Reason from Future: Reverse Thought Chain Enhances LLM Reasoning

Conference: ACL 2025
arXiv: 2506.03673
Code: None
Area: LLM/NLP
Keywords: Reverse Reasoning, Bidirectional Chain of Thought, Search Space Reduction, Goal-driven Reasoning, Chain-of-Thought

TL;DR

Proposes the Reason from Future (RFF) reasoning paradigm, which achieves bidirectional reasoning by alternating between reverse reasoning (decomposing backward from the goal) and forward reasoning (approaching the goal from the current state). It significantly outperforms methods like CoT, ToT, and CR on benchmarks including Game of 24, GSM8K, and MATH-500, while substantially reducing the search space.

Background & Motivation

The reasoning ability of Large Language Models (LLMs) is a key factor that determines their capability boundaries. Existing reasoning paradigms can be classified into several categories:

Chain-of-Thought (CoT): Generates intermediate reasoning steps sequentially, but is inherently forward-looking and local, making it prone to trapping in local optima.

Tree/Search Reasoning (ToT/MCTS): Explores multiple paths via search trees, but suffer from exponentially growing search spaces and massive computational overhead.

Progressive Prompting (PHP/CR): Gradually refines reasoning through multi-round interactions and prompts, but is prone to overthinking.

A common core limitation of these methods is the lack of a global perspective. During each reasoning step, the model concentrates solely on the current state without a macroscopic understanding of the overall structure and solution path of the problem. This leads to two issues: (1) blindly exploring numerous irrelevant branches, thereby wasting computational resources; (2) accumulation of errors in forward reasoning, making subsequent steps difficult to correct.

Human problem-solving approaches differ—research shows that humans first establish a holistic mental model to form a "cognitive roadmap" of the solution path, and then maintain dynamic calibration with the final goal during concrete operations.

Inspired by the "backward deduction from the goal" strategy in maze solving, RFF (Reason from Future) is proposed: decomposing the previous step (last step) from the goal state through reverse reasoning, using this sub-goal to guide forward reasoning, and alternating between the two directions until they converge. The key insight is that reverse reasoning prioritizes identifying core logical relationships and exerts goal-directed constraints on intermediate steps, thereby narrowing the search space and mitigating error accumulation.

Method

Overall Architecture

RFF is a reasoning paradigm that alternately executes reverse and forward reasoning, consisting of three core components: - Last Step Generator G(): The reverse reasoning component, which decomposes the previous sub-goal from the target state. - Stepwise Forward Reasoner R(): The forward reasoning component, which moves a step closer to the new sub-goal from the current state. - State Checker C(): The state checking component, which determines whether the forward reasoning state has reached the target state.

At step \(i\): 1. Reverse reasoning generates a new sub-goal \(T_i = G(p_\theta, S_{i-1}, T_{i-1})\) 2. Forward reasoning advances toward the sub-goal \(S_i = R(p_\theta, S_{i-1}, T_i, A_{i-1})\) 3. State checking verifies if \(S_i\) satisfies \(T_i\)

This process repeats until \(S_i = T_i\) (the forward state meets the target state), at which point the final answer is outputted.

Key Designs

1. Last Step Generator (Reverse Reasoner):

   • Mechanism: Decomposes a target state \(T_i\), combined with the current state \(S_i\), into the previous sub-goal \(T_{i+1}\).
   • Design Motivation: Reverse decomposition forces the model to thin backward from the outcome: "What is required in the last step to reach the goal?", prioritizing the establishment of core logical relationships.
   • The form of the sub-goal depends on the task characteristics: a set of numbers in Game of 24, or intermediate variables to be solved in math word problems.
   • Key constraint: The transition step from \(T_{i+1}\) to \(T_i\) must be explicitly outputted to guarantee the correctness of the reverse decomposition.

2. Two Stepwise Forward Reasoner Strategies:

   • RFF-T (Tree-type): Applicable to search tree problems (e.g., Game of 24, mazes), where the solution is a branch of the tree. It uses an avoidance set \(\{A\}\) to record failed attempts, preventing repeating erroneous paths at the same level: \(S_i \leftarrow R(p_\theta, S_{i-1}, T_i, A_{i-1})\).
   • RFF-G (Graph-type): Applicable to DAG-type problems (e.g., math problems), where all prior computed information is either useful or redundant but harmless. It accumulates all state information: \(S_i \leftarrow S_{i-1} \cup R(p_\theta, S_{i-1}, T_i)\).

3. Two State Checker Strategies:

   • RFF-T: Checks whether \(S_i\) overlaps with \(T_i\) or can reach it in a single step. It is equipped with a Verifier \(V()\) to validate path correctness, backtracking to prior states in case of failure.
   • RFF-G: Checks whether all information required for the target state is already contained in the current state. No backtracking is needed, since each step contributes a node to the DAG.

Loss & Training

• RFF is a training-free reasoning paradigm requiring no additional training or fine-tuning.
• All components are implemented via meticulously designed prompts, leveraging the in-context learning capabilities of LLMs.
• A 1-shot exemplar is used to guide the model to perform formatted reasoning.
• Game of 24 uses a temperature of 0.7 (maintaining search diversity), while math problems use greedy search (excluding the impact of randomness).

Key Experimental Results

Main Results

Game of 24 (Search Tree Task):

Model	Method	Accuracy	Visited States
GPT-4	CoT	3%	1.0
GPT-4	ToT(n=5)	74%	-
GPT-4	CR(n=1)	84%	11.7
GPT-4	CR(n=5)	94%	13.7
GPT-4	RFF(n=5)	95%	9.3
Llama3-8B	CR(n=5)	19%	89.8
Llama3-8B	RFF(n=5)	89%	9.9

RFF achieves 89% with Llama3-8B, outperforming GPT-4 + CR(n=1) at 84%, while the number of visited states is only 1/9 of CR's.

Mathematical Reasoning (DAG Task):

Model	Method	GSM8K	SVAMP	ASDiv	MATH	Average
Llama3-8B	CoT	75.6%	80.5%	82.3%	32.8%	67.8%
Llama3-8B	CR	77.0%	71.2%	84.8%	40.2%	68.3%
Llama3-8B	RFF	83.8%	89.7%	86.7%	41.4%	75.4%
Qwen2.5-7B	CoT	87.2%	92.1%	88.0%	74.6%	85.5%
Qwen2.5-7B	RFF	89.5%	95.1%	92.2%	79.8%	89.1%

Ablation Study

Redundant Thought Study (5-Digit Game of 24):

Model	Method	Accuracy	Visited States	Description
GPT-4	CR(n=5)	76% (↓18%)	7.06	Decreases significantly after adding redundant numbers
GPT-4	RFF(n=5)	89% (↓6%)	5.96	More robust, minimally affected by redundancy
Llama3-8B	CR(n=5)	26% (↓from 19%)	96.56	Search space explosion
Llama3-8B	RFF(n=5)	85% (↓4%)	28.62	Maintains high accuracy

Robustness Study (GSM-Symbolic): - RFF performs more stably across 50 GSM-Symbolic variants, with accuracies more concentrated in high-value intervals. - CoT's average accuracy is significantly lower than RFF's, with a more dispersed distribution.

Commonsense Reasoning:

Method	CommonQA	LogiQA	Average
CoT	73.1%	41.8%	57.5%
CR	75.4%	45.5%	60.5%
RFF	77.1%	45.2%	61.2%

Key Findings

• Search Efficiency Revolution: While achieving the highest accuracy on Game of 24, RFF visits the fewest states (9.3 vs. 13.7 for CR), reducing the search space by approximately 30%.
• Larger Gains on Weaker Models: The improvement of RFF relative to CoT on Llama3-8B (+7.6 average) is significantly greater than that on Qwen2.5-7B (+3.6), indicating that RFF provides stronger complementary benefits to weaker models.
• Resilience against Overthinking: CR performs worse than CoT (71.2% vs. 80.5%) on the simple SVAMP task due to overthinking; in contrast, RFF's State Checker terminates reasoning timely to avoid this issue.
• Redundancy Robustness: When redundant information (such as an extra number) is introduced, CR's performance drops drastically, whereas RFF remains almost unaffected, confirming the efficacy of goal-directed search pruning.
• Robustness on GSM-Symbolic: Although the accuracy of both methods declines across variants, RFF exhibits smaller performance drops and more concentrated distributions.

Highlights & Insights

• Cognitive Science-Inspired Design: Draws inspiration from human holistic mental modeling in problem-solving and backward maze navigation, translating cognitive science insights into an engineerable reasoning paradigm.
• Information Complementarity of Bidirectional Reasoning: Reverse reasoning provides goal constraints (knowing where to go) while forward reasoning provides information accumulation (knowing what is available), forming an effective closed loop of information.
• Two Variants of a Unified Framework: The categorization of RFF-T (search tree) and RFF-G (directed acyclic graph) precisely models the solution space structures of different problem typologies.
• Adaptive Role of State Checker: On simple problems, forward and reverse reasoning converge quickly, degenerating back to simple CoT; on complex problems, it fully exploits bidirectional reasoning, achieving adaptive complexity.

Limitations & Future Work

• Dependency on Model's Reverse Reasoning Capability: When the model has not been specifically trained, the last step of reverse reasoning could fail, leading to overall failure.
• Lack of Verification on Larger Scale Models: Experiments are restricted to 8B and 7B models; the effectiveness on 70B+ models remains unknown.
• Generalization Beyond Math Problems: Performance improvements on commonsense reasoning are relatively modest (+3.7%), requiring validation across more non-mathematical reasoning tasks.
• Absence of Comparison with Latest o1/o3-class Reasoning Models: These models natively integrate similar search and verification mechanisms.
• Future Improvement Directions: Specifically enhancing reverse reasoning capabilities through fine-tuning or reinforcement learning; exploring the possibility of combining RFF with MCTS.

Related Work & Insights

• CoT (Wei et al., 2022) pioneered a new direction in reasoning paradigm design, and RFF is a natural extension of this.
• AoT (Sel et al., 2023) and AoT+ also introduce a global perspective, but it is only acquired during the exploration process (degenerating to conventional ToT), whereas RFF establishes a global perspective from the beginning via reverse reasoning.
• The cumulative reasoning concept of CR (Zhang et al., 2023) is reflected in RFF-G, but RFF introduces additional goal constraints.
• The reverse reasoning mechanism of RFF can be combined with SFT/RLHF to train models to better execute "backward deduction from results".

Rating

• Novelty: ⭐⭐⭐⭐⭐ The integration in bidirectional reasoning is novel and elegant, with precise categorizations of RFF-T/RFF-G.
• Experimental Thoroughness: ⭐⭐⭐⭐ Covers both search tree and DAG tasks, including redundancy and robustness analyses, though lacking experiments on larger models.
• Writing Quality: ⭐⭐⭐⭐ The framework description is clear and the algorithm pseudocode is well-structured, although some experimental analyses are relatively brief.
• Value: ⭐⭐⭐⭐⭐ Proposes a highly versatile, plug-and-play reasoning enhancement paradigm, showing particularly significant gains on weaker models.

Related Papers

• [ACL 2025] Reversal of Thought: Enhancing Large Language Models with Preference-Guided Reverse Reasoning Warm-up
• [ACL 2025] GAMEBoT: Transparent Assessment of LLM Reasoning in Games
• [ACL 2025] LR²Bench: Evaluating Long-chain Reflective Reasoning Capabilities of Large Language Models via Constraint Satisfaction Problems
• [ACL 2025] Structural Reasoning Improves Molecular Understanding of LLM
• [ACL 2025] Stepwise Reasoning Disruption Attack of LLMs