Look Forward Once (LFO): A Design Principle

• Look Forward Once is a design principle emphasizing single-pass computation for front-loaded decision-making in adaptive control and transformer inference.
• It eliminates iterative stages by using one forward pass to extract multiple judgments, rationales, or perform cache optimization.
• LFO spans diverse applications—from MPC and text rationalization to locomotion and audio DSP—though its standardization remains ambiguous.

Searching arXiv for recent and directly relevant papers on “Look Forward Once”, “You Only Forward Once”, and nearby “look once/forward-looking” terminology. Look Forward Once (LFO) is not a standardized arXiv method name. In the most relevant papers, the phrase is best understood as an umbrella for forward-looking or single-pass computation: adaptive MPC that guarantees persistent excitation by only looking forward into the receding prediction horizon, transformer-based systems that extract multiple judgments or rationales from a single forward pass, and locomotion frameworks that use a prior forward traversal to support later backward motion. At the same time, the same acronym can denote the unrelated low frequency oscillator of audio effects. This suggests that LFO is presently better treated as a design principle than as a single canonical architecture (Brüggemann et al., 2020, Zhang et al., 20 Nov 2025, Luo et al., 3 Mar 2026, Mitcheltree et al., 2023).

1. Terminological scope and disambiguation

The closest arXiv usages adjacent to “Look Forward Once” span several distinct literatures.

Exact term	Domain	Relation to LFO
"Forward-looking persistent excitation in model predictive control" (Brüggemann et al., 2020)	Adaptive MPC	Explicitly forward-looking; closest literal match
"You Only Forward Once" (Zhang et al., 20 Nov 2025)	Multimodal judging	Single forward pass yields multiple requirement-level decisions
"You Only Forward Once" (Jiang et al., 2023)	Text rationalization	Prediction and rationale extraction in one forward pass
"LOOK-M" (Wan et al., 2024)	Multimodal long-context inference	One-time prefill compression; adjacent rather than identical
"Look Forward to Walk Backward" (Luo et al., 3 Mar 2026)	Legged locomotion	Prior forward sensing used for later backward locomotion
"Modulation Extraction for LFO-driven Audio Effects" (Mitcheltree et al., 2023)	Audio DSP	LFO means low frequency oscillator

The ambiguity is not merely lexical. In some papers, “forward” means future-oriented optimization or planning; in others it means a single transformer evaluation; in others it means a literal forward traversal of terrain. Conversely, some “look once” titles are only tangential: the sorghum-head counting paper is a YOLOv4-based single-shot detector, not an LFO framework, and the intervertebral-disc paper uses “look once” for a permutation-invariant post-processing module rather than for a method named LFO (Mosley et al., 2020, Azad et al., 2022).

A plausible implication is that LFO, when used informally, denotes a family resemblance among methods that avoid iterative decoding, backward-looking memory constraints, or repeated search. The common pattern is not a shared backbone but a shared computational stance: decide, compress, or certify using one forward-looking computation rather than an iterative or history-preserving procedure.

2. Single-forward-pass prediction, judgment, and rationalization

The clearest “forward once” formulation appears in the multimodal judging paper "You Only Forward Once: An Efficient Compositional Judging Paradigm" (Zhang et al., 20 Nov 2025). There the input is an image $I \in \mathbb{R}^{3\times H\times W}$ and a set of requirements $(\mathbf{p}_i)_{i=1}^N$, with judging function

$$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$

The method inserts each requirement into a structured template, appends an unknown answer token, records the placeholder positions $(pos_i)_{i=1}^N$, runs a single forward pass through a decoder-only MLLM, and extracts requirement-level logits by

$$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$

Inference then compares the probabilities of "yes" and "no" at each answer slot. The architecture remains autoregressive, but inference is not autoregressive in the usual token-by-token sense; the method exploits the fact that next-token logits for every position are available after one teacher-forced forward pass. The paper reports $Acc_\mathrm{property}=92.3\%$ and $Acc_\mathrm{sample}=46.6\%$ for YOFO on the SA-1B-derived validation set with Qwen3-VL, and on LRVS-Fashion reranking it reports $Error_\mathrm{rank}=3.7\%$ and throughput $47.6$ pairs/s for YOFO with Qwen3-VL (Zhang et al., 20 Nov 2025).

The text-rationalization paper "You Only Forward Once: Prediction and Rationalization in A Single Forward Pass" uses the same acronym for a different mechanism (Jiang et al., 2023). Its target is unsupervised rationale extraction for text classification, where prior Rationalizing Neural Prediction systems followed a generate-then-predict factorization

$$P(\mathbf{y}\mid\mathbf{X})=P(\mathbf{y}\mid\mathbf{Z})P(\mathbf{Z}\mid\mathbf{X}),$$

with explicit sparsity and contiguity penalties on a rationale mask. YOFO replaces that two-phase structure with a single-phase formulation over $(\mathbf{p}_i)_{i=1}^N$0, using a pretrained LLM and cumulative layerwise masks. At layer $(\mathbf{p}_i)_{i=1}^N$1, a selector $(\mathbf{p}_i)_{i=1}^N$2 predicts $(\mathbf{p}_i)_{i=1}^N$3, the cumulative mask is updated by $(\mathbf{p}_i)_{i=1}^N$4, and token count decreases monotonically through depth. The paper organizes the transformer into Information Gathering, Rationale Generation, and Performance Boosting stages, and optimizes task loss with sparsity and contiguity regularization. On BeerAdvocate, it reports an improvement of up to $(\mathbf{p}_i)_{i=1}^N$5 in token-level F1 over previous state-of-the-art methods (Jiang et al., 2023).

Taken together, these two YOFO papers show that “forward once” need not mean the same thing across domains. In multimodal judging, it means harvesting many answer logits from designated template positions in one pass. In rationale extraction, it means predicting labels and rationale-supporting token survival within one PLM pass rather than by a generator–predictor pipeline. The shared principle is elimination of an explicit sequential intermediate stage.

3. Forward-looking persistent excitation in adaptive MPC

Among the cited works, the most literal connection to an LFO reading appears in "Forward-looking persistent excitation in model predictive control" (Brüggemann et al., 2020). The paper studies simultaneous regulation and parameter estimation for nonlinear discrete-time systems

$(\mathbf{p}_i)_{i=1}^N$6

with dynamics linear in unknown parameters,

$(\mathbf{p}_i)_{i=1}^N$7

and regressor form

$(\mathbf{p}_i)_{i=1}^N$8

Its central claim is that persistent excitation can be guaranteed by only looking forward in time into the receding prediction horizon, rather than by looking backward and preserving prior regressor data. Persistent excitation is defined by the windowed information-matrix condition

$(\mathbf{p}_i)_{i=1}^N$9

for all $$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$0.

The mechanism is not an online dual-control optimizer but an offline generation of a periodic exciting reference trajectory perturbing the equilibrium. A periodic input perturbation is designed so that the corresponding reference trajectory is persistently exciting; then a standard reference-tracking nonlinear MPC follows that trajectory online. The online cost is

$$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$1

and parameter estimation is performed with a recursive least-squares identifier with forgetting factor. Under the stated local assumptions, the paper proves practical convergence of the tracking error to a robust positively invariant neighborhood and exponential convergence of the parameter-estimation error to a noise-dependent ball.

This forward-looking construction differs sharply from backward-looking PE-MPC schemes that preserve windows of past controls or regressors inside the optimization. In that sense, the paper provides the most direct formalization of an LFO-like idea: excitation is certified from intended future behavior rather than from constraints on stored past data. The limitation, stated explicitly in the paper, is that the results are local in state and parameter-estimate space (Brüggemann et al., 2020).

4. One-shot cache decisions in multimodal long-context inference

The paper "LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference" is not labeled LFO, but it embodies a closely related one-shot decision pattern (Wan et al., 2024). The setting is long-context MLLMs whose multimodal KV cache grows with interleaved text and many images. LOOK-M compresses the cache once during prompt prefill rather than continuously during decoding. The multimodal prompt is written as

$$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$2

and cumulative prompt attention scores are computed by

$$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$3

The core heuristic is a text-prior boost: $$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$4 which strongly favors retention of text tokens. The compressed cache keeps a recent window of size $$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$5 and the top-$$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$6 earlier tokens by score: $$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$7 with $$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$8. To compensate for evicted visual context, LOOK-M merges pruned KV pairs back into retained ones using cosine-similarity matching and one of three strategies: averaged, weighted, or pivotal merging.

The paper characterizes the method as fine-tuning-free and reports that with a 20% budget, memory drops from $$f(I, (\mathbf{p}_i)_{i=1}^N) = (\mathbf{a}_i)_{i=1}^N,\qquad \mathbf{a}_i\in\{\text{yes},\text{no}\}.$$9 GiB to $(pos_i)_{i=1}^N$0 GiB and latency improves from $(pos_i)_{i=1}^N$1 ms/token to $(pos_i)_{i=1}^N$2 ms/token; with a 5% budget, memory drops to $(pos_i)_{i=1}^N$3 GiB and latency to $(pos_i)_{i=1}^N$4 ms/token. The abstract also states reductions of KV cache memory by $(pos_i)_{i=1}^N$5 to $(pos_i)_{i=1}^N$6 in some cases and up to $(pos_i)_{i=1}^N$7 faster decoding (Wan et al., 2024).

This suggests an LFO-style interpretation in cache management: the model uses a single prefill-time analysis to decide what to preserve for all subsequent decode steps. The paper itself, however, is careful to frame LOOK-M as adjacent to, not identical with, any named “Look Forward Once” method.

5. Forward observation as memory for later control

"Look Forward to Walk Backward: Efficient Terrain Memory for Backward Locomotion with Forward Vision" shifts the meaning of “forward” from transformer evaluation or MPC planning to physical traversal (Luo et al., 3 Mar 2026). The method, LF2WB, addresses quadruped locomotion with a forward-facing egocentric depth camera and proprioception. During forward motion, the robot writes a compact associative terrain memory; during backward locomotion, it retrieves that memory to avoid obstacles and negotiate gaps without rearward vision. The main platform is the DEEP Robotics Lite3 quadruped, with a forward-facing RealSense depth camera at $(pos_i)_{i=1}^N$8 Hz and a locomotion controller at $(pos_i)_{i=1}^N$9 Hz.

The architecture comprises an estimator, actor, and critic. Depth and proprioception are encoded as

$$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$0

A stack of DeltaNet-Transformer layers maintains fixed-size recurrent memory states $$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$1. For each layer,

$$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$2

and readout is

$$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$3

The update is explicitly described as a delta-rule selective overwrite that softly removes then writes the memory state along the active subspace.

The significance of LF2WB for an LFO reading lies in its fixed-state, constant-time recurrent deployment. The paper states that training uses hardware-efficient parallel computation, while inference remains recurrent with constant per-step cost and a constant-size state. In simulation, LF2WB outperforms Gated DeltaNet, Linear Attention, LSTM, and Transformer-XL on forward-then-backward protocols. At difficulty $$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$4, for example, success rates are $$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$5 on P1, $$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$6 on P2, $$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$7 on P3, $$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$8 on P4, and $$\mathbf{h}=\mathrm{VLM}(I,\mathbf{t}),\qquad \mathbf{l}_i=\mathbf{h}[pos_i-1].$$9 on P5, all above the corresponding baselines. Real-world experiments on trails nearly $Acc_\mathrm{property}=92.3\%$0 m long include backward traversal over a $Acc_\mathrm{property}=92.3\%$1 m high step, a $Acc_\mathrm{property}=92.3\%$2 m wide gap, and composite step-gap layouts (Luo et al., 3 Mar 2026).

A plausible implication is that LF2WB realizes a physical rather than purely computational LFO idea: a route is looked at in the forward direction once, encoded into compact latent memory, and then exploited later under partial observability.

6. Ambiguities, misconceptions, and boundaries of the label

The strongest misconception is that LFO names a single agreed-upon architecture. The cited papers do not support that reading. In audio work, "Modulation Extraction for LFO-driven Audio Effects" uses LFO in the established DSP sense of low frequency oscillator; the paper is about extracting arbitrary modulation signals for phaser, flanger, and chorus, and explicitly states that LFO means low frequency oscillator, not “look forward once” (Mitcheltree et al., 2023).

A second misconception is that every “look once” title belongs to the same methodological lineage. "Image-Based Sorghum Head Counting When You Only Look Once" is a YOLOv4-based single-shot object detector for UAV phenotyping, with custom anchors and reported test mAP $Acc_\mathrm{property}=92.3\%$3; it does not introduce an LFO method (Mosley et al., 2020). "Intervertebral Disc Labeling With Learning Shape Information, A Look Once Approach" uses a U-Net-based candidate generator plus a permutation-invariant post-processing network that classifies candidate points in one shot; it explicitly says “look once,” not “look forward once,” and introduces no LFO acronym (Azad et al., 2022).

The papers closest to an LFO interpretation also expose important limits. The multimodal judging YOFO is still built on a causal autoregressive backbone, depends on explicit requirement templates, and its full application pipeline includes external query decomposition and score aggregation. The text-rationalization YOFO weakens the classical rationale-faithfulness criterion by treating rationales as support for prediction rather than as sufficient predictors. LOOK-M performs once-only compression during prefill, not end-to-end one-pass reasoning from raw query to final answer. LF2WB requires prior forward exposure to the terrain and does not solve completely novel backward traversal of unseen terrain behind the robot. The adaptive MPC paper gives local, not global, guarantees.

The cumulative picture is therefore precise but non-unified. “Look Forward Once” is a useful interpretive label for methods that front-load decision-making into a single forward-looking computation, whether that computation is a transformer pass, a prefill-time cache analysis, an offline exciting-reference construction, or a forward traversal that writes terrain memory. It is not, in the current cited arXiv literature, a single standardized framework.