TrimSFT Gains 26.9 Points on MATH500 by Targeting Intermediate-Confidence Tokens

A team of researchers at Dartmouth College has introduced TrimSFT, a modified supervised fine-tuning method that concentrates learning signal on the tokens a model finds genuinely instructive — those it has not yet mastered but is plausibly capable of learning. By selectively suppressing gradient updates from both ends of the difficulty spectrum, the approach, accepted at EMNLP 2026, produced gains of up to 26.9 points on MATH500 over standard supervised fine-tuning. The gain required no reference model and no extra forward pass through the network — a substantially lower overhead than competing post-training methods.

The paper, "Which Tokens Should SFT Actually Learn? A Token-Trimming Perspective on Mathematical Reasoning," was submitted to arXiv on September 9, 2026 by Yaning Jia, Chunhui Zhang, Wenxuan Xu, Xingjian Diao, Xiaoyuan Wang, and senior author Soroush Vosoughi, an associate professor of computer science at Dartmouth who holds a PhD from MIT and has received research awards from Google and Amazon.

The Structural Flaw in Standard Supervised Fine-Tuning

To understand why TrimSFT matters, it helps to understand what standard SFT is actually doing at the token level. When a language model undergoes supervised fine-tuning, the training objective is cross-entropy loss: for every token position in every training example, the model is penalized in proportion to how far its predicted probability is from 1.0 on the correct token. Crucially, this loss applies uniformly across every token regardless of whether the model was already getting it right before training began.

That uniformity creates two distinct inefficiencies. First, for tokens the model has already mastered — those where its predicted probability for the correct token is already high — continued training pressure pushes toward over-sharpening, making the model excessively confident in patterns it has already captured and potentially degrading generalization. Second, for tokens that are beyond the model's current capability — cases where the gold token is so unexpected given the model's current representation that the gradient provides little usable learning signal — the loss still applies full weight, meaning noisy or high-magnitude updates dominate training without actually improving model performance on those positions.

This is particularly consequential for mathematical reasoning, where a chain-of-thought solution contains a mixture of structural tokens (notation, operators, formatting) that the model trivially predicts alongside genuinely difficult inferential steps that sit at the frontier of what the model can learn. Standard cross-entropy treats a mundane closing parenthesis and a pivotal algebraic substitution identically.

How TrimSFT's Logit-Gap Gaussian Actually Works

TrimSFT's solution is built around the concept of the logit gap — the difference between the model's raw score (logit) for the correct gold token and the raw score of the single strongest competing alternative token. When the logit gap is large and positive, the model already strongly prefers the gold token; the position is effectively mastered. When the logit gap is near zero or negative, the model is uncertain or predicts the wrong token entirely.

The paper's core contribution is a Gaussian-shaped weight function applied over this logit gap dimension. The weight peaks at an intermediate margin value m — representing the zone of moderate confidence — and falls off on both sides following the shape of a normal distribution with bandwidth parameter τ. Concretely: tokens where the logit gap is very large receive a small weight (easy, already mastered, skip); tokens where the logit gap is near zero or negative also receive a small weight (hard, currently unlearnable, skip); tokens where the logit gap sits in the intermediate productive zone receive a large weight (concentrate learning here).

This bilateral trimming is what distinguishes TrimSFT from its closest predecessors. Focal loss — a technique adapted from computer vision, where it was developed to handle class imbalance in dense object detection — is one-sided: it downweights well-classified examples but upweights difficult ones, which in the SFT context would mean increasing gradient pressure on the already-unlearnable tokens. Dynamic Fine-Tuning (DFT), a recent approach from Wu et al. that has gained traction in the post-training community, reweights the SFT loss using the token's predicted probability rather than the logit gap, and applies this reweighting monotonically — always in one direction. Neither method trims both extremes simultaneously.

A key practical advantage: TrimSFT requires no reference model (unlike methods derived from RLHF or DPO, which need a frozen copy of the policy for KL regularization) and no additional forward pass through the network (unlike importance-sampling methods that compute ratios across multiple passes). The Gaussian weight is computed from the current-step logit values already available during the standard forward pass. This makes it a single-hyperparameter modification to the standard training loop — add the Gaussian weight to the per-token loss scaling, set τ, and train.

Ablations Reveal a Counterintuitive Sensitivity Pattern

The authors' ablation experiments reveal a finding worth pausing on: the bandwidth τ of the Gaussian weight matters substantially more than the exact location of the center margin m. In practice this means a practitioner who gets the width of the productive learning window approximately right — not too narrow, not too wide — can tolerate more imprecision about exactly where the window is centered. Intuitively, this makes sense: any reasonable center will sit in the "middle difficulty" region; what matters is whether the bell curve is sharp enough to actually exclude the extremes.

The authors also tested half-trim variants: configurations that suppress only the easy end (mastered tokens) or only the hard end (currently unlearnable tokens), leaving the other extreme to train normally. Both half-trim variants showed inferior performance trade-offs compared to the full bilateral trim. This finding suggests that both types of wasteful gradient update are materially harmful to training dynamics — removing either one alone captures only part of the gain.

Benchmark Results Across Six Models: What the Numbers Mean

MATH500, the primary benchmark on which TrimSFT reports its headline +26.9-point improvement, is a curated 500-problem subset of the broader MATH dataset developed by Dan Hendrycks and colleagues. It draws from American mathematics competition problems — AMC 10, AMC 12, AIME, and similar — spanning algebra, geometry, number theory, combinatorics, and pre-calculus. The problems require step-by-step reasoning in LaTeX format and are considered substantially more challenging than grade-school benchmarks such as GSM8K. An independent evaluation platform, Artificial Analysis, tracks frontier models on MATH500, and the benchmark has become a standard signal for comparing post-training methods on the reasoning dimension.

The Dartmouth team evaluated TrimSFT against three baselines — the unmodified pretrained base (no SFT), standard SFT, focal-loss SFT (FSFT), and DFT — across six base models: Llama3.2-3B, Llama3.1-8B, DeepSeekMath-7B, Qwen2.5-Math-1.5B, Qwen2.5-Math-7B, and Qwen3-4B-Base. The choice to test base models only — rather than instruction-tuned or chat variants — is a deliberate experimental design decision. It reduces the confounding influence of prior instruction tuning and allows a cleaner comparison of the SFT objectives in isolation. The authors report in the paper that TrimSFT is the top-performing method on MATH500 across all six models, and achieves the best average performance across five of the six models when all five evaluated benchmarks are considered together.

All benchmark results are author-reported from the paper and have not yet been independently reproduced by external evaluators. The paper's acceptance at EMNLP 2026 Findings implies peer review, but independent replication typically follows conference publication.

Where TrimSFT Fits in the Crowded Post-Training Landscape

The 2025–2026 period has produced an unusually dense cluster of papers questioning the assumption that all SFT tokens should be trained equally. InfoSFT, proposed in 2026, emphasizes what the authors call "middle-confidence" tokens using an entropy-based weighting rule and arrives at a qualitatively similar conclusion from a different mathematical starting point: learning and forgetting tradeoffs favor tokens that are informative but not trivially easy. The PEAR approach (Zhang et al., 2026) takes a different angle — optimizing SFT not for raw SFT performance but for the quality of the resulting RL initialization — and demonstrates that the post-SFT RL stage is significantly affected by which SFT objective was used.

This convergence of independent research directions is relevant context for TrimSFT. It suggests that the standard SFT paradigm is broadly recognized as suboptimal, and that the field is actively exploring which properties of the SFT objective matter most. TrimSFT's specific contribution — using the logit gap as the difficulty signal and applying a symmetric Gaussian trim — adds a distinct and empirically competitive entry to this space. Its comparison against DFT is particularly pointed, since DFT has been increasingly cited in the post-training literature: on five of six models, TrimSFT's bilateral Gaussian trim outperforms DFT's monotonic probability-based reweighting.

The Pipeline Implication the Paper Does Not Discuss

Current practice in reasoning model development treats SFT as a first stage and reinforcement learning — typically using methods like GRPO or variants — as a second, separate stage. Research into SFT-RL pipelines has established that SFT checkpoint quality matters substantially for subsequent RL performance: models initialized from stronger SFT checkpoints do not always outperform those initialized from weaker ones, because overfit or poorly calibrated SFT checkpoints reduce what researchers call "RL plasticity" — the model's capacity to improve further under reward-based training.

TrimSFT's paper does not report results from SFT-then-RL pipelines. Its experiments evaluate SFT performance in isolation. But given the documented relationship between SFT calibration and RL plasticity, there is a reasonable hypothesis — not yet tested — that TrimSFT's better-calibrated confidence distribution could translate into larger downstream RL gains than a standard SFT checkpoint of comparable surface accuracy. This is an open research question that follows directly from TrimSFT's core claim: that its confidence reshaping is more balanced, not merely more accurate.

Limitations and Open Questions

The paper is confined to mathematical reasoning. TrimSFT's gains on MATH500 and related benchmarks do not yet establish that bilateral logit-gap trimming works equally well on code generation, language following, or other fine-tuning targets where the token difficulty distribution may have a different shape. Practitioners hoping to apply TrimSFT to domains outside math would need to run their own validation experiments.

The code repository listed in the paper — github.com/karpning/TrimSFT — returned a 404 error at the time of this article's publication, suggesting the repository had not yet been made public. Adoption in production pipelines depends on the code becoming available following the EMNLP 2026 camera-ready process.

The evaluation set also stays within the 1.5B–8B parameter range. Whether the bilateral trim principle holds at 70B+ parameter scales, where gradient dynamics and token difficulty distributions could shift, is not addressed. Similarly, the paper tests base models rather than instruction-tuned or chat-formatted variants; the behavior of the logit-gap signal in those settings may differ.

The single hyperparameter τ requires tuning. While the authors' finding that τ matters more than m simplifies the search, it does not eliminate it. Different model families and training datasets may need different τ values, which means practitioners should treat the hyperparameter as genuinely requiring validation rather than as a fixed constant.

What the Field Should Watch Next

The immediate next milestone is the public release of TrimSFT's code and the camera-ready paper at EMNLP 2026, expected later this fall. Community reproduction — running TrimSFT against existing public baselines on shared benchmark sets — will be the most important validation step. If independent evaluators confirm the MATH500 gains even approximately, TrimSFT has a credible path to becoming a standard component of mathematical reasoning fine-tuning pipelines.

The broader question is whether the logit-gap signal generalizes as a useful difficulty proxy beyond math. The concurrent literature on token-level reweighting — InfoSFT, DFT, PEAR, and others — each proposes a different proxy for "how much should this token contribute to training." The field does not yet have consensus on which signal is most informative across domains. TrimSFT's bilateral Gaussian trim is the most explicit implementation of the intuition that the productive learning zone is bounded on both sides, and that result — held up across five of six model families at peer review — provides stronger empirical grounding for that intuition than most prior work.