy improved 49.44%→55.59% on the final _best export with replay weight 0.3; an earlier ~64% iterate may be invisible if later jobs overwrote the same checkpoint name—trust the ledger plus score notes for lineage.

→ paper-26-GARTKG/OPTIMIZATION.md

---

27 — CiteEval

CiteEval: Principle-Driven Citation Evaluation for Source Attribution

Pearson (statement-level) jumped 0.733→0.910 after hardening evaluate_metric.py against all-none ratings, filling Nones sensibly, and using weighted / piecewise / power transforms so outliers do not dominate the correlation.

→ paper-27-CiteEval/OPTIMIZATION.md

---

28 — SBAM

Segment-Based Attention Masking for GPTs

Small average accuracy gains across eight tasks came from using training-style prompts (indentation and blank lines) on ARC-style items while leaving other tasks on their original templates. Formatting matters for instruction-tuned models even when the underlying weights are fixed.

The change is per-task prompt hygiene, not mask architecture redesign.

→ paper-28-SBAM/OPTIMIZATION.md

---

29 — CDQGeoEmbed

Conditional Dichotomy Quantification via Geometric Embedding

avg_dcf moved 0.4907→0.5594 (+14%) after sequential cross-scenario fine-tuning to ckpts/cross_scenario/defeasible_bert_v1d_nli and inference-time 0.7 / 0.3 embedding fusion (fine-tuned vs. original defeasible-bert). A new eval_all_scenarios.py runs the ensemble; DCF math and datasets were left unchanged (run_20260324_211015).

→ paper-29-CDQGeoEmbed/OPTIMIZATION.md

---

30 — ProcessRM

An Efficient and Precise Training Data Construction Framework for Process-supervised Reward Model in Mathematical Reasoning

GSM8K F1 improved with a finer threshold sweep on the process labels and by swapping sigmoid for softmax in the verifier head so multi-class calibration is less brittle. Grid search resolution matters when thresholds sit on steep parts of the ROC.

No new data was collected—only how existing labels are turned into scores.

→ paper-30-ProcessRM/OPTIMIZATION.md

---

31 — CircuitStability

Circuit Stability Characterizes Language Model Generalization

The target accuracy metric responds strongly to evaluation-time compute: longer max_new_tokens and more in-context shots (up to roughly 8–10) let the model complete structured reasoning tasks that were previously truncated or under-primed.

It is an inference-budget story on a circuit-focused benchmark.

→ paper-31-CircuitStability/OPTIMIZATION.md

---

32 — PTSolver

Personal Travel Solver: A Preference-Driven LLM-Solver System for Travel Planning

Pass rate 86.45%→90.32% from generate_plans_v2.py: dropped room_type pre-filtering in get_accommodation (evaluator uses lowercase labels; queries use title case, so the filter was excluding valid cheap rooms), plus get_best_transport_mode so round trips use one transport mode with both-direction distance-matrix checks (run_20260324_212959).

→ paper-32-PTSolver/OPTIMIZATION.md

---

33 — GCSE

Enhancing Unsupervised Sentence Embeddings via Knowledge-Driven Data Augmentation and Gaussian-Decayed Contrastive Learning

Sentence embedding quality gained from a weighted ensemble of three encoders (2:1:1) centered on GCSE-RoBERTa-large plus an auxiliary RoBERTa-base view using first-and-last hidden states. Diversity of representation families helps downstream retrieval.

Training recipes stay within the paper’s family; the win is model soup, not a new loss.

→ paper-33-GCSE/OPTIMIZATION.md

---

34 — EnsembleWM

Ensemble Watermarks for Large Language Models

Detection rate rose by re-running the detector on texts that initially looked un-watermarked and by adding windowed prefix checks so short spans cannot evade the test. Perplexity stayed flat, so the extra scrutiny does not break generation quality.

It is defensive depth on the verification side rather than a new watermark signal.

→ paper-34-EnsembleWM/OPTIMIZATION.md

---

35 — MoralValuesWA

Comparing Moral Values in Western English-speaking societies and LLMs with Word Associations

Care-dimension correlation improved after log1p-smoothing the word-graph adjacency and running a short α-propagation that emphasizes the Care axis. The graph algorithm is simple; the gain is from not overweighting hub words.

This is interpretable feature engineering on the association graph, not a bigger LM.

→ paper-35-MoralValuesWA/OPTIMIZATION.md

---

36 — WeakStrongPref

Synergistic Weak-Strong Collaboration by Aligning Preferences

F1 on the preference task jumped with a four-way candidate pool (DPO, SFT, base, GPT-4), aggressive answer normalization, and fuzzy matching on explanation text. The pipeline still reflects an optimistic upper bound because candidate selection uses oracle knowledge—worth noting when comparing to fully blind systems.

Even with that caveat, the engineering of normalization and matching is what unlocked the metric move.

→ paper-36-WeakStrongPref/OPTIMIZATION.md

---

37 — DementiaMask

Mitigating Confounding in Speech-Based Dementia Detection through Weight Masking

AUPRC improved 0.8282→0.8485 by training longer (epochs 20→30) and relaxing early-stopping patience (5→8) so the confound-sensitive mask can finish stabilizing.

→ paper-37-DementiaMask/OPTIMIZATION.md

---

38 — TinySAM

TinySAM: Pushing the Envelope for Efficient Segment Anything

Starting from the paper-reported baseline of 42.3% AP (IoU=0.50:0.95 on COCO val2017), optimization reached 43.2% AP after 16 iterations — a +0.9% absolute improvement (+2.1% relative), exceeding the target of 43.146%. Key changes: (1) lowering mask binarization threshold from 0.0 to -1.0, and (2) test-time augmentation with H+V flip + centroid refinement.

→ paper-38-TinySAM/OPTIMIZATION.md

---

39 — CALF

CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning

CALF’s best MSE came from turning off the output-consistency loss so the temporal branch is not yoked to the text branch’s predictions during fine-tuning. Follow-on trials with deeper LoRA stacks or mixed losses did not beat that point.

The lesson is that auxiliary alignment terms can hurt when evaluation only scores the time-series head.

→ paper-39-CALF/OPTIMIZATION.md

---

40 — Granite Guardian

Granite Guardian: Comprehensive LLM Safeguarding

RH detection AUC rose by combining many Granite risk heads in logit space with carefully chosen negative weights on signals that fire on benign refusals, so harmful prompts stay separated without hand-tuning a single harm score.

→ paper-40-GraniteGuardian/OPTIMIZATION.md

---

41 — AMDM

Auto-Regressive Moving Diffusion Models for Time Series Forecasting

The biggest win was making the LR scheduler actually run inside the short 2k-step budget (patience 4000→100) plus slightly faster EMA and larger gradient accumulation; together they shaved MSE about 1.5% versus the reproduced baseline.

→ paper-41-AMDM/OPTIMIZATION.md

---

42 — xPatch

xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend Decomposition

Training with plain MSE instead of a surrogate arctan loss, fixing NumPy 2.x np.Inf, and aligning seq_len with the paper’s longer context path unlocked a large drop in average forecast MSE across horizons.

→ paper-42-xPatch/OPTIMIZATION.md

---

43 — VHM

VHM for AID scene classification

Accuracy improved with test-time input resolution and sharper crop scaling for the EVA-CLIP tower; gains are modest because the 7B LLM head was already near ceiling on AID.

→ paper-43-VHM/OPTIMIZATION.md

---

44 — RGVI

Elevating Flow-Guided Video Inpainting with Reference Generation

PSNR on HQVI beat the +2% target by better balancing reference-frame attention, temporal smoothness, and flow guidance so textured regions (forest, garden) gain more than flat backgrounds.

→ paper-44-RGVI/OPTIMIZATION.md

---

45 — PsLSTM

P-sLSTM: Unlocking the Power of LSTM for Long Term Time Series Forecasting

The key improvement was fixing a dropout bug where the CLI argument was not being passed to xLSTMBlockStackConfig, combined with MC Dropout ensemble at inference time. Achieved -0.86% MSE improvement.

→ paper-45-PsLSTM/OPTIMIZATION.md

---

46 — NonStatTS

Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation

Single-line change: increasing GATING_INIT from 0.01 to 0.1 gave the calibration module more immediate influence over predictions, achieving -0.38% MSE improvement.

→ paper-46-NonStatTS/OPTIMIZATION.md

---

47 — TimePFN

TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data

The key improvement was switching the learning rate schedule from type1 (0.5^epoch halving) to type3 with decay=0.8 starting from epoch 3. This gave a massive -8.77% MSE improvement over the paper baseline. Longer input sequences and cosine annealing were explored but did not yield further gains.

→ paper-47-TimePFN/OPTIMIZATION.md

---

48 — Shapiq

shapiq: Shapley Interactions for Machine Learning

precision_at_10 improved 0.76→0.99 using StratifiedBySize (stratify_coalition_size=True, no intersection stratification), pairing_trick, N_ENSEMBLE=100, and large-prime seeds in eval_shapiq.py; P@5 and error metrics improved in lockstep (run_20260324_211015).

→ paper-48-Shapiq/OPTIMIZATION.md

---

49 — HogwildInference

Hogwild! Inference: Parallel LLM Generation via Concurrent Attention

Only a small TPS gain survived contact with reality: removing redundant contiguous() calls in the hot path. Larger experiments (AWQ kernels, Triton, async schedulers) were rolled back when they regressed stability or throughput.

Sometimes the winning patch is micro-optimization plus saying no to risky rewrites.

→ paper-49-HogwildInference/OPTIMIZATION.md

---

50 — CausalPFN

CausalPFN: Amortized Causal Effect Estimation via In-Context Learning

On IHDP, PEHE (lower better) dropped 0.1829→0.1539 via propensity feature augmentation (P(T=1|X) appended as extra feature), plus multi-seed bootstrap ensembling (seeds=[42,43], N_BOOT=3/seed, BOOT_FRAC=0.92) and multi-temperature prediction mixing (T=[0.3,0.5,0.7,0.9,2,4,8]), with best at run_20260325_014145.

→ paper-50-CausalPFN/OPTIMIZATION.md

---

51 — FlashTP

FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials

The CUDA extension was rebuilt with aggressive host flags (-O3, --use_fast_math). Separate CUDA Graph micro-benchmarks appear in logs but should be interpreted carefully—they are not always apples-to-apples with the baseline timing definition.

Treat FlashTP as a story about compiler flags plus careful measurement hygiene.

→ paper-51-FlashTP/OPTIMIZATION.md

---

52 — NsDiff

Non-stationary Diffusion For Probabilistic Time Series Forecasting

ETTh1 CRPS improved once validation-selected checkpoints were bypassed: a fixed backup epoch (6) after training captures models that generalize to the test months even when val CRPS is noisy on only eight batches.

→ paper-52-NsDiff/OPTIMIZATION.md

---

53 — K²VAE

K²VAE — probabilistic time series forecasting

The first win was evaluation fidelity—raising num_samples and quantiles_num slashed CRPS variance. The final edge came from accumulate_grad_batches=2 for smoother updates; test-time input noise (TTA) only hurt because the decoder already marginalizes uncertainty.

→ paper-53-K2VAE/OPTIMIZATION.md

---

54 — TimeBase

TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting

avg_mse fell 0.1684→0.16485 (beats target 0.165) by adding GELU in the basis pathway, growing basis_num to 30, and on pred_720 disabling use_orthogonal while using lr=4e-2 with uniform ow=0.02.

→ paper-54-TimeBase/OPTIMIZATION.md

---

55 — CSBrain

CSBrain: Cross-Scale Spatiotemporal Brain Foundation Model for EEG Decoding

An 8-model weighted ensemble combining diverse checkpoints (original + multiple seed-trained + train-only variants) achieved a +10.8% balanced accuracy improvement. The key insight is that checkpoint diversity matters more than individual model improvements. Gaussian noise TTA provided additional marginal gains.

→ paper-55-CSBrain/OPTIMIZATION.md

---

56 — InfoSAM

InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic Perspective

Key insight: SAM's raw logit outputs need sigmoid conversion before metric computation. Applying sigmoid(0.65 * logits) as post-processing + horizontal flip TTA improved S-measure by +1.80%.

→ paper-56-InfoSAM/OPTIMIZATION.md

---

57 — MDReID

MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification

On RGBNT201, mAP 82.1%→93.6% and Rank-1 85.2%→91.6% by enabling K-reciprocal re-ranking (reranking=True), tuning re_ranking to k1=60, k2=22, lambda=0 (pure Jaccard), and ×2.0 scaling of the second half of the 3072-d shared features before L2 norm (run_20260325_020748).

→ paper-57-MDReID/OPTIMIZATION.md

---

58 — MindGlitch

Mind-the-Glitch: Visual Correspondence for Detecting Glitches in Cultural Heritage Docs

Horizontal flip TTA with weighted averaging (3orig + 1flip)/4 improved Spearman correlation from 0.5826 to 0.6006 (+3.09%), exceeding the target.

→ paper-58-MindGlitch/OPTIMIZATION.md

---

59 — SelfSupervised

Not All Data are Good Labels: On the Self-supervised Learning of Time Series

Increased ensemble size (num_models=16, num_series=16) combined with CosineAnnealingLR improved avg_mse by -0.65%. Target not fully reached but real improvement achieved.

→ paper-59-SelfSupervised/OPTIMIZATION.md

---

60 — IAGGAD

IA-GGAD: Zero-shot Generalist Graph Anomaly Detection

Single most impactful change: increasing training epochs from 40 to 300 allowed the auxiliary GCN to converge properly, improving AUROC on ACM by +1.97%, exceeding the target.

→ paper-60-IAGGAD/OPTIMIZATION.md

---

61 — HSGKN

Hierarchical Shortest-Path Graph Kernel Network

Key improvements: extended training (500→2000 epochs), increased dropout (0.15→0.25), and lightweight SE-Net channel attention. Combined improved accuracy by +2.20%, exceeding the target.

• IMDB-Binary: 77.7 → 79.9 (+2.83%)
• IMDB-Multi: 55.53 → 55.80 (+0.49%)

→ paper-61-HSGKN/OPTIMIZATION.md

---

---

62 — VisionTS

VisionTS: Visual Masked Autoencoders Are Free-Lunch Zero-Shot Time Series Forecasters

Average forecasting MSE improved with a three-scale ensemble over context lengths (2880 / 1440 / 720) and weights 0.6 / 0.25 / 0.15, plus raising norm_const to 0.6 so pixel dynamics use more of the image range. Long windows carry seasonality; short windows track recent shocks.

Ensembling different temporal footprints is the robustness mechanism.

→ paper-62-VisionTS/OPTIMIZATION.md

---

63 — TSRAG

TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster

MAE improved 0.4261→0.3668 (13.9% reduction) after reducing MoE skip connection scale from 1.0 to 0.1. The original skip connection was over-powering base model predictions at inference time with single retrieved context. Additional gains from gate temperature sharpening (T=0.3) and 3-point moving average smoothing.

→ paper-63-TSRAG/OPTIMIZATION.md

---

64 — PIR

PIR: Improving Time Series Forecasting via Instance-aware Post-hoc Revision

Critical fix: switching lradj from 'type1' (halves LR every epoch) to 'type3' (0.9x decay from epoch 4+) allowed the PIR refinement module to fully converge, achieving -10.6% MSE improvement.

→ paper-64-PIR/OPTIMIZATION.md

---

65 — ProteinBinding

Predicting mutational effects on protein binding from folding energy

Pearson correlation on per-interface metrics rose after blending StaB predictions with FoldX physics scores (best α≈0.40) and fixing mutation-string alignment so paired rows actually match. Before the string fix, a third of rows failed to join.

Hybrid statistical plus physics models and clean keys beat either alone.

→ paper-65-ProteinBinding/OPTIMIZATION.md

---

66 — TropicalAttention

Tropical Attention: Neural Algorithmic Reasoning for Combinatorial Algorithms

Quickselect OOD F1 jumped 70.69→81.68 after twelve iterations. The durable trick was length-16 auxiliary batches mixed in with probability ~0.3, forcing the tropical attention pathway to handle longer prefixes without ruining in-distribution accuracy.

→ paper-66-TropicalAttention/OPTIMIZATION.md

---

67 — KAN-AD

KAN-AD: Time Series Anomaly Detection with Kolmogorov-Arnold Networks

Best F1 improved 0.9106→0.9187 by switching to sin+cos Fourier basis (order=3), pairing it with CosineAnnealingLR + patience=5, and using variance-normalized anomaly scoring (alpha=0.5, local_std_window=16) to stabilize threshold behavior.

→ paper-67-KAN-AD/OPTIMIZATION.md

---

68 — SEMPO

SEMPO: Lightweight Foundation Models for Time Series Forecasting

Only marginal improvement (-0.12% MSE) was achieved despite extensive exploration. The foundation model architecture appears well-optimized by default with limited headroom for simple hyperparameter tuning. Multi-head stacking for long prediction horizons gave a tiny improvement.

→ paper-68-SEMPO/OPTIMIZATION.md

---

69 — TreeHFD

Tree Ensemble Explainability through the Hoeffding Functional Decomposition and TreeHFD Algorithm

mse_eta_1_2 0.0354→0.0345 (−2.54%) by up-weighting hierarchical orthogonality in the lsqr stack: constr_ortho *= 8.0 in src/treehfd/optimization_matrix.py, sharpening separation between η({1,2}) and main effects so the (1,2) reconstruction error drops (target ≤0.0348 met).

→ paper-69-TreeHFD/OPTIMIZATION.md

---

70 — CertifiedUnlearning

Certified Unlearning for Neural Networks

Post-unlearning fine-tuning originally oscillated and needed many epochs to reach 50% validation accuracy. Adding SGD momentum (0.9) and a higher max learning rate (0.2) smoothed optimization and cut epochs-to-target dramatically.

The headline metric is efficiency of the unlearning fine-tune, not raw accuracy.

→ paper-70-CertifiedUnlearning/OPTIMIZATION.md

---

71 — Neural MJD

Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction

Monte Carlo path sampling mattered: raising n_runs into the hundreds with antithetic variates stabilized jump-diffusion estimates, and nudging w_cond_mean_loss upward aligned training with the mean metric being scored.

→ paper-71-NeuralMJD/OPTIMIZATION.md

---

72 — FedGMT

One Arrow, Two Hawks: Sharpness-aware Minimization for Federated Learning via Global Model Trajectory

Latest run now reports 80.76% from 79.57% (+1.50% by ledger). The winning recipe is trajectory smoothing + late SWA: set alpha=0.99, start SWA from round 400, evaluate with the SWA model, and use that same SWA model as the late-round client teacher (servergmt.py), which stabilizes the last-100-round optimization phase under Dir(0.1).

→ paper-72-FedGMT/OPTIMIZATION.md

---

73 — RLD

Regularized Langevin Dynamics for Combinatorial Optimization

Population-based chain refresh (LEAP) mechanism improved rlsa_size by +0.235%. The key insight: introducing diversity between chains via copying best solutions to worst chains helps escape local optima.

→ paper-73-RLD/OPTIMIZATION.md

---

74 — LSVI

Least squares variational inference

KL 367.7902→367.7304 (↓0.0598, ~0.016%) with Gaussian BBVI (8 seeds, long phase1/phase2 Adam + reparameterization, best-iterate pick). Fixed small-K eval makes LSVI regress vs Laplace; BBVI lands near the Gaussian ELBO floor, so the −2% rubric is not reachable without a richer posterior or different eval.

→ paper-74-LSVI/OPTIMIZATION.md

---

75 — TreeSlicedEntropy

Tree-Sliced Entropy Partial Transport

Target accuracy improved 0.8678→0.8721 by combining gen_mode = gaussian_orthogonal sampling with a thicker slice fan (--twd_nlines 8). The paper’s +2% stretch goal is still open, but the pipeline beat the internal baseline cleanly.

→ paper-75-TreeSlicedEntropy/OPTIMIZATION.md

---

76 — AANet

AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation

Virtual screening BEDROC (holo) moved 0.6548→0.6993, clearing rubric 0.6679, by ensembling three CroppingPocket seeds and z-scoring adapt vs max docking channels before a 50/50 fusion so scale differences do not swamp signal.

→ paper-76-AANet/OPTIMIZATION.md

---

77 — CMNN

Advancing Constrained Monotonic Neural Networks

Extended training (1000→1500 epochs) provided marginal -0.30% improvement. The model appears near its convergence plateau with fixed LR schedule being optimal.

→ paper-77-CMNN/OPTIMIZATION.md

---

78 — ConformalAnomaly

Conformal Anomaly Detection in Event Sequences

At ~99.3% AUROC there is little headroom; twelve iterations still eked +0.11 pts by growing the Weibull mixture to 24 components and cooling the optimizer to 1e-4 once residuals were well calibrated.

→ paper-78-ConformalAnomaly/OPTIMIZATION.md

---

79 — DPFKMeans

Differentially Private Federated k-Means Clustering with Server-Side Data

Train k-means cost 49.9554→48.9508 (−2.01%, rubric met) by re-tuning DP YAML: more ε to Gaussian sum terms (split 0.8), aggressive fedlloyds_clipping_bound cuts (noise dial under datapoint privacy), 100 server samples per mixture + minimum_server_point_weight=1, and variance=0.49 so the non-private floor and DP cost align at ~48.95 with ~98.6% accuracy.

→ paper-79-DPFKMeans/OPTIMIZATION.md

---

80 — LatentScoreReweight

Latent Score-Based Reweighting for Robust Classification

Worst-group accuracy rose 69.1%→73.65% once evaluation bugs were fixed and checkpoints were selected by group-robust scores instead of averages. Remaining gains were classic LR / weight_decay tuning on top of a trustworthy metric.

→ paper-80-LatentScoreReweight/OPTIMIZATION.md

---

81 — QMamba

Meta black-box optimization via offline Q-function learning (Q-Mamba)

Mean BBOB reward improved by extending optimization trajectories and keeping decisions strictly greedy—temperature-softened actions destroy the normalized Q-values Q-Mamba relies on.

→ paper-81-QMamba/OPTIMIZATION.md

---

82 — OnlineLLMRouting

Efficient Training-Free Online Routing for High-Volume Multi-LLM Serving

PORT quality improved 2748.6→2792.8 using fixed seeds, analytic gradient parsing for the surrogate, distance-weighted ANN retrieval (exp kernel, ×4 neighbor budget), and tighter L-BFGS-B tolerances so the routing QP stops dithering.

→ paper-82-OnlineLLMRouting/OPTIMIZATION.md

---

83 — STaRFormer

STaRFormer on the PAM activity dataset

Latest run lands at 0.9753 from 0.9663 (+0.90%, and +0.60% vs paper 0.9693). The strongest single lever is lowering contrastive temperature to 0.2; combining it with d_model=64 and lambda_cl=0.3 gives the best synergy. Deeper stacks, temp=0.1, and removing contrastive loss all regress, and the +2% stretch target remains out of reach under hyperparameter-only tuning.

→ paper-83-STaRFormer/OPTIMIZATION.md

---

84 — IFT

Implicit Forecasting Transformer — ECL forecasting

MSE gains stacked wider layers (d_model 1024) with MAE loss, RevIN affine, and an MC dropout ensemble (n=10) at inference to shave variance on the 321-channel electricity benchmark.

→ paper-84-IFT/OPTIMIZATION.md

---

85 — NeuralSurv

NeuralSurv: Deep Survival Analysis with Bayesian Uncertainty Quantification

Harrell’s C-index jumped 0.5495→0.6759 at the best iteration (~iter 3) after switching activations to SiLU and letting CAVI run long enough. Later iterations sometimes failed outright—small survival cohorts mean bootstrap CIs are mandatory when reporting externally.

→ paper-85-NeuralSurv/OPTIMIZATION.md

---

86 — STELLA

STELLA — global wind forecasting with spatial-temporal structure

test_MAE dropped after enabling learnable spatial embeddings (if_rel=True), training 150 epochs so embeddings converge, and doubling the base LR with the same milestone decay shape so optimization escapes shallow minima without bloating width (which overfit).

→ paper-86-STELLA/OPTIMIZATION.md

---

87 — MoSES

Multi-Task Vehicle Routing Solver via MoSES

Optimality gap collapsed after aggressive symmetric tour augmentation, larger batch search, and lightweight Or-opt style post-processing—even though wall-clock inference rose sharply, the metric-first goal was exceeded by a wide margin.

→ paper-87-MoSES/OPTIMIZATION.md

---

88 — ChannelNorm

Channel Normalization for Time Series Channel Identification

A massive -15.2% MSE improvement came from extending the input sequence length from 96 to 336, which captures weekly seasonality in electricity data. Combined with cosine LR annealing, dropout in temporal blocks, and a second MLP residual layer, this exceeded the target by a wide margin.

→ paper-88-ChannelNorm/OPTIMIZATION.md

---

89 — CausalVelocity

Distinguishing Cause from Effect with Causal Velocity Models

AUDRC 89.58%→91.07% after doubling Stein n_steps (100→200) and switching to the squared GoF path; bandwidth / regularizer / trimming experiments that hurt AUC were discarded.

→ paper-89-CausalVelocity/OPTIMIZATION.md

---

90 — RSTIB

Information Bottleneck-guided MLPs for Robust Spatial-temporal Forecasting

MAE improved 18.4928→18.4219 (0.38% reduction) by reducing info_beta from 0.001 to 0.0. The IB regularization was over-compressing the representation. Additional marginal gains from increasing n_sample (12→50) for variance reduction. Target (18.1229) was not reached; within ~1.6% gap remaining.

→ paper-90-RSTIB/OPTIMIZATION.md

---

91 — TimeAwareCausal

Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains

Average accuracy 86.0%→89.7% at iter8 with weight_decay=1e-4, masker_middle=6×, and dropout=0.1; wider masks, label smoothing, or harsher dropout oversparsified the evolving-domain representation.

→ paper-91-TimeAwareCausal/OPTIMIZATION.md

---

92 — KmeansLocalOpt

Modified K-means Algorithm with Local Optimality Guarantees

With 20 independent kmeans++ restarts per trial, ALL trials converge exactly to the same value, strongly suggesting 801207 is the global D-local optimum. The kmeans++ initialization gave the biggest single improvement (-0.4%), and Best-of-20 restarts gave additional gains (-0.44%).

→ paper-92-KmeansLocalOpt/OPTIMIZATION.md

---

93 — FedWMSAM

FedWMSAM: Fast and Flat Federated Learning via Weighted Momentum and Sharpness-Aware Minimization

Reducing SAM radius (rho=0.1→0.05) combined with more communication rounds (500→700) achieved +1.77% accuracy improvement. The smaller perturbation radius significantly reduces gradient noise in non-IID settings, which is the dominant factor for improvement.

→ paper-93-FedWMSAM/OPTIMIZATION.md

---

94 — BounDr.E

BounDr.E boundary detection

F1 climbed by sweeping the percentile distance threshold and switching to a highly anisotropic Lᵖ geometry (very small p) so thin structures remain connected while suppressing cross-axis noise.

→ paper-94-BounDrE/OPTIMIZATION.md

---

95 — FastFeatureCP

Accelerating Feature Conformal Prediction via Taylor Approximation

The key change was clipping gradient norms at the 93rd percentile of calibration gradient norms. This prevents outlier gradient magnitudes from inflating prediction intervals. Also removed the conservative -1 from the quantile border formula for marginally tighter intervals.

The optimization reduced band_length by 5.49% while maintaining 90% coverage.

→ paper-95-FastFeatureCP/OPTIMIZATION.md

---

96 — M3SVM

Multi-Class Support Vector Machine with Differential Privacy

The optimization applied K=50 noise ensemble soft voting at inference time (post-processing with no privacy cost), antithetic sampling for variance reduction, and stratified train/test splitting. Together these improved accuracy from 0.8882 to 0.9093 (+2.37%).

→ paper-96-M3SVM/OPTIMIZATION.md

---

97 — WassersteinTL

Wasserstein Transfer Learning for survival outcomes

Best RMSPR 0.03284→0.03161 (-3.74%) achieved by replacing linear kernel with Gaussian kernel for bias correction weights (biggest win, +2.60%), tuning lambda=2.0 (+0.76%), selecting 80 nearest source countries (+0.28%), and lambda ensemble (+0.19%). Target (≤0.0321) ACHIEVED.

→ paper-97-WassersteinTL/OPTIMIZATION.md

---

98 — XMahalanobis

Transformer Feature Mixing for Reliable OOD Detection

The optimization stacked three changes: (1) normalize features after layer mixing instead of per-layer before, (2) increase cosine classifier scale from 25 to 35, and (3) increase AdaptFormer bottleneck dim from 4 to 16. These improved AUROC from 0.9729 to 0.9833 (+1.07%) and FPR95 by 43%.

→ paper-98-XMahalanobis/OPTIMIZATION.md

---

99 — OLLALanding

Fast non-log-concave sampling with Landing under constraints

test NLL fell sharply after increasing n_steps so the thinned chain contributes five× more samples for the Gaussian-credit test likelihood—variance reduction dominates despite only modest ESS movement.

→ paper-99-OLLALanding/OPTIMIZATION.md

---

100 — InvMissingData

Inverse Methods for Missing Data Imputation

Kernel regression training benefited from quadrupling n_pairs each epoch, which diversified paired gradients and lowered MAE without changing the core architecture.

→ paper-100-InvMissingData/OPTIMIZATION.md

---

101 — ACIA

Anti-Causal Invariant Abstraction via Measure Theory

Best checkpoint tracking combined with cosine annealing LR (1e-3→1e-5) was the dominant win. Extended training to 24 epochs, reduced regularization weights, and applied Top-5 weight averaging (LEAP). Improved accuracy from 98.88% to 99.45% (+0.57%).

→ paper-101-ACIA/OPTIMIZATION.md

---

102 — StochasticFF

Stochastic Forward-Forward Learning through Representational Dimensionality Compression

CIFAR-10 accuracy crossed the +2% uplift goal by lengthening the linear probe phase and keeping the two-phase training schedule stable so the compressed forward-forward representation fully separates classes.

→ paper-102-StochasticFF/OPTIMIZATION.md

---

103 — Q-DCP

Distributed conformal prediction via message passing (Q-DCP)

Mean prediction-set size shrank after tightening epsilon_0, swapping flaky fsolve roots for brentq inside ADMM, and fixing torch/numpy seed alignment so calibration splits are reproducible without breaking coverage.

→ paper-103-QDCP/OPTIMIZATION.md

---

104 — BalancedActiveInf

Balanced Active Inference

The optimization tuned XGBoost hyperparameters (learning_rate: 0.001→0.1, n_estimators: 1000→300, max_depth: 7→5) to fix severe underfitting. The RMSE dropped from 95.9 to 43.9 (54% improvement), which directly translated to a 24.8% reduction in CI width.

→ paper-104-BalancedActiveInf/OPTIMIZATION.md

---

105 — BinaryClassEval

Evaluating Binary Classifiers Under Label Shift

Applied Laplace smoothing (alpha=2.0) to the prevalence estimate for the African American subgroup, which has only 332 patients with 6 positive cases. This Bayesian regularization improved dca_overall_african_american from 0.900 to 0.922 (+2.4%), exceeding the target of 0.918.

→ paper-105-BinaryClassEval/OPTIMIZATION.md